CN115423040A - User portrait recognition method and AI system for interactive marketing platform - Google Patents

Abstract

Translated fromChinese

本发明实施例提供的互动营销平台的用户画像识别方法及AI系统，通过对选定用户的业务行为日志进行业务行为描述向量挖掘得到选定用户的业务行为描述向量，再依据选定用户的业务行为描述向量对临时用户画像进行调整，从而得到选定用户的最终用户画像，本发明实施例采用临时用户画像获取选定用户的最终用户画像，在具有模板指示的情况下，使得选定用户的最终用户画像偏离性低，效率和精确性高。

The user portrait recognition method and AI system of the interactive marketing platform provided by the embodiment of the present invention obtain the business behavior description vector of the selected user by mining the business behavior description vector of the business behavior log of the selected user, and then obtain the business behavior description vector of the selected user according to the business behavior description vector of the selected user. The behavior description vector adjusts the temporary user portrait to obtain the final user portrait of the selected user. In the embodiment of the present invention, the temporary user portrait is used to obtain the final user portrait of the selected user. In the case of template instructions, the selected user's End user portraits have low deviation, high efficiency and accuracy.

Description

Translated fromChinese

互动营销平台的用户画像识别方法及AI系统User portrait recognition method and AI system for interactive marketing platform

技术领域technical field

本申请涉及业务营销、人工智能领域，具体而言，涉及一种互动营销平台的用户画像识别方法及AI系统。This application relates to the fields of business marketing and artificial intelligence, and specifically relates to a user portrait recognition method and AI system for an interactive marketing platform.

背景技术Background technique

随着互联网的高速发展，各种互联网运营平台通过对其平台上的用户的行为习惯进行分析，得到用户们的用户画像，以此进行相应的营销动作。对于高日活的运营平台而言，例如电商品台、视频平台、阅读平台等，海量的用户每天会产生大量的业务行为数据，而用户的需求和业务倾向往往是动态变化的，因此需要对大量的业务行为数据进行及时高效地分析处理，得到对应阶段性的用户画像。而目前对于海量的业务行为数据进行分析和画像描绘效率和准确性还很吃力，需要多样化的解决方案。With the rapid development of the Internet, various Internet operating platforms analyze the behavior habits of users on their platforms to obtain user portraits of users, so as to carry out corresponding marketing actions. For Gao Rihuo’s operating platforms, such as e-commerce platforms, video platforms, reading platforms, etc., a large number of users will generate a large amount of business behavior data every day, and user needs and business tendencies often change dynamically, so it is necessary to analyze a large number of Analyze and process business behavior data in a timely and efficient manner, and obtain corresponding phased user portraits. At present, it is still very difficult to analyze massive business behavior data and describe the efficiency and accuracy of the portrait, which requires a variety of solutions.

发明内容Contents of the invention

本发明的目的在于提供一种互动营销平台的用户画像识别方法及AI系统，以提升用户画像的描绘效率和准确性。The purpose of the present invention is to provide a user portrait recognition method and an AI system for an interactive marketing platform, so as to improve the efficiency and accuracy of depicting user portraits.

本申请实施例是这样实现的：The embodiment of this application is implemented in this way:

第一方面，本申请实施例提供了一种互动营销平台的用户画像识别方法，应用于互动营销平台，该互动营销平台与业务终端设备通信连接，该方法包括：获取业务终端设备发送的选定用户的业务行为日志；对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量；依据选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像。In the first aspect, the embodiment of the present application provides a user portrait recognition method of an interactive marketing platform, which is applied to an interactive marketing platform, and the interactive marketing platform is connected to a service terminal device through communication. The method includes: obtaining the selected The user's business behavior log; the business behavior description vector mining of the business behavior log to obtain the business behavior description vector of the selected user; the temporary user portrait is adjusted according to the business behavior description vector of the selected user to obtain the end user of the selected user portrait.

本申请实施例中，通过对选定用户的业务行为日志进行业务行为描述向量挖掘得到选定用户的业务行为描述向量，再依据选定用户的业务行为描述向量对临时用户画像进行调整，从而得到选定用户的最终用户画像，本发明实施例采用临时用户画像获取选定用户的最终用户画像，具有模板指示的情况下，获得的选定用户的最终用户画像可以保证偏离性低，精确性高。In the embodiment of this application, the business behavior description vector of the selected user is obtained by mining the business behavior description vector of the business behavior log of the selected user, and then the temporary user portrait is adjusted according to the business behavior description vector of the selected user, so as to obtain For the end user portrait of the selected user, the embodiment of the present invention uses the temporary user portrait to obtain the end user portrait of the selected user. In the case of template instructions, the obtained end user portrait of the selected user can ensure low deviation and high accuracy .

作为一种可行的实施方式，对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量，包括：通过画像生成模型中的向量挖掘模块对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量；依据选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像，包括：通过画像生成模型中的画像调整单元依据向量挖掘模块产生的选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像。As a feasible implementation, the business behavior description vector mining is carried out on the business behavior log to obtain the business behavior description vector of the selected user, including: performing business behavior description vector mining on the business behavior log through the vector mining module in the portrait generation model , get the business behavior description vector of the selected user; adjust the temporary user portrait according to the business behavior description vector of the selected user, and obtain the final user portrait of the selected user, including: mining according to the vector through the portrait adjustment unit in the portrait generation model The business behavior description vector of the selected user generated by the module adjusts the temporary user portrait to obtain the final user portrait of the selected user.

本申请实施例中，通过画像生成模型对业务行为日志进行业务行为描述向量挖掘并根据业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像，该重新描绘的过程通过人工智能模型执行，效率高，解放了人力。In the embodiment of the present application, business behavior description vector mining is performed on business behavior logs through the portrait generation model, and temporary user portraits are adjusted according to the business behavior description vectors to obtain end user portraits of selected users. The redrawing process is carried out through artificial intelligence Model execution has high efficiency and liberates manpower.

作为一种可行的实施方式，向量挖掘模块包括多个依次连接的归类分析单元以及多个依次连接的识别优化单元，画像调整单元包括多个依次连接的调整单元，每个调整单元对应一个识别优化单元；通过画像生成模型中的向量挖掘模块对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量，包括：通过多个依次连接的归类分析单元对业务行为日志进行归类分析，并通过多个识别优化单元对归类分析单元生成的结果进行识别优化，得到每个所述识别优化单元关于选定用户的业务行为描述向量；通过画像生成模型中的画像调整单元依据向量挖掘模块产生的选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像，包括：对于画像调整单元中的首个调整单元，通过和调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对临时用户画像进行调整，得到调整单元对应的调整临时用户画像；对于首个调整单元之外的每一调整单元，通过和每个调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对前一个调整单元产生的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像；其中，最末一个调整单元获得的调整临时用户画像作为选定用户的最终用户画像。As a feasible implementation, the vector mining module includes a plurality of sequentially connected classification and analysis units and a plurality of sequentially connected identification optimization units, and the portrait adjustment unit includes a plurality of sequentially connected adjustment units, and each adjustment unit corresponds to a recognition An optimization unit; through the vector mining module in the portrait generation model, the business behavior description vector mining is carried out on the business behavior log, and the business behavior description vector of the selected user is obtained, including: the business behavior log is processed by multiple sequentially connected classification and analysis units Categorizing and analyzing, and identifying and optimizing the results generated by the categorizing and analyzing units through a plurality of identifying and optimizing units to obtain the business behavior description vector of each of the identified and optimized units about the selected user; through the portrait adjustment unit in the portrait generation model According to the business behavior description vector of the selected user generated by the vector mining module, the temporary user portrait is adjusted to obtain the final user portrait of the selected user, including: for the first adjustment unit in the portrait adjustment unit, through the identification corresponding to the adjustment unit The business behavior description vector of the selected user generated by the optimization unit adjusts the temporary user portrait to obtain the adjusted temporary user portrait corresponding to the adjustment unit; for each adjustment unit other than the first adjustment unit, by corresponding to each adjustment unit The business behavior description vector of the selected user generated by the identification and optimization unit of , adjusts the adjusted temporary user portrait generated by the previous adjustment unit, and obtains the adjusted temporary user portrait corresponding to the adjustment unit; among them, the adjusted temporary user portrait obtained by the last adjustment unit The persona serves as the end user persona for the selected user.

本申请实施例中，画像调整单元接收从向量挖掘模块输出的不同维度的业务行为描述向量，然后对临时用户画像进行阶段性地调整，令产生的最终用户画像的各画像维度下的代表标签表征得准确可靠。In the embodiment of the present application, the portrait adjustment unit receives the business behavior description vectors of different dimensions output from the vector mining module, and then adjusts the temporary user portraits in stages, so that the representative labels in each portrait dimension of the generated end user portraits represent be accurate and reliable.

作为一种可行的实施方式，每个调整单元包括业务行为描述向量挖掘单元以及分析单元，对于画像调整单元中的首个调整单元，通过和调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对临时用户画像进行调整，得到调整单元对应的调整临时用户画像，包括：通过调整单元的业务行为描述向量挖掘单元，对临时用户画像和与调整单元对应的识别优化单元产生的选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到调整单元针对选定用户的业务行为完善描述向量；通过调整单元的分析单元依据业务行为完善描述向量对临时用户画像进行调整，得到后一个调整单元进行第二业务行为描述向量挖掘和调整中运用的调整临时用户画像；对于首个调整单元之外的每一调整单元，通过和每一调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对前一个调整单元产生的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像，包括：通过调整单元的业务行为描述向量挖掘单元，对前一个调整单元获得的业务行为完善描述向量和调整临时用户画像、和与调整单元对应的识别优化单元产生的选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到调整单元针对选定用户的业务行为完善描述向量；通过调整单元的分析单元依据业务行为完善描述向量对前一个调整单元获得的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像。As a feasible implementation, each adjustment unit includes a business behavior description vector mining unit and an analysis unit. For the first adjustment unit in the portrait adjustment unit, the business of the selected user generated by the identification optimization unit corresponding to the adjustment unit The behavior description vector adjusts the temporary user portrait to obtain the adjusted temporary user portrait corresponding to the adjustment unit. The business behavior description vector of a given user is used to mine the business behavior description vector, and the adjustment unit obtains the perfect description vector for the business behavior of the selected user; the analysis unit of the adjustment unit adjusts the temporary user portrait according to the business behavior perfect description vector, and obtains the latter The adjustment unit conducts the second business behavior description vector mining and adjusts the temporary user profile used in the adjustment; for each adjustment unit other than the first adjustment unit, the selected user’s profile generated by the identification and optimization unit corresponding to each adjustment unit The business behavior description vector adjusts the adjusted temporary user portrait generated by the previous adjustment unit to obtain the adjusted temporary user portrait corresponding to the adjustment unit, including: through the business behavior description vector mining unit of the adjustment unit, the business behavior obtained by the previous adjustment unit The behavior perfect description vector and the adjusted temporary user portrait, and the business behavior description vector of the selected user generated by the identification optimization unit corresponding to the adjustment unit are used to mine the business behavior description vector, and the business behavior description vector of the selected user by the adjustment unit is obtained; The analysis unit of the adjustment unit adjusts the adjusted temporary user portrait obtained by the previous adjustment unit according to the perfect description vector of business behavior, and obtains the adjusted temporary user portrait corresponding to the adjustment unit.

本申请实施例中，基于业务行为描述向量挖掘单元对业务行为描述向量和临时用户画像，或前一个调整单元产生的调整临时用户画像，进行业务行为描述向量挖掘，得到选定用户的业务行为完善描述向量，如此，调整获得的最终用户画像更加精确可靠。In the embodiment of the present application, based on the business behavior description vector mining unit, the business behavior description vector and the temporary user portrait, or the adjusted temporary user portrait generated by the previous adjustment unit, are used to mine the business behavior description vector to obtain the complete business behavior of the selected user. The description vector, in this way, is adjusted to obtain a more accurate and reliable portrait of the end user.

作为一种可行的实施方式，该方法中，还包括对画像生成模型的调校过程，该调校过程包括如下步骤：获取包含选定用户的业务行为日志样本；通过画像生成模型依据业务行为日志样本和事先生成的临时用户画像，得到选定用户的多个调校用户画像，多个调校用户画像包括画像调整单元中各个调整单元产生的调校用户画像；依据多个调校用户画像确定代价值loss1；基于代价值loss1调节画像生成模型的模型系数。As a feasible implementation, the method also includes an adjustment process for the portrait generation model, and the adjustment process includes the following steps: obtaining a business behavior log sample containing a selected user; using the portrait generation model according to the business behavior log The samples and the pre-generated temporary user portraits are used to obtain multiple calibration user portraits of the selected user. The multiple calibration user portraits include the calibration user portraits generated by each adjustment unit in the portrait adjustment unit; it is determined based on the multiple calibration user portraits Cost value loss1; adjust the model coefficient of the portrait generation model based on the cost value loss1.

本申请上述实施例中，通过依据多个调校用户画像获得代价值loss1，让得到的代价值准确可靠。In the above embodiments of the present application, the cost value loss1 is obtained based on multiple adjusted user portraits, so that the obtained cost value is accurate and reliable.

作为一种可行的实施方式，每个调校用户画像的描绘体系由多个调校画像维度组成，依据多个调校用户画像确定代价值loss1，包括：针对每个调校用户画像，依据调校用户画像的描绘体系中各调校画像维度的代表标签和画像维度样本的代表标签之间的矢量差代价值，画像维度样本为选定用户的实际用户画像的描绘体系中的画像维度；以及，获取调校用户画像每个调校画像维度的规范化代价值，规范化代价值与每个调校画像维度之间的调整内容关联；通过各个调校用户画像对应的矢量差代价值和规范化代价值，得到代价值loss1。As a feasible implementation, the depiction system of each adjustment user portrait is composed of multiple adjustment portrait dimensions, and the cost value loss1 is determined according to the multiple adjustment user portraits, including: for each adjustment user portrait, according to the adjustment Correct the vector difference cost value between the representative labels of each adjusted portrait dimension in the depiction system of user portraits and the representative labels of the portrait dimension samples, and the portrait dimension samples are the portrait dimensions in the depiction system of the actual user portrait of the selected user; and , to obtain the normalized cost value of each adjustment profile dimension of the adjustment user profile, and the normalization cost value is associated with the adjustment content between each adjustment profile dimension; through the vector difference cost value and the normalization cost value corresponding to each adjustment user profile , get the cost value loss1.

本申请实施例中，通过获取矢量差代价值和规范化代价值，可以衡量各个层次的画像修正情形，令获取到的代价值loss1精确可靠。In the embodiment of the present application, by obtaining the vector difference cost value and the normalized cost value, the image correction situation at each level can be measured, so that the obtained cost value loss1 is accurate and reliable.

作为一种可行的实施方式，通过各个调校用户画像对应的矢量差代价值和规范化代价值，获得代价值loss1，包括：针对每个调校用户画像，将调校用户画像对应的矢量差代价值和规范化代价值进行合并，得到调校用户画像的代价值；将各个调校用户画像的代价值求和，得到代价值loss1；其中，各个调校用户画像的规范化代价值相应的权值和预设参考条件反向关联，预设参考条件包括以下条件中的至少一个：获取调校用户画像的调校循环轮数；获取到调校用户画像的调整单元在画像调整单元中的顺次。As a feasible implementation, the cost value loss1 is obtained through the vector difference cost value and the normalized cost value corresponding to each adjustment user portrait, including: for each adjustment user portrait, the vector difference corresponding to the adjustment user portrait is represented by value and the normalized cost value are combined to obtain the cost value of the adjusted user portrait; the cost value of each adjusted user portrait is summed to obtain the cost value loss1; among them, the corresponding weight value and the normalized cost value of each adjusted user portrait The preset reference conditions are reversely correlated, and the preset reference conditions include at least one of the following conditions: obtaining the number of adjustment cycles for adjusting the user portrait; obtaining the sequence of the adjustment unit for adjusting the user portrait in the portrait adjustment unit.

本申请实施例中，采用了变化的规范化方式，使得模型从画像框架上（即画像维度上）开始训练，然后在各个画像维度的具体构成上进行训练，模型的学习是渐进式地，有助于学习的精度和可靠性。In the embodiment of the present application, a changed standardization method is adopted, so that the model starts training from the portrait frame (that is, the portrait dimension), and then trains on the specific composition of each portrait dimension. The learning of the model is gradual, which is helpful for learning accuracy and reliability.

作为一种可行的实施方式，上述方法还包括：通过画像生成模型得到业务行为日志样本对于选定用户的行为向量聚类识别结果，行为向量聚类识别结果依据向量挖掘模块中最末一个识别优化单元产生的选定用户的业务行为描述向量获取；依据行为向量聚类识别结果以及业务行为日志样本关于选定用户的实际行为向量聚类识别结果，获得代价值loss2；基于代价值loss1调节画像生成模型的模型系数，包括：通过代价值loss1和代价值loss2调节画像生成模型的系数。As a feasible implementation, the above method further includes: obtaining the behavior vector clustering recognition result of the business behavior log sample for the selected user through the portrait generation model, and the behavior vector clustering recognition result is optimized according to the last recognition in the vector mining module Acquire the business behavior description vector of the selected user generated by the unit; obtain the cost value loss2 based on the cluster recognition result of the behavior vector and the cluster recognition result of the business behavior log sample on the actual behavior vector of the selected user; adjust the portrait generation based on the cost value loss1 The model coefficient of the model, including: adjusting the coefficient of the portrait generation model through the cost value loss1 and the cost value loss2.

本申请实施例中，依照联合行为向量聚类识别结果相应的代价值loss2，调节画像生成模型的系数，使得更新后的画像生成模型能更准确地挖掘业务行为描述向量。In the embodiment of the present application, the coefficients of the portrait generation model are adjusted according to the corresponding cost value loss2 of the joint behavior vector clustering recognition result, so that the updated portrait generation model can mine business behavior description vectors more accurately.

作为一种可行的实施方式，该方法还包括：通过画像生成模型得到各个行为向量聚类对应的矢量域差值，矢量域差值依据向量挖掘模块中最末一个识别优化单元产生的选定用户的业务行为描述向量获取；依据矢量域差值和业务行为日志样本关于选定用户的实际矢量域差值，得到代价值loss3；基于代价值loss1调节画像生成模型的模型系数，包括：通过代价值loss1、代价值loss2和代价值loss3获得期望代价值；基于期望代价值调节画像生成模型的模型系数。As a feasible implementation, the method further includes: obtaining the vector domain difference corresponding to each behavior vector cluster through the portrait generation model, and the vector domain difference is based on the selected user generated by the last identification and optimization unit in the vector mining module. The business behavior description vector is acquired; the cost value loss3 is obtained according to the vector field difference value and the actual vector field difference value of the business behavior log sample for the selected user; the model coefficient of the portrait generation model is adjusted based on the cost value loss1, including: through the cost value loss1, cost value loss2 and cost value loss3 obtain the expected cost value; adjust the model coefficient of the portrait generation model based on the expected cost value.

本申请实施例中，依照联合矢量域差值对应的代价值loss3调节画像生成模型的模型系数，使得更新后的画像生成模型可以更加精确地挖掘业务行为描述向量。In the embodiment of the present application, the model coefficients of the portrait generation model are adjusted according to the cost value loss3 corresponding to the joint vector domain difference, so that the updated portrait generation model can mine business behavior description vectors more accurately.

作为一种可行的实施方式，通过画像生成模型依据业务行为日志样本和事先生成的临时用户画像，得到选定用户的多个调校用户画像之后，该方法还可以包括：对画像调整单元最末一个调整单元产生的调校用户画像进行整合操作，整合操作包括以下操作中的至少一个：对所述调校用户画像的画像维度进行维度向量统一对所述调校用户画像的画像维度卷积操作；将基于整合操作后的调校用户画像作为后一轮调校的预设临时用户画像，以对画像生成模型进行调校。As a feasible implementation, after the multiple adjusted user portraits of the selected user are obtained through the portrait generation model based on the business behavior log samples and the temporary user portraits generated in advance, the method may also include: finalizing the portrait adjustment unit The adjustment user portrait generated by an adjustment unit performs an integration operation, and the integration operation includes at least one of the following operations: performing a dimension vector on the portrait dimension of the adjustment user portrait and uniformly performing a convolution operation on the portrait dimension of the adjustment user portrait ; Use the adjusted user portrait based on the integration operation as the preset temporary user portrait for the next round of adjustment, so as to adjust the portrait generation model.

本申请实施例中，通过对最末一个调整单元产生的调校用户画像进行整合操作，并将整合操作后的调校用户画像作为后一轮调校的预设临时用户画像，对画像生成模型进行调校，可以自主地描绘出针对选定用户的临时用户画像。In the embodiment of the present application, the adjustment user portrait generated by the last adjustment unit is integrated, and the adjustment user portrait after the integration operation is used as the preset temporary user portrait for the next round of adjustment to generate a model for the portrait. With tuning, temporary user personas for selected users can be drawn autonomously.

作为一种可行的实施方式，该方法还可以包括：当画像生成模型的代价值符合事先设定的标准，将画像调整单元中最末一个调整单元产生的调校用户画像作为临时用户画像。As a feasible implementation, the method may further include: when the cost value of the portrait generation model meets the preset standard, using the adjusted user portrait generated by the last adjustment unit in the portrait adjustment unit as the temporary user portrait.

本申请实施例中，不但可以对画像生成模型进行调校，还可以获取选定用户对应的临时用户画像，帮助后面的调校可以对照选定用户的样本经验，从而增加最终用户画像的精确可靠。In the embodiment of this application, not only can the portrait generation model be adjusted, but also the temporary user portrait corresponding to the selected user can be obtained, which can help subsequent adjustments to compare the sample experience of the selected user, thereby increasing the accuracy and reliability of the end user portrait .

作为一种可行的实施方式，该方法还可以包括用户画像的存储过程：获取最终用户画像和最终用户画像对应的联合分析信息；其中，联合分析信息包括最终用户画像对应的业务行为日志的获取时间和产生业务行为日志的第一选定用户的区域归属特征；确定最终用户画像与对比用户画像集合中的每个对比用户画像的标签匹配度，并通过联合分析信息和每个对比用户画像的联合对比分析信息，得到最终用户画像与每个对比用户画像的完善匹配度；其中，完善匹配度用于从标签指示信息和时域交集情况两方面一起衡量用户画像的匹配度；基于最终用户画像与对比用户画像集合中的每个对比用户画像的标签匹配度和完善匹配度，确定和最终用户画像对应的目标对比用户画像；将最终用户画像更新为新的对比用户画像，然后在对比用户画像集合中构建最终用户画像和预存的目标对比用户画像对应的目标用户画像类型的映射。As a feasible implementation, the method may also include a process of storing user portraits: acquiring end user portraits and joint analysis information corresponding to end user portraits; where the joint analysis information includes the acquisition time of business behavior logs corresponding to end user portraits and the regional attribution characteristics of the first selected user that generates the business behavior log; determine the label matching degree of the end user portrait and each comparison user portrait in the comparison user portrait collection, and jointly analyze the information and the combination of each comparison user portrait Comparing and analyzing information to obtain the perfect matching degree between the end user portrait and each comparison user portrait; among them, the perfect matching degree is used to measure the matching degree of the user portrait from the two aspects of label indication information and time domain intersection; based on the end user portrait and the Compare the label matching degree and perfect matching degree of each comparison user portrait in the user portrait collection, determine the target comparison user portrait corresponding to the end user portrait; update the end user portrait to a new comparison user portrait, and then compare the user portrait collection Construct the mapping between the end user portrait and the target user portrait type corresponding to the pre-stored target comparison user portrait.

本申请实施例中，通过事先构建对比用户画像集合，在对拟进行保存的最终用户画像进行保存时，和对比用户画像集合中的对比用户画像进行对比，如果确定出和拟存储的用户画像相匹配的对比用户画像，则将用户画像保存在对比用户画像集合中，同时构建映射关系方便取用，实现了用户画像的按类保存。In the embodiment of this application, by constructing a comparison user portrait set in advance, when saving the end user portrait to be saved, compare it with the comparison user portrait in the comparison user portrait collection, if it is determined that it is consistent with the user portrait to be stored For matching comparison user portraits, the user portraits are stored in the comparison user portrait collection, and at the same time, the mapping relationship is constructed for easy access, realizing the storage of user portraits by category.

第二方面，本申请实施例提供一种AI系统，包括相互连接的处理器和存储器，存储器存储有计算机程序，当处理器运行该计算机程序时，执行上述第一方面提供的方法。In the second aspect, the embodiment of the present application provides an AI system, including a processor and a memory connected to each other, the memory stores a computer program, and when the processor runs the computer program, it executes the method provided in the first aspect above.

本申请实施例提供的互动营销平台的用户画像识别方法及AI系统，通过对选定用户的业务行为日志进行业务行为描述向量挖掘得到选定用户的业务行为描述向量，再依据选定用户的业务行为描述向量对临时用户画像进行调整，从而得到选定用户的最终用户画像，本发明实施例采用临时用户画像获取选定用户的最终用户画像，在具有模板指示的情况下，使得选定用户的最终用户画像偏离性低，效率和精确性高。The user portrait recognition method and AI system of the interactive marketing platform provided by the embodiment of the present application obtain the business behavior description vector of the selected user by mining the business behavior description vector of the business behavior log of the selected user, and then obtain the business behavior description vector of the selected user according to the business behavior description vector of the selected user. The behavior description vector adjusts the temporary user portrait to obtain the final user portrait of the selected user. In the embodiment of the present invention, the temporary user portrait is used to obtain the final user portrait of the selected user. In the case of template instructions, the selected user's End user portraits have low deviation, high efficiency and accuracy.

在后面的描述中，将部分地陈述其他的特征。在检查后面内容和附图时，本领域的技术人员将部分地发现这些特征，或者可以通过生产或运用了解到这些特征。通过实践或使用后面所述详细示例中列出的方法、工具和组合的各个方面，当前申请中的特征可以被实现和获得。In the following description, other features will be partly stated. These features will be discovered in part by those skilled in the art upon examination of the following content and drawings, or may be learned from production or use. The features of the present application can be implemented and obtained by practicing or using the various aspects of the methods, means and combinations set forth in the detailed examples described below.

附图说明Description of drawings

为了更清楚地说明本申请实施例中的技术方案，下面将对本申请实施例描述中所需要使用的附图作简单地介绍。In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings that need to be used in the description of the embodiments of the present application.

图1是根据本申请的一些实施例所示的一种互动营销平台的用户画像识别方法的流程图。Fig. 1 is a flow chart of a user portrait recognition method of an interactive marketing platform according to some embodiments of the present application.

图2是本申请实施例提供的用户画像识别装置的架构示意图。Fig. 2 is a schematic structural diagram of a user portrait recognition device provided by an embodiment of the present application.

图3是根据本申请的一些实施例所示的一种AI系统中硬件和软件组成的示意图。Fig. 3 is a schematic diagram of hardware and software components in an AI system according to some embodiments of the present application.

具体实施方式detailed description

下面结合本申请实施例中的附图对本申请实施例进行描述。本申请实施例的实施方式部分使用的术语仅用于对本申请的具体实施例进行解释，而非旨在限定本申请。Embodiments of the present application are described below with reference to the drawings in the embodiments of the present application. The terms used in the implementation of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

本申请实施例中互动营销平台的用户画像识别方法的执行主体为互动营销平台，互动营销平台中包含与业务终端设备通信连接的电子设备，电子设备可以包括但不限于单个网络服务器、多个网络服务器组成的服务器组或于云计算的由大量计算机或网络服务器构成的云，其中，云计算是分布式计算的一种，由一群松散耦合的计算机集组成的一个超级虚拟计算机。其中，电子设备可单独运行来实现本申请，也可接入网络并通过与网络中的其他电子设备的交互操作来实现本申请。其中，电子设备所处的网络包括但不限于互联网、广域网、城域网、局域网、VPN网络等。其中，业务终端设备可以是包括但不限于电脑、智能手机、PAD等，用户通过业务终端设备进行操作，产生业务行为数据并形成业务行为日志，业务行为日志的具体内容不做限定。In the embodiment of the application, the user portrait recognition method of the interactive marketing platform is executed by the interactive marketing platform. The interactive marketing platform includes electronic devices that communicate with business terminal devices. The electronic devices may include, but are not limited to, a single network server, multiple network A server group composed of servers or a cloud composed of a large number of computers or network servers in cloud computing, where cloud computing is a type of distributed computing, a super virtual computer composed of a group of loosely coupled computer sets. Wherein, the electronic device can operate independently to realize the present application, and can also access the network and realize the present application through interactive operation with other electronic devices in the network. Wherein, the network where the electronic device is located includes but is not limited to the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, and the like. Among them, business terminal equipment may include but not limited to computers, smart phones, PADs, etc. Users operate through business terminal equipment to generate business behavior data and form business behavior logs. The specific content of business behavior logs is not limited.

本申请实施例提供了一种互动营销平台的用户画像识别方法，该方法应用于电子设备，如图1所示，该方法包括：The embodiment of the present application provides a user portrait recognition method for an interactive marketing platform, which is applied to an electronic device, as shown in Figure 1, and the method includes:

步骤10：获取业务终端设备发送的选定用户的业务行为日志。Step 10: Obtain the service behavior log of the selected user sent by the service terminal device.

本申请实施例中，业务行为日志是业务终端设备收集的选定用户的交互行为数据构成的集合，例如对于电商场景而言，业务行为日志是选定用户在电商平台的相关行为数据的集合，例如商品浏览、链接点击、商品购买、商品评价、交互会话等，容易理解，业务行为日志中的业务行为数据可以是针对选定用户在预设的采集周期中获取的数据，该预设的周期可以根据实际情况而设定，本申请实施例对此不做限定。其中，选定用户即需要进行针对性用户画像描绘的用户。另外，对于其他应用场景，业务行为日志可以是其他业务行为数据的集合，例如，对于短视频平台用户，业务行为日志可以是用户浏览视频的浏览记录、互动记录和交易记录等，此处不再对其他应用场景及其对应的行为数据进行举例说明。In this embodiment of the application, the business behavior log is a collection of interactive behavior data of selected users collected by business terminal devices. Collections, such as product browsing, link clicks, product purchases, product evaluations, interactive sessions, etc., are easy to understand. The business behavior data in the business behavior log can be the data obtained in the preset collection cycle for selected users. The period of can be set according to the actual situation, which is not limited in this embodiment of the present application. Among them, the selected user is the user who needs to be described with a targeted user portrait. In addition, for other application scenarios, the business behavior log can be a collection of other business behavior data. For example, for short video platform users, the business behavior log can be the browsing record, interaction record, and transaction record of the user’s browsing video, which will not be discussed here. Give examples of other application scenarios and their corresponding behavioral data.

本申请实施例中，用户画像为用以描述用户需求的工具，其表现形式为一系列画像维度下的标签的集合，通过不同画像维度及对应的标签信息量化用户的特征属性，抽象出用户的需求信息。其中，关于画像维度，可以根据具体需要进行配置，举例而言，对于电商场景，画像维度可以包括用户个体属性和行为属性，其中，个体属性可以细化为年龄、性别、地域、受教育程度、职业、收入、生活习惯、消费习惯等，行为属性可以细化为商品种类、活跃频率、商品倾向、产品驱动、使用习惯、产品消费等。以上仅进行了电商场景的用户画像举例，容易理解，对于其他应用场景，可以根据产品或服务的推广中心进行适应性地维度设计本申请对此不做限定。In the embodiment of this application, the user portrait is a tool used to describe the user's needs, and its expression form is a collection of tags under a series of portrait dimensions. The user's characteristic attributes are quantified through different portrait dimensions and corresponding label information, and the user's profile is abstracted. demand information. Among them, the portrait dimension can be configured according to specific needs. For example, for e-commerce scenarios, the portrait dimension can include user individual attributes and behavior attributes, where individual attributes can be refined into age, gender, region, and education level , Occupation, income, living habits, consumption habits, etc. Behavioral attributes can be refined into commodity types, activity frequency, commodity tendencies, product drivers, usage habits, product consumption, etc. The above is only an example of user portraits in e-commerce scenarios, which is easy to understand. For other application scenarios, adaptive dimension design can be carried out according to the promotion center of the product or service. This application does not limit this.

步骤20：对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量。Step 20: Mining the business behavior description vectors from the business behavior logs to obtain the business behavior description vectors of the selected users.

本申请实施例中，对业务行为日志进行业务行为描述向量挖掘，可以采用事先调校好的画像生成模型对业务行为日志进行业务行为描述向量挖掘，业务行为描述向量是能够体现业务行为的特征的矢量结果。In the embodiment of this application, the business behavior description vector mining is performed on the business behavior log, and the pre-adjusted portrait generation model can be used to mine the business behavior description vector on the business behavior log, and the business behavior description vector can reflect the characteristics of the business behavior Vector results.

步骤30：依据选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像。Step 30: Adjust the temporary user portrait according to the business behavior description vector of the selected user to obtain the final user portrait of the selected user.

本申请实施例中，将选定用户的业务行为描述向量作为对临时用户画像的调整因子，依据该调整因子对临时用户画像进行调整。其中，临时用户画像可以是针对选定用户而事先生成的，临时用户画像可以认为是一个用户画像模板，其不代表选定用户的真实用户画像，经过选定用户的业务行为描述向量对临时用户画像进行调整后，使之成为选定用户的真实用户画像。依据选定用户的业务行为描述向量对临时用户画像进行调整，具体可以为依据选定用户的业务行为描述向量对目标临时用户画像进行调整，目标临时用户画像是针对选定用户事先生成的。In the embodiment of the present application, the business behavior description vector of the selected user is used as an adjustment factor for the temporary user portrait, and the temporary user portrait is adjusted according to the adjustment factor. Among them, the temporary user portrait can be generated in advance for the selected user, and the temporary user portrait can be considered as a user portrait template, which does not represent the real user portrait of the selected user. The profile is adjusted so that it becomes a real user profile of the selected user. The temporary user portrait is adjusted according to the business behavior description vector of the selected user, specifically, the target temporary user portrait is adjusted according to the business behavior description vector of the selected user, and the target temporary user portrait is generated in advance for the selected user.

以上步骤10-30通过对选定用户的业务行为日志进行业务行为描述向量挖掘得到选定用户的业务行为描述向量，再依据选定用户的业务行为描述向量对临时用户画像进行调整，从而得到选定用户的最终用户画像，其中，通过临时用户画像获取选定用户的最终用户画像，在具有模板指示的情况下，使得选定用户的最终用户画像偏离性低，精确性高。In the above steps 10-30, the business behavior description vector of the selected user is obtained by mining the business behavior description vector of the business behavior log of the selected user, and then the temporary user portrait is adjusted according to the business behavior description vector of the selected user, so as to obtain the selected user's business behavior description vector. The final user portrait of the selected user, wherein the temporary user portrait is used to obtain the final user portrait of the selected user, and in the case of template instructions, the deviation of the final user portrait of the selected user is low and the accuracy is high.

作为一种实施方式，步骤20具体可以包括以下步骤：As an implementation manner, step 20 may specifically include the following steps:

通过画像生成模型中的向量挖掘模块对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量。步骤30具体包括以下步骤：通过画像生成模型中的画像调整单元依据向量挖掘模块产生的选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像。通过画像生成模型对业务行为日志进行业务行为描述向量挖掘并通过选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像，该重新描绘的过程通过人工智能模型执行，效率高，解放了人力。Through the vector mining module in the portrait generation model, the business behavior description vector is mined from the business behavior log to obtain the business behavior description vector of the selected user.Step 30 specifically includes the following steps: using the portrait adjustment unit in the portrait generation model to adjust the temporary user portrait according to the business behavior description vector of the selected user generated by the vector mining module, to obtain the final user portrait of the selected user. Use the portrait generation model to mine the business behavior description vectors from the business behavior logs and adjust the temporary user portraits through the business behavior description vectors of the selected users to obtain the final user portraits of the selected users. , high efficiency, liberated manpower.

作为一种实施方式，向量挖掘模块包括多个依次连接的归类分析单元以及多个依次连接的识别优化单元。画像调整单元包括多个依次连接的调整单元，其中，每个调整单元对应一个识别优化单元。As an implementation, the vector mining module includes a plurality of sequentially connected classification and analysis units and a plurality of sequentially connected identification optimization units. The portrait adjustment unit includes a plurality of sequentially connected adjustment units, wherein each adjustment unit corresponds to a recognition optimization unit.

作为一种实施方式，通过画像生成模型中的向量挖掘模块对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量的过程可以包括：通过多个依次连接的归类分析单元对业务行为日志进行归类分析，并通过多个识别优化单元对归类分析单元生成的结果进行识别优化，得到每个识别优化单元对于选定用户的业务行为描述向量。其中，向量挖掘模块可以是具备残差连接的UNet或SegNet构造。归类分析单元是一种encoder，具体可以包括卷积层、池化层和BN层，卷积层获取业务行为日志中数据的特征，池化层对数据进行下采样，然后经由BN层进行归一化，完成归类分析。对应的，识别优化单元是一种decoder，用于将归类分析得到的特征信息进行识别优化，先对特征信息进行上采样，再进行卷积操作，完善特征信息，填补归类分析单元的池化过程形成的信息缺失。其中，归类分析单元和识别优化单元中的卷积层的算法相同。As an implementation, the process of mining business behavior description vectors from business behavior logs through the vector mining module in the portrait generation model to obtain the business behavior description vectors of selected users may include: through multiple sequentially connected classification and analysis units The business behavior logs are classified and analyzed, and the results generated by the classification and analysis units are identified and optimized by multiple identification and optimization units to obtain the business behavior description vector of each identification and optimization unit for the selected user. Among them, the vector mining module can be a UNet or SegNet structure with residual connections. The classification analysis unit is a kind of encoder, which can specifically include a convolutional layer, a pooling layer, and a BN layer. The convolutional layer obtains the characteristics of the data in the business behavior log, and the pooling layer down-samples the data, and then performs classification via the BN layer. One to complete the classification analysis. Correspondingly, the identification optimization unit is a decoder, which is used to identify and optimize the feature information obtained by the classification analysis. First, the feature information is up-sampled, and then the convolution operation is performed to improve the feature information and fill the pool of the classification analysis unit. The information formed by the chemicalization process is missing. Wherein, the algorithm of the convolutional layer in the classification analysis unit and the identification optimization unit is the same.

通过画像生成模型中的画像调整单元依据向量挖掘模块产生的选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像的过程可以包括：Through the portrait adjustment unit in the portrait generation model to adjust the temporary user portrait according to the business behavior description vector of the selected user generated by the vector mining module, the process of obtaining the final user portrait of the selected user may include:

针对画像调整单元中的首个调整单元，通过和调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对临时用户画像进行调整，得到调整单元对应的调整临时用户画像。For the first adjustment unit in the portrait adjustment unit, the temporary user portrait is adjusted through the business behavior description vector of the selected user generated by the identification optimization unit corresponding to the adjustment unit, and the adjusted temporary user portrait corresponding to the adjustment unit is obtained.

对于画像调整单元中首个调整单元之外的每一调整单元，通过和每一调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对前一个调整单元产生的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像。其中，最末一个调整单元获得的调整临时用户画像作为选定用户的最终用户画像。For each adjustment unit other than the first adjustment unit in the portrait adjustment unit, the business behavior description vector of the selected user generated by the identification optimization unit corresponding to each adjustment unit is used to adjust the temporary user portrait generated by the previous adjustment unit Adjustment is performed to obtain an adjusted temporary user portrait corresponding to the adjustment unit. Among them, the adjusted temporary user portrait obtained by the last adjustment unit is used as the final user portrait of the selected user.

画像调整单元接收从向量挖掘模块不同维度的业务行为描述向量，然后对临时用户画像进行阶段性地调整，令产生的最终用户画像的各画像维度下的代表标签表征得更加准确可靠。The portrait adjustment unit receives the business behavior description vectors of different dimensions from the vector mining module, and then adjusts the temporary user portraits in stages, so that the representation of the representative labels in each portrait dimension of the generated end user portrait is more accurate and reliable.

具体而言，各个调整单元包括业务行为描述向量挖掘单元和分析单元。通过调整单元的业务行为描述向量挖掘单元，对临时用户画像和与调整单元对应的识别优化单元产生的选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到调整单元针对选定用户的业务行为完善描述向量。同时，通过调整单元的分析单元依据业务行为完善描述向量对前一个调整单元获得的调整临时用户画像进行调整，得到该调整单元对应的调整临时用户画像。上述向量挖掘模块和画像调整单元均为卷积网络。Specifically, each adjustment unit includes a business behavior description vector mining unit and an analysis unit. Through the business behavior description vector mining unit of the adjustment unit, the temporary user portrait and the business behavior description vector of the selected user generated by the identification and optimization unit corresponding to the adjustment unit are mined to obtain the business behavior description vector of the adjustment unit for the selected user. Behavior perfection description vector. At the same time, the analysis unit of the adjustment unit adjusts the adjusted temporary user portrait obtained by the previous adjustment unit according to the perfect description vector of business behavior to obtain the adjusted temporary user portrait corresponding to the adjustment unit. The above-mentioned vector mining module and portrait adjustment unit are both convolutional networks.

作为一种实施范式，首个归类分析单元包括一个或多个卷积层，通过首个归类分析单元得到多通道业务描述向量，不是首个的归类分析单元各包含一个或多个卷积层合一个maxpool层。其中，不是最末一个的识别优化单元包含一个或多个卷积层，一个反卷积层和一个Self-Attention层。其中，首个归类分析单元表示没有和其连接的前一个归类分析单元的归类分析单元，而最末一个识别优化单元表示没有和其连接的后一个归类分析单元的归类分析单元。将N通道业务行为描述向量加载到首个识别优化单元的Self-Attention层，再把Self-Attention层生成的结果加载至反卷积层，最后将获得的结果加载至卷积层，根据在识别优化单元中加入Self-Attention层，可以帮助模型突出再次生成的业务行为描述向量，同时阻抑无关业务行为描述向量，此处所提及的模型突出再次生成的业务行为描述向量是对业务行为日志进行业务行为描述向量挖掘获得的选定用户的业务行为描述向量，无关业务行为描述向量是对业务行为日志进行业务行为描述向量挖掘获得的与选定用户没有关联的业务行为描述向量。As an implementation paradigm, the first classification analysis unit includes one or more convolutional layers, and the multi-channel service description vector is obtained through the first classification analysis unit, and the classification analysis units that are not the first include one or more volumes The stack is combined with a maxpool layer. Among them, the recognition optimization unit that is not the last one includes one or more convolutional layers, a deconvolutional layer and a Self-Attention layer. Among them, the first classification analysis unit represents the classification analysis unit without the previous classification analysis unit connected with it, and the last recognition optimization unit represents the classification analysis unit without the subsequent classification analysis unit connected with it . Load the N-channel business behavior description vector to the Self-Attention layer of the first identification optimization unit, then load the result generated by the Self-Attention layer to the deconvolution layer, and finally load the obtained result to the convolution layer, according to the recognition Adding a Self-Attention layer to the optimization unit can help the model highlight the regenerated business behavior description vectors, while suppressing irrelevant business behavior description vectors. The model mentioned here highlights the regenerated business behavior description vectors is an important part of the business behavior log The business behavior description vector of the selected user is obtained by mining the business behavior description vector, and the irrelevant business behavior description vector is the business behavior description vector not related to the selected user obtained by mining the business behavior description vector from the business behavior log.

画像调整单元可以包括N个调整单元，，可以在顶层特征维度上设置多个调整单元，即多个调整单元和向量挖掘模块中最末一个识别优化单元相匹配，从而能够将信息完全利用。画像调整单元接收源于向量挖掘模块不同维度的业务行为描述向量，再对临时用户画像进行阶段性地调整，令产生的最终用户画像的各画像维度下的代表标签表征得准确可靠。The portrait adjustment unit can include N adjustment units, and multiple adjustment units can be set on the top-level feature dimension, that is, multiple adjustment units match the last recognition optimization unit in the vector mining module, so that information can be fully utilized. The portrait adjustment unit receives the business behavior description vectors from different dimensions of the vector mining module, and then adjusts the temporary user portraits in stages, so that the representative labels in each portrait dimension of the generated end user portraits can be represented accurately and reliably.

本申请实施例中，首个调整单元的业务行为描述向量挖掘单元包括pooling层、加强层和分析层，对于首个调整单元中pooling层，其输入为对应识别优化单元的Self-Attention层生成的结果和临时用户画像，pooling层对Self-Attention层生成的结果和临时用户画像采样，得到选定用户的业务行为描述向量的采样结果，加强层通过临时用户画像对采样结果进行增强，在此基础上，分析层再进行调整分析，得到选定用户的业务行为完善描述向量。选定用户的该业务行为完善描述向量为对临时用户画像调整时的调整系数。分析单元将选定用户的业务行为完善描述向量和临时用户画像融合后即获得调整单元对应的调整临时用户画像。In the embodiment of the present application, the business behavior description vector mining unit of the first adjustment unit includes a pooling layer, a reinforcement layer and an analysis layer. For the pooling layer in the first adjustment unit, its input is generated by the Self-Attention layer of the corresponding identification optimization unit. Results and temporary user portraits, the pooling layer samples the results generated by the Self-Attention layer and temporary user portraits, and obtains the sampling results of the business behavior description vectors of selected users, and the enhancement layer enhances the sampling results through temporary user portraits. On the top, the analysis layer performs adjustment and analysis to obtain the perfect description vector of the selected user's business behavior. The perfect description vector of the business behavior of the selected user is an adjustment coefficient when adjusting the temporary user portrait. The analysis unit fuses the business behavior perfect description vector of the selected user with the temporary user portrait to obtain the adjusted temporary user portrait corresponding to the adjustment unit.

除首个调整单元外的调整单元的业务行为描述向量挖掘单元中，pooling层的输入为前一个调整单元产生的调整临时用户画像和对应识别优化单元的Self-Attention层生成的结果。除首个调整单元外的调整单元中业务行为描述向量挖掘单元的pooling层通过前一个调整单元产生的调整临时用户画像和对应识别优化单元的Self-Attention层生成的结果进行采样，得到选定用户的业务行为描述向量的采样结果，加强层通过前一个调整单元中加强层输出和前一个调整单元产生的调整临时用户画像，对该pooling层的输出结果进行加强，然后，该调整单元的分析层依据该加强层的输出得到选定用户的业务行为完善描述向量。分析单元将本个调整单元的分析层产生的选定用户的业务行为完善描述向量和前一个调整单元产生的调整临时用户画像融合后获得本个调整单元对应的调整临时用户画像。将最末一个调整单元产生的调整临时用户画像确定为选定用户的最终用户画像。首个调整单元即没有和其相连的前一个调整单元的调整单元。最末一个调整单元即没有和其相连的后一个调整单元的调整单元。通过业务行为描述向量挖掘单元对业务行为描述向量和临时用户画像或前一个调整单元产生的调整临时用户画像进行业务行为描述向量挖掘，得到选定用户的业务行为完善描述向量，从而令调整获得的最终用户画像准确可靠。In the business behavior description vector mining unit of the adjustment unit other than the first adjustment unit, the input of the pooling layer is the adjusted temporary user portrait generated by the previous adjustment unit and the result generated by the Self-Attention layer of the corresponding recognition optimization unit. The pooling layer of the business behavior description vector mining unit in the adjustment unit other than the first adjustment unit samples the adjusted temporary user portrait generated by the previous adjustment unit and the result generated by the Self-Attention layer of the corresponding identification optimization unit to obtain the selected user The sampling result of the business behavior description vector, the enhancement layer strengthens the output result of the pooling layer through the output of the enhancement layer in the previous adjustment unit and the adjusted temporary user portrait generated by the previous adjustment unit, and then, the analysis layer of the adjustment unit According to the output of the strengthening layer, the business behavior perfect description vector of the selected user is obtained. The analysis unit fuses the perfect description vector of the selected user's business behavior generated by the analysis layer of this adjustment unit with the adjusted temporary user portrait generated by the previous adjustment unit to obtain the adjusted temporary user portrait corresponding to this adjustment unit. The adjusted temporary user portrait generated by the last adjustment unit is determined as the final user portrait of the selected user. The first adjustment unit is the adjustment unit that has no previous adjustment unit connected to it. The last adjustment unit is the adjustment unit that has no subsequent adjustment unit connected to it. Through the business behavior description vector mining unit, the business behavior description vector and the temporary user portrait or the adjusted temporary user portrait generated by the previous adjustment unit are used to mine the business behavior description vector, and the business behavior perfect description vector of the selected user is obtained, so that the adjustment obtained End user portraits are accurate and reliable.

作为一些实施方式，向量挖掘模块中最末一个识别优化单元后，可以连接聚类识别层以及矢量域差值获取层。通过聚类识别层获取该业务行为日志对于选定用户的行为向量聚类识别结果，以及通过矢量域差值获取层获取业务行为日志中各个行为向量聚类结果中的行为向量和行为向量聚类结果的中心行为向量（即具备代表性的行为向量）间的矢量差值。As some implementation manners, after the last identification optimization unit in the vector mining module, the cluster identification layer and the vector domain difference acquisition layer may be connected. Obtain the behavior vector clustering identification result of the business behavior log for the selected user through the clustering identification layer, and obtain the behavior vector and behavior vector clustering in each behavior vector clustering result in the business behavior log through the vector domain difference acquisition layer The vectorial difference between the center action vectors of the results (i.e. the representative action vectors).

作为一种实施方式，画像生成模型中的画像调整单元为多个，多个画像调整单元与向量挖掘模块连接，不同的画像调整单元收集业务行为描述向量对选定用户中不同的层次进行用户画像修正。将业务行为日志加载到画像生成模型，画像生成模型中的各画像调整单元输出选定用户各个层次维度对应的用户画像，将各层次维度的最终用户画像合并，得到选定用户的最终用户画像。As an implementation, there are multiple portrait adjustment units in the portrait generation model, and the multiple portrait adjustment units are connected to the vector mining module. Different portrait adjustment units collect business behavior description vectors to perform user portraits on different levels of selected users. fix. The business behavior log is loaded into the portrait generation model, and each portrait adjustment unit in the portrait generation model outputs the user portrait corresponding to each dimension of the selected user, and merges the end user portraits of each dimension to obtain the end user portrait of the selected user.

本申请实施例提供的互动营销平台的用户画像识别方法还包括对画像生成模型的调校过程，包括以下步骤：The user portrait recognition method of the interactive marketing platform provided in the embodiment of the present application also includes a process of adjusting the portrait generation model, including the following steps:

步骤100：获取选定用户的业务行为日志样本。Step 100: Acquiring business behavior log samples of selected users.

步骤200：通过画像生成模型依据业务行为日志样本和事先生成的临时用户画像，得到选定用户的多个调校用户画像，多个调校用户画像包括画像调整单元中各个调整单元产生的调校用户画像。Step 200: Obtain multiple adjusted user portraits of the selected user based on the business behavior log samples and pre-generated temporary user portraits through the portrait generation model. The multiple adjusted user portraits include the adjustments generated by each adjustment unit in the portrait adjustment unit User portrait.

步骤300：依据多个调校用户画像确定代价值loss1。Step 300: Determine the cost value loss1 according to multiple adjusted user profiles.

每个调校用户画像的描绘体系由多个调校画像维度组成。其中，调校画像维度可以参照前述关于用户画像维度的介绍。The description system of each adjustment user portrait consists of multiple adjustment portrait dimensions. Wherein, to adjust the portrait dimension, you can refer to the aforementioned introduction about the user portrait dimension.

作为一种实施方式，获取代价值loss1的过程可以包括：针对每个调校用户画像，依据调校用户画像的描绘体系中各调校画像维度的代表标签与画像维度样本的代表标签之间的矢量差代价值。画像维度样本为选定用户的实际用户画像的描绘体系中的画像维度。可选的，业务行为日志样本可以注解各行为向量聚类识别的实际结果，通过聚类得到的某一向量集合构成用户画像的一画像维度下的标签集合，其中，矢量结果突出的某一标签可以作为代表标签。As an implementation, the process of obtaining the cost value loss1 may include: for each adjusted user portrait, according to the relationship between the representative labels of each adjusted portrait dimension and the representative labels of the portrait dimension samples in the description system of the adjusted user portrait Vector difference cost value. The portrait dimension sample is the portrait dimension in the depiction system of the actual user portrait of the selected user. Optionally, business behavior log samples can annotate the actual results of each behavior vector clustering recognition, and a certain vector set obtained through clustering constitutes a label set of user portraits in one portrait dimension. Among them, a certain label with prominent vector results Can be used as a representative label.

以及，获取调校用户画像上各调校画像维度的规范化代价值。规范化代价值与各调校画像维度之间的调整内容关联。通过获取规范化代价，为代价函数赋予限制，规范接下来的循环迭代，防止模型自我膨胀和过拟合，对代价函数增加一个限制条件，限制较高次的参数值过大，作为一些实施方式，规范化代价值可以包括以下内容：标签数量代价、标签长度代价和标签方向代价，其中，标签数量代价和维度中标签的数量关联，用于规范超出预设数量范围区间的标签数量，保证调校用户画像中画像维度的标签数量保持在合理范围；标签长度代价用于规范调校画像维度中标签长度超过预设长度的标签长度，使得画像维度中的标签长度具有归一性；因为同一用户的同一画像维度中的画像标签表征的倾向性往往是一致的，不会有大的偏差，标签方向代价用于规范画像维度中标签的矢量方向偏离目标方向的标签方向，保证画像维度的统一性，该目标方向例如是同一画像维度中所有标签的方向均值。相应的权值作为一种实施方式，可以为标签数量代价、标签长度代价和标签方向代价分别赋予相应的权值，然后加权求和获得上述规范化代价值。再通过各调校用户画像对应的矢量差代价值和规范化代价值，得到代价值loss1。具体而言，针对每个调校用户画像，将调校用户画像对应的矢量差代价值和规范化代价值进行合并，得到该调校用户画像的代价值。再将各调校用户画像的代价值求和，得到代价值loss1。其中，每个调校用户画像的规范化代价值相应的权值与预设参考条件反向关联。预设参考条件包括以下条件中的至少一个：获取调校用户画像的调校循环轮数；获取到调校用户画像的调整单元在画像调整单元中的顺次。调校循环轮数的获取可以是将代价值loss1调节画像生成模型中的数值加上一得到的数值。规范化项的权值不能过大或过小，多小则画像维度不均衡，过大，则调校用户画像的细节缺失，本申请实施例中采用了变化的规范化方式，使得模型从画像框架上（即画像维度上）开始训练，然后在各个画像维度的具体构成上进行训练，模型的学习是渐进式地，有助于学习的精度和可靠性。And, the normalized cost value of each adjustment profile dimension on the adjustment user profile is obtained. The adjustment content association between the normalized cost value and each adjustment profile dimension. By obtaining the normalized cost, assign restrictions to the cost function, standardize the next loop iterations, prevent model self-expansion and overfitting, add a restriction to the cost function, and limit the value of higher-order parameters from being too large. As some implementation methods, The normalized cost value can include the following contents: label quantity cost, label length cost and label direction cost. Among them, the label quantity cost is associated with the quantity of labels in a dimension, and is used to standardize the quantity of labels beyond the preset quantity range to ensure that the adjustment user The number of tags in the portrait dimension in the portrait is kept within a reasonable range; the tag length cost is used to standardize and adjust the tag length in the portrait dimension that exceeds the preset length, so that the tag length in the portrait dimension is normalized; because the same user's same The tendency of the portrait label representation in the portrait dimension is often consistent, and there will be no large deviation. The label direction cost is used to standardize the label direction in which the vector direction of the label in the portrait dimension deviates from the target direction, so as to ensure the uniformity of the portrait dimension. The target direction is, for example, the mean value of the directions of all labels in the same profile dimension. Corresponding weights As an implementation manner, corresponding weights can be assigned to the tag quantity cost, tag length cost, and tag direction cost, and then weighted and summed to obtain the above normalized cost value. Then, the cost value loss1 is obtained by adjusting the vector difference cost value and the normalized cost value corresponding to each user profile. Specifically, for each adjusted user profile, the vector difference cost value corresponding to the adjusted user profile and the normalized cost value are combined to obtain the cost value of the adjusted user profile. Then sum the cost values of each adjusted user profile to obtain the cost value loss1. Wherein, the weight corresponding to the normalized cost value of each adjusted user portrait is inversely associated with the preset reference condition. The preset reference conditions include at least one of the following conditions: obtaining the number of adjustment cycles for adjusting the user portrait; obtaining the order of the adjustment unit for adjusting the user portrait in the portrait adjustment unit. The acquisition of the number of rounds of adjustment cycles may be a value obtained by adding one to the value in the adjustment image generation model of the cost value loss1. The weight of the normalization item cannot be too large or too small. If it is too small, the dimensions of the portrait will be unbalanced. If it is too large, the details of the adjustment user portrait will be missing. (that is, in the portrait dimension) to start training, and then train on the specific composition of each portrait dimension. The learning of the model is gradual, which is conducive to the accuracy and reliability of learning.

通过获取矢量差代价值和规范化代价值，可以衡量多方面的修正情形，令获取到的代价值loss1精确度高。By obtaining the vector difference cost value and the normalized cost value, various correction situations can be measured, so that the obtained cost value loss1 is highly accurate.

作为一种实施方式，互动营销平台的用户画像识别方法中画像生成模型的调校过程还可包括：As an implementation, the adjustment process of the portrait generation model in the user portrait recognition method of the interactive marketing platform may also include:

通过画像生成模型获得业务行为日志样本中选定用户的行为向量聚类识别结果。行为向量聚类识别结果依据向量挖掘模块中最末一个识别优化单元产生的选定用户的业务行为描述向量得到，行为向量聚类识别结果是对挖掘得到的用户行为描述向量进行聚类后得到的结果，容易理解，该结果的个数可以为多个。之后依据行为向量聚类识别结果和业务行为日志样本关于选定用户的实际行为向量聚类识别结果，得到代价值loss2。画像生成模型中向量挖掘模块中最末一个识别优化单元后，还可耦合聚类识别层，通过聚类识别层可以获取业务行为日志关于选定用户的行为向量聚类识别结果。代价值loss2涉及到的代价类型不做限定，例如交叉熵代价。联合行为向量聚类识别结果相应的代价值loss2调节画像生成模型的模型系数，使得模型能更准确地挖掘用户行为描述向量。The behavior vector clustering recognition results of selected users in the business behavior log samples are obtained through the portrait generation model. The behavior vector clustering recognition result is obtained based on the business behavior description vector of selected users generated by the last recognition optimization unit in the vector mining module, and the behavior vector clustering recognition result is obtained after clustering the user behavior description vectors obtained by mining As a result, it is easy to understand that there may be multiple results. Then, according to the behavior vector clustering recognition result and the business behavior log sample's actual behavior vector clustering recognition result of the selected user, the cost value loss2 is obtained. After the last identification and optimization unit in the vector mining module in the portrait generation model, the clustering identification layer can also be coupled, through which the behavior vector clustering identification results of the selected users in the business behavior log can be obtained. The type of cost involved in the cost value loss2 is not limited, such as cross-entropy cost. The corresponding cost value loss2 of the joint behavior vector clustering recognition results adjusts the model coefficients of the portrait generation model, so that the model can mine user behavior description vectors more accurately.

作为一种实施方式，画像生成模型的调校过程还可包括：As an implementation, the adjustment process of the portrait generation model may also include:

通过画像生成模型获取各行为向量聚类对应的矢量域差值。画像生成模型中向量挖掘模块中最末一个识别优化单元后还耦合矢量域差值获取层，通过矢量域差值获取层可以获取业务行为日志中各个行为向量聚类结果中的行为向量和行为向量聚类结果的中心行为向量间的矢量差值。再依据矢量域差值和业务行为日志样本对于选定用户的实际矢量域差值，获得代价值loss3。例如，代价值loss3为L1代价，联合矢量域差值对应的代价值loss3调节画像生成模型的模型系数，使得模型能更准确地挖掘用户行为描述向量。The vector field difference corresponding to each behavior vector cluster is obtained through the portrait generation model. After the last identification and optimization unit in the vector mining module in the portrait generation model, the vector domain difference acquisition layer is also coupled, through which the vector domain difference acquisition layer can obtain the behavior vector and behavior vector in the clustering results of each behavior vector in the business behavior log The vector difference between the central behavior vectors of the clustering results. Then, the cost value loss3 is obtained according to the vector field difference value and the actual vector field difference value of the business behavior log sample for the selected user. For example, the cost value loss3 is the L1 cost, and the cost value loss3 corresponding to the joint vector field difference adjusts the model coefficients of the portrait generation model, so that the model can mine user behavior description vectors more accurately.

步骤400：基于代价值loss1调节画像生成模型的模型系数。Step 400: Adjust the model coefficients of the portrait generation model based on the cost value loss1.

作为一些实施方式，基于代价值loss1调节画像生成模型的模型系数。作为另一些实施方式，通过代价值loss1和代价值loss2调节画像生成模型的模型系数。作为其他的实施方式，通过代价值loss1和代价值loss3调节画像生成模型的模型系数。又或者在其他实施方式中，通过代价值loss1、代价值loss2和代价值loss3调节画像生成模型的模型系数。As some implementation manners, the model coefficients of the portrait generation model are adjusted based on the cost value loss1. As some other implementation manners, the model coefficients of the portrait generation model are adjusted through the cost value loss1 and the cost value loss2. As another implementation manner, the model coefficients of the portrait generation model are adjusted through the cost value loss1 and the cost value loss3. Or in other implementation manners, the model coefficients of the portrait generation model are adjusted through the cost value loss1 , the cost value loss2 and the cost value loss3 .

例如，通过代价值loss1、代价值loss2和代价值loss3，得到期望代价值，再基于期望代价值调节画像生成模型的模型系数。其中，期望代价值的计算方式可以参考以下公式：For example, the expected cost value is obtained through the cost value loss1, the cost value loss2 and the cost value loss3, and then the model coefficient of the portrait generation model is adjusted based on the expected cost value. Among them, the calculation method of expected cost value can refer to the following formula:

D=aA₃+bA₂+A₁；A₁=

；G=dG₁+eG₂+fG₃D=aA₃ +bA₂ +A₁ ; A₁ =

;G=dG₁ +eG₂ +fG₃

D为期望代价值；A₃为代价值loss3；A₂为代价值loss2、A₁为代价值loss1；C为第一个调整单元产生的调校用户画像对应的矢量差代价值；G为第一个调整单元产生的调校用户画像对应的规范化代价值；n为画像调整单元中的调整单元的顺次；a、b、c分别为对应代价值分配的权值；G₁为标签数量代价；G₂为标签长度代价；G₃为标签方向代价，d、e、f为对应代价值分配的权值。D is the expected cost value; A₃ is the cost value loss3; A₂ is the cost value loss2, A₁ is the cost value loss1; C is the vector difference cost value corresponding to the adjusted user portrait generated by the first adjustment unit; G is the cost value of the first adjustment unit The normalized cost value corresponding to the adjusted user portrait generated by an adjustment unit; n is the order of the adjustment units in the portrait adjustment unit; a, b, and c are the weights assigned to the corresponding cost value; G₁ is the label quantity cost ; G₂ is the label length cost; G₃ is the label direction cost, and d, e, f are the weights assigned to the corresponding cost values.

实际应用中，依据样本调校各业务行为日志样本对应的选定用户全部行为的行为向量聚类识别结果得到代价值loss2，依据样本调校各业务行为日志样本对应的选定用户全部行为的矢量域差值得到代价值loss3，依据样本调校各业务行为日志样本对应的选定用户全部行为的调校用户画像得到代价值loss1。基于代价值loss1、代价值loss2和代价值loss3获取期望代价值，再基于期望代价值调节画像生成模型的模型系数。In practical applications, adjust the behavior vectors of all the behaviors of selected users corresponding to each business behavior log sample according to the samples. The cost value loss3 is obtained from the domain difference value, and the cost value loss1 is obtained by adjusting the user profile of all behaviors of the selected users corresponding to each business behavior log sample according to the sample adjustment. The expected cost value is obtained based on the cost value loss1, the cost value loss2 and the cost value loss3, and then the model coefficient of the portrait generation model is adjusted based on the expected cost value.

实际应用中，可能画像生成模型包含多个和向量挖掘模块连接的画像调整单元，此时，逐一按序调校画像调整单元和其中一个向量挖掘模块连接获得的网络。当第一个画像调整单元和向量挖掘模块构建而成的网络对应的代价值符合事先设定的标准时，继续调校下一个画像调整单元和向量挖掘模块构建而成的网络，循环之下直到全部画像调整单元和向量挖掘模块构建而成的网络的代价值符合事先设定的标准，例如代价值满足预设代价值。In practical applications, the portrait generation model may include multiple portrait adjustment units connected to the vector mining module. At this time, the network obtained by connecting the portrait adjustment units to one of the vector mining modules is adjusted one by one. When the cost value corresponding to the network constructed by the first portrait adjustment unit and the vector mining module meets the preset standard, continue to adjust the network constructed by the next portrait adjustment unit and the vector mining module, until all The cost value of the network constructed by the portrait adjustment unit and the vector mining module conforms to a preset standard, for example, the cost value satisfies a preset cost value.

作为一些实施方式，步骤200后还可以包括：对画像调整单元最末一个调整单元产生的调校用户画像进行整合操作。整合操作包括以下操作中的至少一个：对调校用户画像的画像维度进行维度向量统一；对调校用户画像的画像维度卷积操作。维度向量统一化的过程为设置一个参考向量，然后优化画像维度中全部维度向量向该参考向量趋近完成统一。画像维度卷积操作可以是利用现有算法进行卷积操作，例如加一平滑算法。再将基于整合操作后的调校用户画像确定为后一轮调校的预设临时用户画像对画像生成模型进行调校。通过将本轮最末一个调整单元产生的调校用户画像经过整合操作确定为后一轮调校的预设临时用户画像，在调校的过程中反复修正更新，以适应各个选定用户的临时用户画像。As some implementation manners, after step 200, it may further include: performing an integration operation on the adjusted user portrait generated by the last adjustment unit of the portrait adjustment unit. The integration operation includes at least one of the following operations: performing dimension vector unification on the portrait dimension of the adjusted user portrait; convolution operation on the portrait dimension of the adjusted user portrait. The process of dimension vector unification is to set a reference vector, and then optimize all dimension vectors in the image dimension to approach the reference vector to complete the unification. The image dimension convolution operation may be performed by using an existing algorithm, such as an plus-one smoothing algorithm. Then, the adjusted user portrait based on the integration operation is determined as the preset temporary user portrait for the next round of calibration to adjust the portrait generation model. The adjustment user portrait generated by the last adjustment unit of this round is determined as the preset temporary user portrait of the next round of adjustment through integration operation, and it is revised and updated repeatedly during the adjustment process to adapt to the temporary of each selected user User portrait.

本申请实施例中，通过对最末一个调整单元产生的调校用户画像进行整合操作，将整合后的调校用户画像确定为后一轮调校的预设临时用户画像，对画像生成模型进行调校，可以灵活建立匹配各种选定用户的临时用户画像。In the embodiment of the present application, by integrating the adjustment user portrait generated by the last adjustment unit, the integrated adjustment user portrait is determined as the preset temporary user portrait for the next round of adjustment, and the portrait generation model is Tuning can flexibly create temporary user portraits that match various selected users.

作为一种实施方式，当画像生成模型的代价值符合事先设定的标准时，将画像调整单元中最末一个调整单元产生的调校用户画像作为临时用户画像。换言之，将画像调整单元中最末一个调整单元产生的调校用户画像作为步骤30中用于依据选定用户的业务行为描述向量得到选定用户的最终用户画像的目标临时用户画像。事先设定的标准可以为画像生成模型的代价值不大于预设代价值，上述过程可以对画像生成模型进行调校，同时还可以获取选定用户对应的临时用户画像，帮助后面的调校可以对照选定用户的样本经验，从而增加最终用户画像的精确可靠。As an implementation, when the cost value of the portrait generation model meets the preset standard, the adjusted user portrait generated by the last adjustment unit in the portrait adjustment unit is used as the temporary user portrait. In other words, the adjusted user portrait generated by the last adjustment unit in the portrait adjustment unit is used as the target temporary user portrait for obtaining the final user portrait of the selected user according to the business behavior description vector of the selected user instep 30 . The pre-set standard can be that the cost value of the portrait generation model is not greater than the preset cost value. The above process can adjust the portrait generation model, and at the same time obtain the temporary user portrait corresponding to the selected user, which can help subsequent adjustments Contrast with the sample experience of selected users, thereby increasing the accuracy and reliability of the end user portrait.

作为一些实施方式，获取到最终用户画像后，本申请实施例提供的用户画像识别方法还可以包括将用户画像分类存储的步骤，包括：As some implementations, after the end user portrait is obtained, the user portrait recognition method provided in the embodiment of the present application may also include the step of classifying and storing the user portrait, including:

步骤40：获取最终用户画像和最终用户画像对应的联合分析信息，其中，联合分析信息包括最终用户画像对应的业务行为日志的获取时间和产生业务行为日志的第一选定用户的区域归属特征；确定最终用户画像与对比用户画像集合中的每个对比用户画像的标签匹配度，并通过联合分析信息和每个对比用户画像的联合对比分析信息，得到最终用户画像与每个对比用户画像的完善匹配度；其中，完善匹配度用于从标签指示信息和时域交集情况两方面一起衡量用户画像的匹配度；基于最终用户画像与对比用户画像集合中的每个对比用户画像的标签匹配度和完善匹配度，确定和最终用户画像对应的目标对比用户画像；将最终用户画像更新为新的对比用户画像，然后在对比用户画像集合中构建最终用户画像和预存的目标对比用户画像对应的目标用户画像类型的映射。Step 40: Obtain the end user portrait and the joint analysis information corresponding to the end user portrait, wherein the joint analysis information includes the acquisition time of the business behavior log corresponding to the end user portrait and the area attribution characteristics of the first selected user who generated the business behavior log; Determine the label matching degree of the end user portrait and each comparison user portrait in the comparison user portrait collection, and obtain the perfection of the end user portrait and each comparison user portrait through the joint analysis information and the joint comparison analysis information of each comparison user portrait Matching degree; among them, the perfect matching degree is used to measure the matching degree of the user portrait from the two aspects of label indication information and time domain intersection; based on the tag matching degree and Improve the matching degree, determine the target comparison user portrait corresponding to the end user portrait; update the end user portrait to a new comparison user portrait, and then construct the target user corresponding to the end user portrait and the pre-stored target comparison user portrait in the comparison user portrait collection Mapping of image types.

在产品或服务的推送时，往往是针对一类用户画像的用户进行推送，那么，合理地将用户画像进行按类存储是增加后续高效推送的前提，上述步骤40，通过标签层面和时域层面进行归类分析，确定匹配的目标对比用户画像，并在确定到时，将最终用户画像存储在对比用户画像对应的位置并建立映射信息，归类合理性和准确性高。When products or services are pushed, they are often pushed for users of a type of user portrait. Then, reasonably storing user portraits by category is a prerequisite for increasing subsequent efficient pushes. The above step 40, through the label level and the time domain level Carry out classification analysis, determine the matching target and compare the user portrait, and when determined, store the end user portrait in the corresponding position of the comparison user portrait and establish mapping information, the classification is reasonable and accurate.

其中，通过联合分析信息和每个对比用户画像的联合对比分析信息，确定最终用户画像与每个对比用户画像的完善匹配度的步骤，可以包括：对于每个对比用户画像，基于联合分析信息和对比用户画像的联合对比分析信息，确定关于最终用户画像与该对比用户画像的联合分析特征；其中，联合分析特征表示最终用户画像与该对比用户画像的获取时间和选定用户之间的联系；对于各个对比用户画像，通过最终用户画像和该对比用户画像的标签指示信息，确定关于最终用户画像与该对比用户画像的标签特征；对于每个对比用户画像，通过最终用户画像与该对比用户画像的联合分析特征和标签特征，确定最终用户画像与该对比用户画像的完善匹配度。Wherein, the step of determining the perfect matching degree between the end user portrait and each comparison user portrait through the joint analysis information and the joint comparison analysis information of each comparison user portrait may include: for each comparison user portrait, based on the joint analysis information and Comparing the joint comparative analysis information of the user portraits, and determining the joint analysis features about the end user portrait and the comparison user portrait; wherein, the joint analysis features represent the connection between the end user portrait and the acquisition time of the comparison user portrait and the selected user; For each comparison user portrait, through the end user portrait and the tag indication information of the comparison user portrait, determine the label characteristics about the end user portrait and the comparison user portrait; for each comparison user portrait, through the end user portrait and the comparison user portrait The joint analysis features and label features of , to determine the perfect match between the end user portrait and the comparison user portrait.

作为一种实施方式，联合分析特征包括以下特征的一个或多个：第一选定用户与产生该对比用户画像的第二选定用户之间的区域跨度（所在位置之间的距离）、最终用户画像和该对比用户画像的获取时间间的差值以及第一选定用户与第二选定用户之间的关联预测度；其中，关联预测度表示第一选定用户和第二选定用户之间存在关联关系的概率。标签特征包括以下各个特征中的一个或多个：最终用户画像和对比用户画像的标签匹配度、最终用户画像和对比用户画像的特定画像维度特征，其中，特定画像维度特征包括目标画像维度中的标签向量。As an implementation, the joint analysis features include one or more of the following features: the area span (the distance between the locations) between the first selected user and the second selected user who generated the comparison user portrait, the final The difference between the acquisition time of the user portrait and the comparison user portrait and the association prediction degree between the first selected user and the second selected user; wherein, the association prediction degree represents the relationship between the first selected user and the second selected user The probability that there is a relationship between them. The tag features include one or more of the following features: the tag matching degree of the end user portrait and the comparison user portrait, the specific portrait dimension features of the end user portrait and the comparison user portrait, wherein the specific portrait dimension features include the target portrait dimension label vector.

在一些实施方式中，即使画像的标签共性程度高，但是因为画像对应的选定用户的所在地域差异大，或者分析时间跨度大，用户画像也不适宜划分到一个集合中，而本申请通过选定用户之间的区域跨度用户画像的获取时间差，能够综合时域信息进行画像匹配度的考量，而非仅仅从画像的标签共性进行衡量，增加了画像分类存储的合理性和准确性。In some implementations, even if the labels of the portraits have a high degree of commonality, the user portraits are not suitable to be divided into one set because the geographical differences of the selected users corresponding to the portraits are large, or the analysis time span is large. Given the region span between users and the acquisition time difference of user portraits, time-domain information can be integrated to consider the portrait matching degree, instead of just measuring from the label commonality of portraits, which increases the rationality and accuracy of portrait classification and storage.

综上所述，本申请实施例提供的互动营销平台的用户画像识别方法，通过对选定用户的业务行为日志进行业务行为描述向量挖掘得到选定用户的业务行为描述向量，再依据选定用户的业务行为描述向量对临时用户画像进行调整，从而得到选定用户的最终用户画像，本发明实施例采用临时用户画像获取选定用户的最终用户画像，在具有模板指示的情况下，使得选定用户的最终用户画像偏离性低，精确性高。To sum up, the user portrait recognition method of the interactive marketing platform provided by the embodiment of the present application obtains the business behavior description vector of the selected user by mining the business behavior description vector of the business behavior log of the selected user, and then according to the selected user The business behavior description vector of the temporary user portrait is adjusted to obtain the final user portrait of the selected user. In the embodiment of the present invention, the temporary user portrait is used to obtain the final user portrait of the selected user. In the case of template instructions, the selected The user's end user portrait has low deviation and high accuracy.

基于与图1中所示方法相同的原理，本申请实施例中还提供了一种用户画像识别装置10，如图2所示，该装置10包括：Based on the same principle as the method shown in Figure 1, an embodiment of the present application also provides a userportrait recognition device 10, as shown in Figure 2, thedevice 10 includes:

获取模块11，用于获取业务终端设备发送的选定用户的业务行为日志。The acquiringmodule 11 is configured to acquire the service behavior log of the selected user sent by the service terminal device.

挖掘模块12，用于对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量；Themining module 12 is used to mine the business behavior description vector of the business behavior log to obtain the business behavior description vector of the selected user;

画像获取模块113，用于依据选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像。The portrait acquisition module 113 is configured to adjust the temporary user portrait according to the business behavior description vector of the selected user to obtain the final user portrait of the selected user.

其中，挖掘模块12具体用于：通过画像生成模型中的向量挖掘模块对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量；依据选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像，包括：通过画像生成模型中的画像调整单元，依据向量挖掘模块产生的选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像。向量挖掘模块包括多个依次连接的归类分析单元以及多个依次连接的识别优化单元，画像调整单元包括多个依次连接的调整单元，每个调整单元对应一个识别优化单元；通过画像生成模型中的向量挖掘模块对业务行为日志进行业务行为描述向量挖掘，得到选定用户的业务行为描述向量，包括：通过多个依次连接的归类分析单元对业务行为日志进行归类分析，并通过多个识别优化单元对归类分析单元生成的结果进行识别优化，得到每个识别优化单元针对选定用户的业务行为描述向量；通过画像生成模型中的画像调整单元依据向量挖掘模块产生的选定用户的业务行为描述向量对临时用户画像进行调整，得到选定用户的最终用户画像，包括：对于画像调整单元中的首个调整单元，通过和调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对临时用户画像进行调整，得到调整单元对应的调整临时用户画像；对于首个调整单元之外的每个调整单元，通过和每个调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对前一个调整单元产生的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像；其中，最末一个调整单元获得的调整临时用户画像被确定为选定用户的最终用户画像。每个调整单元包括业务行为描述向量挖掘单元和分析单元，对于画像调整单元中的首个调整单元，通过和调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对临时用户画像进行调整，得到调整单元对应的调整临时用户画像，包括：通过调整单元的业务行为描述向量挖掘单元，对临时用户画像和与调整单元对应的识别优化单元产生的选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到调整单元针对选定用户的业务行为完善描述向量；通过调整单元的分析单元依据业务行为完善描述向量对临时用户画像进行调整，得到后一个调整单元进行第二业务行为描述向量挖掘和调整中运用的调整临时用户画像；对于首个调整单元之外的每个调整单元，通过和每个调整单元对应的识别优化单元产生的选定用户的业务行为描述向量，对前一个调整单元产生的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像，包括：通过调整单元的业务行为描述向量挖掘单元，对前一个调整单元获得的业务行为完善描述向量和调整临时用户画像、和与调整单元对应的识别优化单元产生的选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到调整单元针对选定用户的业务行为完善描述向量；通过调整单元的分析单元依据业务行为完善描述向量对前一个调整单元获得的调整临时用户画像进行调整，得到调整单元对应的调整临时用户画像。Wherein, themining module 12 is specifically used to: carry out business behavior description vector mining on the business behavior log through the vector mining module in the portrait generation model to obtain the business behavior description vector of the selected user; The user portrait is adjusted to obtain the final user portrait of the selected user, including: through the portrait adjustment unit in the portrait generation model, the temporary user portrait is adjusted according to the business behavior description vector of the selected user generated by the vector mining module, and the selected user portrait is obtained. The user's end user persona. The vector mining module includes a plurality of sequentially connected classification analysis units and a plurality of sequentially connected recognition optimization units, and the portrait adjustment unit includes a plurality of sequentially connected adjustment units, each adjustment unit corresponds to a recognition optimization unit; The vector mining module of the business behavior log performs business behavior description vector mining on the business behavior log, and obtains the business behavior description vector of the selected user, including: classifying and analyzing the business behavior log through multiple sequentially connected classification and analysis units, and through multiple The identification optimization unit performs identification optimization on the results generated by the classification analysis unit, and obtains the business behavior description vector of each identification optimization unit for the selected user; through the portrait adjustment unit in the portrait generation model, according to the selected user's profile generated by the vector mining module The business behavior description vector adjusts the temporary user portrait to obtain the final user portrait of the selected user, including: for the first adjustment unit in the portrait adjustment unit, the business behavior of the selected user generated by the identification optimization unit corresponding to the adjustment unit The description vector is used to adjust the temporary user portrait to obtain the adjusted temporary user portrait corresponding to the adjustment unit; for each adjustment unit other than the first adjustment unit, the selected user’s ID generated by the identification optimization unit corresponding to each adjustment unit Business behavior description vector, which adjusts the adjusted temporary user portrait generated by the previous adjustment unit to obtain the adjusted temporary user portrait corresponding to the adjustment unit; among them, the adjusted temporary user portrait obtained by the last adjustment unit is determined as the final selected user User portrait. Each adjustment unit includes a business behavior description vector mining unit and an analysis unit. For the first adjustment unit in the portrait adjustment unit, the business behavior description vector of the selected user generated by the identification and optimization unit corresponding to the adjustment unit is used to analyze the temporary user portrait The adjustment is performed to obtain the adjusted temporary user portrait corresponding to the adjustment unit, including: through the business behavior description vector mining unit of the adjustment unit, the temporary user portrait and the business behavior description vector of the selected user generated by the identification optimization unit corresponding to the adjustment unit are processed. The business behavior description vector is mined to obtain the adjustment unit’s perfect description vector for the selected user’s business behavior; the analysis unit of the adjustment unit adjusts the temporary user portrait according to the business behavior perfect description vector, and obtains the latter adjustment unit for the second business behavior description The adjusted temporary user portrait used in vector mining and adjustment; for each adjustment unit other than the first adjustment unit, the business behavior description vector of the selected user generated by the identification and optimization unit corresponding to each adjustment unit, and the previous one Adjust the adjusted temporary user portrait generated by the adjustment unit to obtain the adjusted temporary user portrait corresponding to the adjustment unit, including: through the business behavior description vector mining unit of the adjustment unit, improve the description vector of the business behavior obtained by the previous adjustment unit and adjust the temporary user The portrait and the business behavior description vector of the selected user generated by the identification and optimization unit corresponding to the adjustment unit are used to mine the business behavior description vector, and the adjustment unit can obtain the perfect description vector for the business behavior of the selected user; through the analysis unit of the adjustment unit, according to the business The behavior perfection description vector adjusts the adjusted temporary user portrait obtained by the previous adjustment unit to obtain the adjusted temporary user portrait corresponding to the adjustment unit.

上述实施例从虚拟模块的角度介绍了用户画像识别装置10，下述从实体模块的角度介绍一种电子设备——AI系统，具体如下所示：The above embodiment introduces the userportrait recognition device 10 from the perspective of a virtual module. The following describes an electronic device——AI system from the perspective of a physical module, as follows:

本申请实施例提供了一种AI系统，如图3所示，AI系统100包括：处理器101和存储器103。其中，处理器101和存储器103相连，如通过总线102相连。可选地，AI系统100还可以包括收发器104。需要说明的是，实际应用中收发器104不限于一个，该AI系统100的结构并不构成对本申请实施例的限定。An embodiment of the present application provides an AI system. As shown in FIG. 3 , theAI system 100 includes: aprocessor 101 and amemory 103 . Wherein, theprocessor 101 is connected to thememory 103 , such as through abus 102 . Optionally, theAI system 100 may further include atransceiver 104 . It should be noted that, in practical applications, thetransceiver 104 is not limited to one, and the structure of theAI system 100 does not limit the embodiment of the present application.

处理器101可以是CPU，通用处理器，GPU，DSP，ASIC，FPGA或者其他可编程逻辑器件、晶体管逻辑器件、硬件部件或者其任意组合。其可以实现或执行结合本申请公开内容所描述的各种示例性的逻辑方框，模块和电路。处理器101也可以是实现计算功能的组合，例如包含一个或多个微处理器组合，DSP和微处理器的组合等。Theprocessor 101 may be a CPU, a general processor, a GPU, a DSP, an ASIC, an FPGA or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can implement or execute the various illustrative logical blocks, modules and circuits described in connection with the present disclosure. Theprocessor 101 may also be a combination that implements computing functions, for example, a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

总线102可包括一通路，在上述组件之间传送信息。总线102可以是PCI总线或EISA总线等。总线102可以分为地址总线、数据总线、控制总线等。为便于表示，图3中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。Bus 102 may include a path for communicating information between the components described above. Thebus 102 may be a PCI bus or an EISA bus or the like. Thebus 102 can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is used in FIG. 3 , but it does not mean that there is only one bus or one type of bus.

存储器103可以是ROM或可存储静态信息和指令的其他类型的静态存储设备，RAM或者可存储信息和指令的其他类型的动态存储设备，也可以是EEPROM、CD-ROM或其他光盘存储、光碟存储(包括压缩光碟、激光碟、光碟、数字通用光碟、蓝光光碟等)、磁盘存储介质或者其他磁存储设备、或者能够用于携带或存储具有指令或数据结构形式的期望的程序代码并能够由计算机存取的任何其他介质，但不限于此。Memory 103 can be ROM or other types of static storage devices that can store static information and instructions, RAM or other types of dynamic storage devices that can store information and instructions, and can also be EEPROM, CD-ROM or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or can be used to carry or store desired program code in the form of instructions or data structures and can be programmed by a computer Any other medium accessed, but not limited to.

存储器103用于存储执行本申请方案的应用程序代码，并由处理器101来控制执行。处理器101用于执行存储器103中存储的应用程序代码，以实现前述任一方法实施例所示的内容。Thememory 103 is used to store the application program code for executing the solution of the present application, and the execution is controlled by theprocessor 101 . Theprocessor 101 is configured to execute the application program code stored in thememory 103, so as to implement the content shown in any one of the foregoing method embodiments.

本申请实施例提供了一种AI系统，本申请实施例中的AI系统包括：一个或多个处理器；存储器；一个或多个计算机程序，其中一个或多个计算机程序被存储在存储器中并被配置为由一个或多个处理器执行，一个或多个程序被处理器执行时，执行上述提供的方法。本申请所提供的技术方案，通过对选定用户的业务行为日志进行业务行为描述向量挖掘得到选定用户的业务行为描述向量，再依据选定用户的业务行为描述向量对临时用户画像进行调整，从而得到选定用户的最终用户画像，本发明实施例采用临时用户画像获取选定用户的最终用户画像，在具有模板指示的情况下，使得选定用户的最终用户画像偏离性低，精确性高。An embodiment of the present application provides an AI system, the AI system in the embodiment of the present application includes: one or more processors; memory; one or more computer programs, wherein one or more computer programs are stored in the memory and It is configured to be executed by one or more processors, and when one or more programs are executed by the processors, the method provided above is executed. The technical solution provided by this application obtains the business behavior description vector of the selected user by mining the business behavior description vector of the business behavior log of the selected user, and then adjusts the temporary user portrait according to the business behavior description vector of the selected user, In order to obtain the end user portrait of the selected user, the embodiment of the present invention uses the temporary user portrait to obtain the end user portrait of the selected user, and in the case of template instructions, the end user portrait of the selected user has low deviation and high accuracy .

本申请实施例提供了一种计算机可读存储介质，该计算机可读存储介质上存储有计算机程序，当计算机程序在处理器上运行时，使得处理器可以执行前述方法实施例中相应的内容。An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program runs on a processor, the processor can execute the corresponding content in the foregoing method embodiments.

应该理解的是，虽然附图的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，其可以以其他的顺序执行。而且，附图的流程图中的至少一部分步骤可以包括多个子步骤或者多个阶段，这些子步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，其执行顺序也不必然是依次进行，而是可以与其他步骤或者其他步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。It should be understood that although the various steps in the flow chart of the accompanying drawings are displayed sequentially according to the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but may be executed at different times, and the order of execution is also It is not necessarily performed sequentially, but may be performed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

以上所述仅是本申请的部分实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本申请原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本申请的保护范围。The above descriptions are only some implementations of the present application. It should be pointed out that for those of ordinary skill in the art, some improvements and modifications can be made without departing from the principle of the application. These improvements and modifications are also It should be regarded as the protection scope of this application.

Claims (11)

Translated fromChinese

1.一种互动营销平台的用户画像识别方法，其特征在于，应用于互动营销平台，所述互动营销平台与业务终端设备通信连接，所述方法包括：1. A user portrait recognition method for an interactive marketing platform, characterized in that it is applied to an interactive marketing platform, and the interactive marketing platform is connected to a service terminal device through communication, and the method includes:获取所述业务终端设备发送的选定用户的业务行为日志；Obtaining the service behavior log of the selected user sent by the service terminal device;对所述业务行为日志进行业务行为描述向量挖掘，得到所述选定用户的业务行为描述向量；Performing business behavior description vector mining on the business behavior log to obtain the business behavior description vector of the selected user;依据所述选定用户的业务行为描述向量对临时用户画像进行调整，得到所述选定用户的最终用户画像。The temporary user portrait is adjusted according to the business behavior description vector of the selected user to obtain the final user portrait of the selected user.2.根据权利要求1所述的方法，其特征在于，所述对所述业务行为日志进行业务行为描述向量挖掘，得到所述选定用户的业务行为描述向量，包括：2. The method according to claim 1, wherein the said business behavior description vector mining is carried out to said business behavior log to obtain the business behavior description vector of said selected user, comprising:通过画像生成模型中的向量挖掘模块对所述业务行为日志进行业务行为描述向量挖掘，得到所述选定用户的业务行为描述向量；Mining the business behavior description vectors of the business behavior logs through the vector mining module in the portrait generation model to obtain the business behavior description vectors of the selected users;所述依据所述选定用户的业务行为描述向量对临时用户画像进行调整，得到所述选定用户的最终用户画像，包括：The step of adjusting the temporary user portrait according to the business behavior description vector of the selected user to obtain the final user portrait of the selected user includes:通过所述画像生成模型中的画像调整单元，依据所述向量挖掘模块产生的选定用户的业务行为描述向量对所述临时用户画像进行调整，得到所述选定用户的最终用户画像。The portrait adjustment unit in the portrait generation model adjusts the temporary user portrait according to the business behavior description vector of the selected user generated by the vector mining module to obtain the final user portrait of the selected user.3.根据权利要求2所述的方法，其特征在于，所述向量挖掘模块包括多个依次连接的归类分析单元以及多个依次连接的识别优化单元，所述画像调整单元包括多个依次连接的调整单元，每个所述调整单元对应一个所述识别优化单元；3. The method according to claim 2, wherein the vector mining module includes a plurality of sequentially connected classification analysis units and a plurality of sequentially connected identification optimization units, and the portrait adjustment unit includes a plurality of sequentially connected adjustment units, each of the adjustment units corresponds to one of the identification optimization units;所述通过画像生成模型中的向量挖掘模块对所述业务行为日志进行业务行为描述向量挖掘，得到所述选定用户的业务行为描述向量，包括：The business behavior description vector mining of the business behavior log by the vector mining module in the portrait generation model is obtained to obtain the business behavior description vector of the selected user, including:通过所述多个依次连接的归类分析单元对所述业务行为日志进行归类分析，并通过多个识别优化单元对所述归类分析单元生成的结果进行识别优化，得到每个所述识别优化单元针对所述选定用户的业务行为描述向量；Classify and analyze the business behavior logs through the plurality of sequentially connected classification and analysis units, and perform recognition and optimization on the results generated by the classification and analysis units through a plurality of identification optimization units, to obtain each identification The optimization unit describes a vector for the business behavior of the selected user;所述通过所述画像生成模型中的画像调整单元依据所述向量挖掘模块产生的选定用户的业务行为描述向量对所述临时用户画像进行调整，得到所述选定用户的最终用户画像，包括：The portrait adjustment unit in the portrait generation model adjusts the temporary user portrait according to the business behavior description vector of the selected user generated by the vector mining module to obtain the final user portrait of the selected user, including :对于所述画像调整单元中的首个所述调整单元，通过和所述调整单元对应的识别优化单元产生的所述选定用户的业务行为描述向量，对所述临时用户画像进行调整，得到所述调整单元对应的调整临时用户画像；For the first adjustment unit in the portrait adjustment unit, adjust the temporary user portrait through the business behavior description vector of the selected user generated by the identification optimization unit corresponding to the adjustment unit, to obtain the adjust the temporary user portrait corresponding to the adjustment unit;对于所述首个调整单元之外的每个所述调整单元，通过和每个所述调整单元对应的识别优化单元产生的所述选定用户的业务行为描述向量，对前一个所述调整单元产生的调整临时用户画像进行调整，得到所述调整单元对应的调整临时用户画像；其中，最末一个所述调整单元获得的调整临时用户画像被确定为所述选定用户的最终用户画像。For each adjustment unit other than the first adjustment unit, the business behavior description vector of the selected user generated by the identification and optimization unit corresponding to each adjustment unit, for the previous adjustment unit The generated adjusted temporary user portrait is adjusted to obtain the adjusted temporary user portrait corresponding to the adjustment unit; wherein, the last adjusted temporary user portrait obtained by the adjustment unit is determined to be the final user portrait of the selected user.4.根据权利要求3所述的方法，其特征在于，每个所述调整单元包括业务行为描述向量挖掘单元和分析单元，所述对于所述画像调整单元中的首个所述调整单元，通过和所述调整单元对应的识别优化单元产生的所述选定用户的业务行为描述向量，对所述临时用户画像进行调整，得到所述调整单元对应的调整临时用户画像，包括：4. The method according to claim 3, wherein each of the adjustment units includes a business behavior description vector mining unit and an analysis unit, and for the first adjustment unit in the portrait adjustment unit, by The business behavior description vector of the selected user generated by the identification optimization unit corresponding to the adjustment unit adjusts the temporary user portrait to obtain the adjusted temporary user portrait corresponding to the adjustment unit, including:通过所述调整单元的业务行为描述向量挖掘单元，对所述临时用户画像和与所述调整单元对应的识别优化单元产生的所述选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到所述调整单元针对所述选定用户的业务行为完善描述向量；Through the business behavior description vector mining unit of the adjustment unit, perform business behavior description vector mining on the temporary user portrait and the business behavior description vector of the selected user generated by the identification optimization unit corresponding to the adjustment unit, to obtain The adjustment unit perfects the description vector for the business behavior of the selected user;通过所述调整单元的分析单元依据所述业务行为完善描述向量对所述临时用户画像进行调整，得到后一个调整单元进行第二业务行为描述向量挖掘和调整中运用的调整临时用户画像；The analysis unit of the adjustment unit adjusts the temporary user portrait according to the business behavior perfect description vector, and obtains the adjusted temporary user portrait used in the second business behavior description vector mining and adjustment by the latter adjustment unit;所述对于所述首个调整单元之外的每个所述调整单元，通过和每个所述调整单元对应的识别优化单元产生的所述选定用户的业务行为描述向量，对前一个所述调整单元产生的调整临时用户画像进行调整，得到所述调整单元对应的调整临时用户画像，包括：For each of the adjustment units other than the first adjustment unit, the business behavior description vector of the selected user generated by the identification and optimization unit corresponding to each of the adjustment units is used for the previous one. The adjusted temporary user portrait generated by the adjustment unit is adjusted to obtain the adjusted temporary user portrait corresponding to the adjustment unit, including:通过所述调整单元的业务行为描述向量挖掘单元，对前一个所述调整单元获得的业务行为完善描述向量和所述调整临时用户画像、和与所述调整单元对应的识别优化单元产生的所述选定用户的业务行为描述向量进行业务行为描述向量挖掘，得到所述调整单元针对所述选定用户的业务行为完善描述向量；Through the business behavior description vector mining unit of the adjustment unit, the business behavior description vector obtained by the previous adjustment unit and the adjusted temporary user portrait, and the identification and optimization unit corresponding to the adjustment unit are generated. Mining the business behavior description vector of the selected user to obtain a perfect description vector for the business behavior of the selected user by the adjustment unit;通过所述调整单元的分析单元依据所述业务行为完善描述向量对前一个所述调整单元获得的调整临时用户画像进行调整，得到所述调整单元对应的调整临时用户画像。The analysis unit of the adjustment unit adjusts the adjusted temporary user portrait obtained by the previous adjustment unit according to the business behavior perfect description vector to obtain the adjusted temporary user portrait corresponding to the adjustment unit.5.根据权利要求4所述的方法，其特征在于，所述方法还包括对所述画像生成模型的调校过程，所述调校过程包括以下步骤：5. method according to claim 4, is characterized in that, described method also comprises the tuning process to described portrait generation model, and described tuning process comprises the following steps:获取所述选定用户的业务行为日志样本；Obtain a sample of the business behavior log of the selected user;通过所述画像生成模型依据所述业务行为日志样本和事先生成的临时用户画像，得到所述选定用户的多个调校用户画像，所述多个调校用户画像包括所述画像调整单元中各个所述调整单元产生的调校用户画像；According to the business behavior log sample and the pre-generated temporary user portrait through the portrait generation model, multiple adjusted user portraits of the selected user are obtained, and the multiple adjusted user portraits are included in the portrait adjustment unit Adjustment user portraits generated by each adjustment unit;依据所述多个调校用户画像确定代价值loss1；Determine the cost value loss1 according to the plurality of adjusted user portraits;基于所述代价值loss1调节所述画像生成模型的模型系数；Adjusting the model coefficients of the portrait generation model based on the cost value loss1;每个所述调校用户画像的描绘体系由多个调校画像维度组成，所述依据所述多个调校用户画像确定代价值loss1，包括：The depiction system of each adjusted user portrait is composed of multiple adjusted portrait dimensions, and the cost value loss1 is determined according to the multiple adjusted user portraits, including:针对每个所述调校用户画像，依据所述调校用户画像的描绘体系中各调校画像维度的代表标签与画像维度样本的代表标签之间的矢量差代价值，其中，所述画像维度样本为所述选定用户的实际用户画像的描绘体系中的画像维度；For each of the adjusted user portraits, according to the vector difference cost value between the representative label of each adjusted portrait dimension and the representative label of the portrait dimension sample in the depiction system of the adjusted user portrait, wherein the portrait dimension The sample is the portrait dimension in the depiction system of the actual user portrait of the selected user;以及，获取所述调校用户画像的每个调校画像维度的规范化代价值，所述规范化代价值与每个所述调校画像维度之间的调整内容关联；And, obtain the normalized cost value of each adjusted portrait dimension of the adjusted user portrait, and the normalized cost value is associated with the adjustment content between each of the adjusted portrait dimensions;通过各个所述调校用户画像对应的所述矢量差代价值和所述规范化代价值，得到所述代价值loss1。The cost value loss1 is obtained through the vector difference cost value and the normalized cost value corresponding to each of the adjusted user portraits.6.根据权利要求5所述的方法，其特征在于，所述通过各个所述调校用户画像对应的所述矢量差代价值和所述规范化代价值，获得所述代价值loss1，包括：6. The method according to claim 5, wherein the obtaining the cost value loss1 through the vector difference cost value and the normalized cost value corresponding to each of the adjusted user portraits includes:针对每个所述调校用户画像，将所述调校用户画像对应的所述矢量差代价值和所述规范化代价值进行合并，得到所述调校用户画像的代价值；For each of the adjustment user portraits, merging the vector difference cost value corresponding to the adjustment user portrait and the normalized cost value to obtain the cost value of the adjustment user portrait;将各个所述调校用户画像的代价值求和，得到所述代价值loss1；Summing the cost values of each of the adjusted user portraits to obtain the cost value loss1;其中，各个所述调校用户画像的规范化代价值相应的权值和预设参考条件反向关联，所述预设参考条件包括以下条件中的至少一个：获取所述调校用户画像的调校循环轮数；获取到所述调校用户画像的调整单元在所述画像调整单元中的顺次；Wherein, the weight corresponding to the normalized cost value of each of the adjusted user portraits is inversely associated with preset reference conditions, and the preset reference conditions include at least one of the following conditions: obtaining the adjustment of the adjusted user portrait The number of rounds; the order of the adjustment unit that adjusts the user portrait in the portrait adjustment unit is obtained;所述方法还包括：The method also includes:通过所述画像生成模型得到所述业务行为日志样本对于所述选定用户的行为向量聚类识别结果，所述行为向量聚类识别结果依据所述向量挖掘模块中最末一个识别优化单元产生的选定用户的业务行为描述向量获取；Obtain the behavior vector clustering recognition result of the business behavior log sample for the selected user through the portrait generation model, and the behavior vector clustering recognition result is generated according to the last recognition optimization unit in the vector mining module Acquire the business behavior description vector of the selected user;依据所述行为向量聚类识别结果以及所述业务行为日志样本对于所述选定用户的实际行为向量聚类识别结果，获得代价值loss2；Obtaining a cost value loss2 according to the behavior vector clustering identification result and the actual behavior vector clustering identification result of the selected user for the business behavior log sample;所述基于所述代价值loss1调节所述画像生成模型的模型系数，包括：The adjustment of the model coefficients of the portrait generation model based on the cost value loss1 includes:通过所述代价值loss1和代价值loss2调节所述画像生成模型的系数。The coefficients of the portrait generation model are adjusted through the cost value loss1 and the cost value loss2.7.根据权利要求6所述的方法，其特征在于，所述方法还包括：7. The method according to claim 6, further comprising:通过所述画像生成模型得到各个所述行为向量聚类对应的矢量域差值，所述矢量域差值依据所述向量挖掘模块中最末一个所述识别优化单元产生的选定用户的业务行为描述向量获取；The vector domain difference corresponding to each of the behavior vector clusters is obtained through the portrait generation model, and the vector domain difference is based on the business behavior of the selected user generated by the last identification and optimization unit in the vector mining module Description vector acquisition;依据所述矢量域差值和所述业务行为日志样本关于所述选定用户的实际矢量域差值，得到代价值loss3。A cost value loss3 is obtained according to the vector domain difference and the actual vector domain difference of the service behavior log sample with respect to the selected user.8.根据权利要求7所述的方法，其特征在于，所述基于所述代价值loss1调节所述画像生成模型的模型系数，包括：8. The method according to claim 7, wherein the adjusting the model coefficients of the portrait generation model based on the cost value loss1 comprises:通过所述代价值loss1、所述代价值loss2和所述代价值loss3获得期望代价值；Obtaining an expected cost value through the cost value loss1, the cost value loss2 and the cost value loss3;基于所述期望代价值调节所述画像生成模型的模型系数。Model coefficients of the profile generation model are adjusted based on the expected cost value.9.根据权利要求8所述的方法，其特征在于，所述通过所述画像生成模型依据所述业务行为日志样本和事先生成的临时用户画像，得到所述选定用户的多个调校用户画像之后，所述方法还包括：9. The method according to claim 8, wherein the portrait generation model is used to obtain a plurality of adjustment users of the selected user according to the business behavior log sample and the temporary user portrait generated in advance After the portrait, the method also includes:对所述画像调整单元最末一个所述调整单元产生的调校用户画像进行整合操作，所述整合操作包括以下操作中的至少一个：对所述调校用户画像的画像维度进行维度向量统一；对所述调校用户画像的画像维度卷积操作；Perform an integration operation on the adjustment user portrait generated by the last adjustment unit of the portrait adjustment unit, and the integration operation includes at least one of the following operations: unify the dimension vectors of the portrait dimensions of the adjustment user portrait; Convolute the portrait dimension of the adjusted user portrait;将基于所述整合操作后的调校用户画像作为后一轮调校的预设临时用户画像，对所述画像生成模型进行调校；Using the adjusted user portrait based on the integration operation as the preset temporary user portrait for the next round of adjustment, and adjust the portrait generation model;当所述画像生成模型的代价值符合事先设定的标准时，将所述画像调整单元中最末一个所述调整单元产生的调校用户画像作为所述临时用户画像。When the cost value of the portrait generation model meets the preset standard, the adjusted user portrait generated by the last adjustment unit among the portrait adjustment units is used as the temporary user portrait.10.根据权利要求1所述的方法，其特征在于，所述方法还包括：10. The method of claim 1, further comprising:获取所述最终用户画像和所述最终用户画像对应的联合分析信息；其中，所述联合分析信息包括所述最终用户画像对应的业务行为日志的获取时间和产生所述业务行为日志的第一选定用户的区域归属特征；Acquiring the end user portrait and joint analysis information corresponding to the end user portrait; wherein the joint analysis information includes the acquisition time of the business behavior log corresponding to the end user portrait and the first choice for generating the business behavior log. Determine the user's regional attribution characteristics;确定所述最终用户画像与对比用户画像集合中的每个对比用户画像的标签匹配度，并通过所述联合分析信息和每个对比用户画像的联合对比分析信息，得到所述最终用户画像与每个对比用户画像的完善匹配度；其中，所述完善匹配度用于从标签指示信息和时域交集情况两方面一起衡量用户画像的匹配度；Determine the label matching degree of the end user portrait and each comparison user portrait in the comparison user portrait set, and obtain the end user portrait and each comparison user portrait through the joint analysis information and the joint comparison analysis information of each comparison user portrait A comparison of the perfect matching degree of the user portrait; wherein, the perfect matching degree is used to measure the matching degree of the user portrait from the two aspects of label indication information and time domain intersection;基于所述最终用户画像与对比用户画像集合中的每个对比用户画像的标签匹配度和完善匹配度，确定和所述最终用户画像对应的目标对比用户画像；Based on the tag matching degree and perfect matching degree of the end user portrait and each comparison user portrait in the comparison user portrait set, determine a target comparison user portrait corresponding to the end user portrait;将所述最终用户画像更新为新的对比用户画像，然后在所述对比用户画像集合中构建所述最终用户画像和预存的所述目标对比用户画像对应的目标用户画像类型的映射。The end user portrait is updated to a new comparison user portrait, and then a mapping between the end user portrait and the pre-stored target user portrait type corresponding to the target comparison user portrait is constructed in the comparison user portrait set.11.一种AI系统，其特征在于，包括相互连接的处理器和存储器，所述存储器存储有计算机程序，当所述处理器运行所述计算机程序时，执行如权利要求1-9任一项所述的方法。11. An AI system, characterized in that it includes a processor and a memory connected to each other, the memory stores a computer program, and when the processor runs the computer program, it executes any one of claims 1-9 the method described.

Images