技术领域Technical Field
本申请涉及人工智能技术领域，尤其涉及一种客户资源确定方法、模型训练方法、电子设备及存储介质。The present application relates to the field of artificial intelligence technology, and in particular to a customer resource determination method, a model training method, an electronic device and a storage medium.
背景技术Background technique
目前，金融产品提供方所提供的产品并不是单一的，即存在形式各异的产品。同时，客户在不同的产品上的需求程度也不相同，例如部分客户可能对一些产品存在比较强的需求，另一部分客户则可能对这些产品的需求较弱。如果我们使用同一量纲来衡量客户对不同产品的需求，那么客户资源(即客户在相应产品上的潜在价值)则可以通过利润和需求进行计算得到。At present, the products provided by financial product providers are not single, that is, there are products of different forms. At the same time, customers have different demands for different products. For example, some customers may have strong demands for some products, while other customers may have weak demands for these products. If we use the same dimension to measure customer demand for different products, then customer resources (that is, the potential value of customers for the corresponding products) can be calculated through profits and demand.
实际应用中，通过统一权重来预测客户对不同产品的需求，其中，每一需求对应预测模型的一个目标的结果。然而，目前在通过预测模型涉及的统一权重进行多目标预测时，存在跷跷板效应。即对一个目标的预测效果提升通常会导致其他目标的预测效果下降，使得由此预测得到的客户资源不准确，即客户资源确定不准确。In practical applications, unified weights are used to predict customer demand for different products, where each demand corresponds to the result of one target of the prediction model. However, when performing multi-target predictions using unified weights involved in the prediction model, a seesaw effect exists. That is, an improvement in the prediction effect for one target usually leads to a decrease in the prediction effect for other targets, making the customer resources predicted inaccurate, that is, the customer resources are not accurately determined.
因此，实际应用中需要一种能够提高客户资源的预测准确性的方案。Therefore, a solution that can improve the prediction accuracy of customer resources is needed in practical applications.
发明内容Summary of the invention
本申请的主要目的在于提供一种客户资源确定方法、模型训练方法、电子设备及存储介质，旨在解决客户资源的预测准确性不高的技术问题。The main purpose of this application is to provide a customer resource determination method, a model training method, an electronic device and a storage medium, aiming to solve the technical problem of low prediction accuracy of customer resources.
为实现上述目的，本申请提供一种客户资源确定方法，所述客户资源确定方法包括以下步骤：To achieve the above object, the present application provides a method for determining customer resources, the method comprising the following steps:
当接收到客户资源确定请求时，基于所述客户资源确定请求包含的客户标识，确定客户的客户特征向量；When receiving a client resource determination request, determining a client feature vector of the client based on a client identifier included in the client resource determination request;
将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果；Inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert sub-model output result and the shared expert sub-model output result of each target respectively through the exclusive expert sub-model and the shared expert sub-model of each target in the preset multi-objective resource prediction model;
基于所述各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述共享专家子模型输出结果和所述独占专家子模型输出结果，确定所述客户的资源表示结果。The resource representation result of the client is determined based on the first difference weight of the respective objectives for the shared expert sub-model, the second difference weight for the exclusive expert sub-model, the shared expert sub-model output result and the exclusive expert sub-model output result.
示例性的，所述基于所述各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述共享专家子模型输出结果和所述独占专家子模型输出结果，确定所述客户的资源表示结果，包括：Exemplarily, the determining the resource representation result of the client based on the first difference weight of the shared expert sub-model, the second difference weight of the exclusive expert sub-model, the output result of the shared expert sub-model, and the output result of the exclusive expert sub-model for each target includes:
分别基于所述各个目标针对所述共享专家子模型的第一差异权重对所述共享专家子模型输出结果进行加权处理，得到第一加权处理结果；Performing weighted processing on the output results of the shared expert sub-model based on the first difference weights of the shared expert sub-model for each of the objectives to obtain a first weighted processing result;
分别基于所述各个目标针对所述独占专家子模型的第二差异权重对所述独享专家子模型输出结果进行加权处理，得到第二加权处理结果；Performing weighted processing on the output results of the exclusive expert sub-model based on the second difference weights of the exclusive expert sub-model for each of the objectives to obtain a second weighted processing result;
分别基于所述第一加权处理结果和所述第二加权处理结果，得到各个目标的资源表示结果，以及基于所述各个目标的资源表示结果，确定所述客户的资源表示结果。Based on the first weighted processing result and the second weighted processing result, resource representation results of each target are obtained respectively, and based on the resource representation results of each target, the resource representation result of the client is determined.
示例性的，所述独占专家子模型包括多个独占专家组，一个所述独占专家组包括多个独占专家，其中，一个所述独占专家组对应一个所述目标；所述共享专家子模型包括一个共享专家组、所述共享专家组包括多个共享专家，其中，各所述目标共享一个所述共享专家组；所述将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果，包括：Exemplarily, the exclusive expert sub-model includes multiple exclusive expert groups, one exclusive expert group includes multiple exclusive experts, wherein one exclusive expert group corresponds to one target; the shared expert sub-model includes one shared expert group, the shared expert group includes multiple shared experts, wherein each target shares one shared expert group; the inputting of the customer feature vector into the preset multi-target resource prediction model, and respectively determining the output results of the exclusive expert sub-model and the shared expert sub-model of each target in the preset multi-target resource prediction model, include:
将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；Inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert output result and the shared expert output result of each objective respectively through the exclusive experts and the shared experts corresponding to each objective in the preset multi-objective resource prediction model;
基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果。Based on the third difference weight of each exclusive expert and the fourth weight of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, the exclusive expert sub-model output results and the shared expert sub-model output results of each target are determined respectively.
示例性的，所述独占专家子模型包括多个独占专家组分别对应的第一独占专家门控网络，所述共享专家子模型包括一个共享专家组对应的第一共享专家门控网络；所述基于各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的独占专家输出结果和所述共享专家输出结果，分别确定各个目标的独占专家子模型输出结果和共享专家子模型输出结果，包括：Exemplarily, the exclusive expert sub-model includes first exclusive expert gating networks corresponding to multiple exclusive expert groups, and the shared expert sub-model includes a first shared expert gating network corresponding to a shared expert group; the exclusive expert sub-model output results and the shared expert sub-model output results of each target are respectively determined based on the third difference weights of each exclusive expert and the fourth weights of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, including:
通过所述第一独占专家门控网络，基于各个目标针对各个独占专家的第三差异权重对所述各个目标的独占专家输出结果进行加权处理，分别确定所述各个目标的独占专家子模型输出结果；Through the first exclusive expert gating network, based on the third difference weight of each target for each exclusive expert, the exclusive expert output results of each target are weighted, and the exclusive expert sub-model output results of each target are determined respectively;
通过所述第一共享专家门控网络，基于所述各个目标针对各个共享专家的第四权重对所述各个目标的共享专家输出结果进行加权处理，分别确定所述各个目标的共享专家子模型输出结果。The shared expert output results of each target are weighted based on the fourth weights of each target for each shared expert through the first shared expert gating network, and the shared expert sub-model output results of each target are determined respectively.
示例性的，所述预设多目标资源预测模型还包括所述各个目标的独占专家子模型对应的第二独占门控网络和所述各个目标的共享专家子模型对应的第二共享专家门控网络，所述基于所述各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述共享专家子模型输出结果和所述独占专家子模型输出结果，确定所述客户的资源表示结果，包括：Exemplarily, the preset multi-objective resource prediction model further includes a second exclusive gating network corresponding to the exclusive expert sub-model of each objective and a second shared expert gating network corresponding to the shared expert sub-model of each objective, and the determining of the resource representation result of the client based on the first difference weight of each objective for the shared expert sub-model, the second difference weight for the exclusive expert sub-model, the output result of the shared expert sub-model and the output result of the exclusive expert sub-model includes:
通过所述第二独占门控网络，分别基于所述各个目标针对所述独占专家子模型的第二差异权重对所述独占专家子模型输出结果进行加权处理；以及通过所述第二共享专家门控网络，分别基于所述各个目标针对所述独占专家子模型的第一差异权重对所述共享专家子模型输出结果进行加权处理，以确定所述客户的资源表示结果。The output results of the exclusive expert sub-model are weighted by the second exclusive gating network based on the second difference weights of the exclusive expert sub-model for each of the objectives; and the output results of the shared expert sub-model are weighted by the second shared expert gating network based on the first difference weights of the exclusive expert sub-model for each of the objectives to determine the resource representation result of the client.
示例性的，为实现上述目的，本申请还提供一种模型训练方法，所述模型训练方法包括以下步骤：Exemplarily, to achieve the above purpose, the present application also provides a model training method, which includes the following steps:
获取待训练预测模型；所述待训练预测模型包括待训练共享专家子模型和待训练独占专家子模型，所述待训练共享专家子模型和所述待训练独占专家子模型分别对应多个目标；Obtaining a prediction model to be trained; the prediction model to be trained includes a shared expert sub-model to be trained and an exclusive expert sub-model to be trained, and the shared expert sub-model to be trained and the exclusive expert sub-model to be trained correspond to multiple targets respectively;
基于不同的所述目标，分别为所述待训练共享专家子模型和所述待训练独占专家子模型分配初始的第一差异权重和初始的第二差异权重；Based on the different objectives, respectively assigning an initial first difference weight and an initial second difference weight to the shared expert sub-model to be trained and the exclusive expert sub-model to be trained;
基于所述初始的第一差异权重、所述初始的第二差异权重、所述待训练共享专家子模型和所述待训练独占专家子模型进行迭代训练，得到多目标资源预测模型。Iterative training is performed based on the initial first difference weight, the initial second difference weight, the shared expert sub-model to be trained and the exclusive expert sub-model to be trained to obtain a multi-objective resource prediction model.
示例性的，所述基于所述初始的第一差异权重、所述初始的第二差异权重、所述待训练共享专家子模型和所述待训练独占专家子模型进行迭代训练，得到多目标资源预测模型，包括：Exemplarily, the iterative training based on the initial first difference weight, the initial second difference weight, the shared expert sub-model to be trained, and the exclusive expert sub-model to be trained to obtain the multi-objective resource prediction model includes:
获取样本数据和所述样本数据对应的各个目标的资源真实表示；Acquire sample data and resource real representations of each target corresponding to the sample data;
将所述样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中各个目标的所述待训练独占专家子模型和所述待训练共享专家子模型分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果；Inputting the sample feature vector corresponding to the sample data into the prediction model to be trained, and determining the output result of the exclusive expert sub-model to be trained and the output result of the shared expert sub-model to be trained of each target in the prediction model to be trained respectively through the exclusive expert sub-model to be trained and the shared expert sub-model to be trained of each target;
分别基于所述各个目标针对所述待训练共享专家子模型的初始的第一差异权重对所述共享专家子模型输出结果进行加权处理，得到第三加权处理结果；Performing weighted processing on the output results of the shared expert sub-model based on the initial first difference weights of the shared expert sub-model to be trained based on the various objectives, to obtain a third weighted processing result;
分别基于所述各个目标针对所述待训练独占专家子模型的初始的第二差异权重对所述独享专家子模型输出结果进行加权处理，得到第四加权处理结果；performing weighted processing on the output results of the exclusive expert sub-model based on the initial second difference weights of the exclusive expert sub-model to be trained based on the respective objectives to obtain a fourth weighted processing result;
分别基于所述第三加权处理结果和所述第四加权处理结果，得到各个目标的资源预测表示；Obtaining resource forecast representations of respective targets based on the third weighted processing result and the fourth weighted processing result respectively;
根据所述各个目标的资源真实表示和所述各个目标的资源预测表示，迭代训练所述待训练预测模型，所述第一差异权重和所述第二差异权重，直到满足预设迭代条件，得到所述多目标资源预测模型。According to the real resource representation of each target and the predicted resource representation of each target, the prediction model to be trained, the first difference weight and the second difference weight are iteratively trained until a preset iteration condition is met to obtain the multi-target resource prediction model.
示例性的，所述待训练独占专家子模型包括多个独占专家组，一个所述独占专家组包括多个独占专家，其中，一个所述独占专家组对应一个所述目标；所述待训练共享专家子模型包括一个共享专家组、所述共享专家组包括多个共享专家，其中，各所述目标共享一个所述共享专家组；所述将所述样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中各个目标的所述待训练独占专家子模型和所述待训练共享专家子模型分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果，包括：Exemplarily, the exclusive expert sub-model to be trained includes multiple exclusive expert groups, one of the exclusive expert groups includes multiple exclusive experts, wherein one exclusive expert group corresponds to one target; the shared expert sub-model to be trained includes one shared expert group, the shared expert group includes multiple shared experts, wherein each target shares one shared expert group; the sample feature vector corresponding to the sample data is input into the prediction model to be trained, and the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained of each target in the prediction model to be trained are respectively determined, including:
将样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；Inputting the sample feature vector corresponding to the sample data into the prediction model to be trained, and determining the exclusive expert output result and the shared expert output result of each target respectively through the exclusive expert and the shared expert corresponding to each target in the prediction model to be trained;
基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果。Based on the third difference weight of each exclusive expert and the fourth weight of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained of each target are determined respectively.
示例性的，所述独占专家子模型包括多个独占专家组分别对应的第一独占专家门控网络，所述共享专家子模型包括一个共享专家组对应的第一共享专家门控网络；所述基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果，包括：Exemplarily, the exclusive expert sub-model includes first exclusive expert gating networks corresponding to multiple exclusive expert groups respectively, and the shared expert sub-model includes a first shared expert gating network corresponding to a shared expert group; the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained for each target are determined based on the third difference weights of each exclusive expert and the fourth weights of each shared expert for each target, and the output results of the exclusive experts and the shared experts for each target, respectively, including:
通过所述第一独占专家门控网络，基于各个目标针对各个独占专家的第三差异权重对所述各个目标的独占专家输出结果进行加权处理，分别确定所述各个目标的待训练独占专家子模型输出结果；Through the first exclusive expert gating network, weighted processing is performed on the exclusive expert output results of each target based on the third difference weight of each target for each exclusive expert, and the output results of the exclusive expert sub-model to be trained for each target are determined respectively;
通过所述第一共享专家门控网络，基于所述各个目标针对各个共享专家的第三差异权重对所述各个目标的共享专家输出结果进行加权处理，分别确定所述各个目标的待训练共享专家子模型输出结果。Through the first shared expert gating network, the shared expert output results of each target are weighted based on the third difference weight of each target for each shared expert, and the output results of the shared expert sub-model to be trained for each target are determined respectively.
示例性的，为实现上述目的，本申请还提供一种电子设备，所述设备包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的客户资源确定程序，所述客户资源确定程序配置为实现如上所述的客户资源确定方法的步骤。Exemplarily, to achieve the above-mentioned purpose, the present application also provides an electronic device, comprising: a memory, a processor, and a customer resource determination program stored in the memory and executable on the processor, wherein the customer resource determination program is configured to implement the steps of the customer resource determination method as described above.
示例性的，为实现上述目的，本申请还提供一种计算机可读存储介质，所述计算机可读存储介质上存储有客户资源确定程序，所述客户资源确定程序被处理器执行时实现如上所述的客户资源确定方法的步骤。Exemplarily, to achieve the above-mentioned purpose, the present application also provides a computer-readable storage medium, on which a customer resource determination program is stored, and when the customer resource determination program is executed by a processor, the steps of the customer resource determination method described above are implemented.
当接收到客户资源确定请求时，基于所述客户资源确定请求包含的客户标识，确定客户的客户特征向量；When receiving a client resource determination request, determining a client feature vector of the client based on a client identifier included in the client resource determination request;
将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定各个目标的独占专家输出结果和共享专家输出结果；Inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert output result and the shared expert output result of each target respectively through the exclusive expert sub-model and the shared expert sub-model of each target in the preset multi-objective resource prediction model;
基于各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述独占专家输出结果和所述共享专家输出结果，确定所述客户的资源表示结果。The resource representation result of the client is determined based on the first difference weights of the respective objectives for the shared expert sub-model, the second difference weights for the exclusive expert sub-model, the exclusive expert output result and the shared expert output result.
本申请实施例提出了在预设多目标资源预测模型中分别设置各个目标的独占专家子模型和共享专家子模型，并且独占专家子模型和共享专家子模型分别具有对应的第二差异权重和第一差异权重，使得不同的目标对应的共享专家输出结果在该目标的模型输出结果中的权重不同，相当于对于不同的目标，共享专家的框架网络的权值也是不同的，而非实际应用中统一或相同的权值。可以理解，本申请的不同之处在于模型结构不同，即区分了独占专家子模型和共享专家子模型，且综合二者的差异权重，可以实现无差别地对待各个目标，降低各个目标的预测偏差，避免出现跷跷板效应，从而提高模型输出结果的准确性，进而提高客户资源的预测准确性。The embodiment of the present application proposes to set up exclusive expert sub-models and shared expert sub-models for each target in a preset multi-target resource prediction model, and the exclusive expert sub-model and the shared expert sub-model have corresponding second difference weights and first difference weights, respectively, so that the weights of the shared expert output results corresponding to different targets in the model output results of the target are different, which is equivalent to that for different targets, the weights of the shared expert framework network are also different, rather than unified or the same weights in actual applications. It can be understood that the difference of the present application lies in the different model structure, that is, the exclusive expert sub-model and the shared expert sub-model are distinguished, and the difference weights of the two are combined to achieve indiscriminate treatment of each target, reduce the prediction deviation of each target, avoid the seesaw effect, thereby improving the accuracy of the model output results, and then improve the prediction accuracy of customer resources.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
图1为本申请实施例方案涉及的MMoE的结构示意图；FIG1 is a schematic diagram of the structure of the MMoE involved in the embodiment of the present application;
图2为本申请实施例方案涉及的CGC的结构示意图；FIG2 is a schematic diagram of the structure of the CGC involved in the embodiment of the present application;
图3为本申请客户资源确定方法第一实施例的流程示意图；FIG3 is a flow chart of a first embodiment of a method for determining customer resources of the present application;
图4为本申请客户资源确定方法实施例中待训练预测模型的一种模型框架示意图；FIG4 is a schematic diagram of a model framework of a prediction model to be trained in an embodiment of the customer resource determination method of the present application;
图5为本申请客户资源确定方法实施例中多目标预测模型的一种模型框架示意图；FIG5 is a schematic diagram of a model framework of a multi-objective prediction model in an embodiment of the customer resource determination method of the present application;
图6为本申请实施例方案涉及的硬件运行环境的结构示意图。FIG6 is a schematic diagram of the structure of the hardware operating environment involved in the embodiment of the present application.
本申请目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings.
具体实施方式Detailed ways
应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
为了更好地理解本申请实施例，以下对本申请实施例进行简要说明：In order to better understand the embodiments of the present application, the embodiments of the present application are briefly described below:
实际应用中，参照图1，图1是MMoE(Multi-gate Mixture-of-Experts，双层多门控专家混合模型)的结构示意图。Gate A和Gate B分别是目标A的门控网络和目标B的门控网络，Expert0、Expert1和Expert2是三个不同的共享专家，通过Gate A可以对Expert0、Expert1和Expert2的三个专家输出结果进行加权融合得到目标A的专家混合输出结果，通过Gate B可以对Expert0、Expert1和Expert2的三个专家输出结果进行加权融合得到目标B的专家混合输出结果。其中，每一共享专家为一个框架网络，其内部运算过程通过全连接神经网络加上激活层实现。在MMoE模型的训练过程中，通过反向传播(BP，Back Propagation)算法来更新权值。而在通过反向传播算法更新权值的过程中，无法通过目标A的模型输出结果和目标B的模型输出结果同时对上述三个共享专家中的权值进行修改，而是一次只能通过一个目标的模型输出结果对该三个共享专家中的权值进行修改。那么，由于三个共享专家的框架网络的权值被目标A和目标B所共享，在通过目标A的模型输出结果对该权值进行修改后，通过更新权值后的共享专家对目标A进行预测时，会使模型的预测准确性得到提高，而通过更新权值后的共享专家对目标B进行预测时，必然会导致模型的预测准确性降低。In practical applications, refer to Figure 1, which is a schematic diagram of the structure of MMoE (Multi-gate Mixture-of-Experts, a two-layer multi-gate expert mixture model). Gate A and Gate B are the gated network of target A and the gated network of target B respectively. Expert0, Expert1 and Expert2 are three different shared experts. Through Gate A, the three expert output results of Expert0, Expert1 and Expert2 can be weighted and fused to obtain the expert mixed output result of target A. Through Gate B, the three expert output results of Expert0, Expert1 and Expert2 can be weighted and fused to obtain the expert mixed output result of target B. Among them, each shared expert is a framework network, and its internal operation process is realized by a fully connected neural network plus an activation layer. During the training process of the MMoE model, the weights are updated by the back propagation (BP, Back Propagation) algorithm. In the process of updating the weights through the back-propagation algorithm, the weights of the three shared experts cannot be modified through the model output results of target A and the model output results of target B at the same time. Instead, the weights of the three shared experts can only be modified through the model output results of one target at a time. Then, since the weights of the framework network of the three shared experts are shared by target A and target B, after the weights are modified through the model output results of target A, when the shared experts with updated weights predict target A, the prediction accuracy of the model will be improved, while when the shared experts with updated weights predict target B, the prediction accuracy of the model will inevitably be reduced.
即目标A的模型输出结果的准确性提高的同时，目标B的模型输出结果的准确性会被降低，这一现象被称之为“跷跷板”现象。That is, while the accuracy of the model output results of target A increases, the accuracy of the model output results of target B will be reduced. This phenomenon is called the "seesaw" phenomenon.
针对“跷跷板”现象，实际应用中提出了CGC，参照图2，图2是CGC的结构示意图。不难看出，相对于MMoE，CGC引入了独占专家的概念，即目标A的专家混合输出结果包括目标A对应的独占专家的专家输出结果和共享专家的专家输出结果，目标B的专家混合输出结果包括目标B对应的独占专家的专家输出结果和共享专家的专家输出结果，可以看出，独占专家的权值是不被多个目标共享的。具体是类似于MMoE，目标A和目标B都对应有一个门控网络，且门控网络用于为目标对应的独占专家和共享专家进行权重分配，即CGC相当于在MMoE的基础上增加了更多的专家，且增加的专家是各个目标独占的而非共享的。也即在通过反向传播算法更新权值时，目标A对应的模型输出结果会作用到共享专家的框架网络的权值的更新，而不会作用到目标B对应的独占专家的权值的更新，使得通过更新权值后的共享专家和未更新权值的独占专家对目标B进行预测时，只有共享专家的专家输出结果会受到目标A对应的模型输出结果的影响，而独占专家的专家输出结果则不会受到其影响。因此，通过引入独占专家，弱化了共享专家的专家输出结果对各个目标的模型输出结果的影响，从而弱化了“跷跷板”现象。In response to the "seesaw" phenomenon, CGC was proposed in practical applications. Refer to Figure 2, which is a schematic diagram of the structure of CGC. It is not difficult to see that, compared with MMoE, CGC introduces the concept of exclusive experts, that is, the expert mixed output result of target A includes the expert output result of the exclusive expert corresponding to target A and the expert output result of the shared expert, and the expert mixed output result of target B includes the expert output result of the exclusive expert corresponding to target B and the expert output result of the shared expert. It can be seen that the weight of the exclusive expert is not shared by multiple targets. Specifically, it is similar to MMoE. There is a gating network corresponding to target A and target B, and the gating network is used to assign weights to the exclusive experts and shared experts corresponding to the target. That is, CGC is equivalent to adding more experts on the basis of MMoE, and the added experts are exclusive to each target rather than shared. That is, when updating the weights through the back-propagation algorithm, the model output results corresponding to target A will affect the update of the weights of the shared expert's framework network, but will not affect the update of the weights of the exclusive expert corresponding to target B. When the shared expert with updated weights and the exclusive expert without updated weights make predictions about target B, only the expert output results of the shared expert will be affected by the model output results corresponding to target A, while the expert output results of the exclusive expert will not be affected. Therefore, by introducing exclusive experts, the influence of the expert output results of the shared expert on the model output results of each target is weakened, thereby weakening the "seesaw" phenomenon.
然而，申请人研究发现，虽然CGC通过引入独占专家的方式弱化了“跷跷板”现象，但弱化的效果仍然不佳。究其原因，CGC是从侧面入手，而非直击导致“跷跷板”现象的核心，即共享专家的框架网络的权值还是所有目标共用的，在通过各个目标的模型输出结果进行反向传播更新权值时，均会对共享专家的框架网络的权值进行更新。However, the applicant found that although CGC weakened the "seesaw" phenomenon by introducing exclusive experts, the weakening effect was still not good. The reason is that CGC started from the side, rather than directly attacking the core of the "seesaw" phenomenon, that is, the weights of the framework network of the shared experts are still shared by all targets. When back-propagating and updating the weights through the model output results of each target, the weights of the framework network of the shared experts will be updated.
为了从直击“跷跷板”现象的问题核心的角度来弱化“跷跷板”现象，从而提高弱化的效果，本申请实施例提出了在预设多目标资源预测模型中分别设置各个目标的独占专家子模型和共享专家子模型，并且独占专家子模型和共享专家子模型分别具有对应的第二差异权重和第一差异权重，使得不同的目标对应的共享专家输出结果在该目标的模型输出结果中的权重不同，相当于对于不同的目标，共享专家的框架网络的权值也是不同的，而非实际应用中统一或相同的权值。可以理解，本申请的不同之处在于模型结构不同，即区分了独占专家子模型和共享专家子模型，且综合二者的差异权重，可以实现无差别地对待各个目标，降低各个目标的预测偏差，避免出现跷跷板效应，从而提高模型输出结果的准确性，进而提高客户资源的预测准确性。In order to weaken the "seesaw" phenomenon from the perspective of directly hitting the core of the problem of the "seesaw" phenomenon, thereby improving the weakening effect, the embodiment of the present application proposes to set up exclusive expert sub-models and shared expert sub-models for each target in the preset multi-target resource prediction model, and the exclusive expert sub-model and the shared expert sub-model have corresponding second difference weights and first difference weights, respectively, so that the shared expert output results corresponding to different targets have different weights in the model output results of the target, which is equivalent to that for different targets, the weights of the framework network of shared experts are also different, rather than unified or the same weights in actual applications. It can be understood that the difference of the present application lies in the different model structure, that is, the exclusive expert sub-model and the shared expert sub-model are distinguished, and the difference weights of the two are combined to achieve indiscriminate treatment of each target, reduce the prediction deviation of each target, avoid the seesaw effect, thereby improving the accuracy of the model output results, and then improve the prediction accuracy of customer resources.
本申请提供一种客户资源确定方法，参照图3，图3为本申请客户资源确定方法第一实施例的流程示意图。The present application provides a method for determining customer resources. Referring to FIG. 3 , FIG. 3 is a flow chart of a first embodiment of the method for determining customer resources of the present application.
本申请实施例提供了客户资源确定方法的实施例，需要说明的是，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。需要说明的是，本申请是在CGC的基础上实现的，即同样引入了独占专家的概念，该客户资源确定方法包括：The embodiment of the present application provides an embodiment of a method for determining customer resources. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that here. It should be noted that the present application is implemented on the basis of CGC, that is, the concept of exclusive experts is also introduced. The method for determining customer resources includes:
步骤S310，当接收到客户资源确定请求时，基于所述客户资源确定请求包含的客户标识，确定客户的客户特征向量。Step S310: When a customer resource determination request is received, a customer feature vector of the customer is determined based on the customer identifier included in the customer resource determination request.
可以理解，客户资源确定请求用于请求客户的资源表示结果。其中，客户包括但不限于个人、企业。相应地，通过该客户标识可以查询得到该客户对应的相关数据。其中，相关数据包括但不限于个人相关数据、企业相关数据。需要说明的是，相关数据包括多个特征数据，即多个维度上的数据，其既能体现特征维度，又能体现特征值。例如相关数据中包括数据：贷款额度为20万，则该数据的特征维度为贷款额度，特征值为20万。其中，该客户特征向量包括该特征维度和特征值。It can be understood that the customer resource determination request is used to request the resource representation result of the customer. Among them, customers include but are not limited to individuals and enterprises. Accordingly, the relevant data corresponding to the customer can be queried through the customer identifier. Among them, the relevant data includes but is not limited to personal related data and enterprise related data. It should be noted that the relevant data includes multiple feature data, that is, data on multiple dimensions, which can reflect both feature dimensions and feature values. For example, the relevant data includes data: the loan amount is 200,000, then the feature dimension of the data is the loan amount, and the feature value is 200,000. Among them, the customer feature vector includes the feature dimension and the feature value.
在一实施例中，客户为个人，该个人相关数据一部分来自于APP(Application，应用程序)获取的该个人授权的数据，另一部分来自于该个人与金融产品提供方的交互行为。In one embodiment, the customer is an individual, and part of the individual-related data comes from the data authorized by the individual obtained by the APP (Application), and the other part comes from the interaction between the individual and the financial product provider.
在另一实施例中，客户为企业，该企业相关数据一部分来源于公开公示的信息，另一部分来源于该企业与金融产品提供方之间的交互行为。In another embodiment, the client is an enterprise, and part of the enterprise-related data comes from publicly disclosed information, and the other part comes from the interaction between the enterprise and the financial product provider.
示例性的，公开公式的信息为工商信息，其通常有公开的信息来源，该工商信息包含了企业注册后产生的一些基本信息，该基本信息包括但不限于行业、地域、注册资本。Exemplarily, the publicly available information is business information, which usually has a public information source. The business information includes some basic information generated after the company is registered, and the basic information includes but is not limited to the industry, region, and registered capital.
需要说明的是，通过交互行为得到的企业相关数据更加丰富。交互行为包括但不限于通过APP、公众号或者小程序进行访问、企业的相关人员与金融产品提供方的营销人员的沟通交流。其中，沟通交流的方式包括但不限于线上通过电话营销、线下进行拜访。相应地，通过APP、公众号或者小程序进行访问时，企业相关数据为访问记录中的数据。It should be noted that the enterprise-related data obtained through interactive behaviors are richer. Interactive behaviors include but are not limited to visits through APP, official accounts or mini programs, and communication between relevant personnel of the enterprise and marketing personnel of the financial product provider. Among them, the methods of communication include but are not limited to online telephone marketing and offline visits. Accordingly, when visiting through APP, official accounts or mini programs, the enterprise-related data is the data in the access record.
示例性的，相关数据可以是与客户的信用相关的数据，例如与客户历史的借款行为和还款行为相关，该相关数据包括但不限于客户的用款规模大小、用款周期、逾期风险。这些相关数据往往反映了客户的习惯和偏好。For example, the relevant data may be data related to the credit of the customer, such as the customer's historical borrowing and repayment behaviors, including but not limited to the customer's loan size, loan cycle, and overdue risk. These relevant data often reflect the customer's habits and preferences.
示例性的，相关数据可以是与企业资质、企业司法等相关的数据。For example, the relevant data may be data related to enterprise qualifications, enterprise legal status, etc.
示例性的，相关数据可以是与客户访问保险相关页面、客户访问信贷相关页面、客户访问基金相关页面等相关的数据。Exemplarily, the relevant data may be data related to a customer visiting an insurance-related page, a customer visiting a credit-related page, a customer visiting a fund-related page, and the like.
示例性的，相关数据可以是与客户购保、客户借款的借据形态、客户购买基金等相关的数据。其中，客户借款的借据形态包括但不限于提供给客户的当前利率以及还款形式。该还款形式包括但不限于定期、随借、是否有锁期。For example, the relevant data may be data related to the customer's insurance purchase, the form of the loan note of the customer's loan, the customer's purchase of funds, etc. Among them, the form of the loan note of the customer's loan includes but is not limited to the current interest rate provided to the customer and the repayment form. The repayment form is not limited to regular, flexible, and whether there is a lock-in period.
可以理解，相关数据与预设多目标预测模型相关，相关数据包括预设多目标预测模型中共享专家组对应的相关子数据和独占专家组对应的相关子数据。其中，独占专家组的数量与预设多目标预测模型所需预测的目标的数量相同，每一独占专家组对应一个不同的相关子数据。其中，独占专家组包括多个独占专家，用于处理相关数据中对应目标所独有的数据；共享专家组包括多个共享专家，用于处理相关数据中所有目标所共有的数据。It can be understood that the relevant data is related to the preset multi-objective prediction model, and the relevant data includes the relevant sub-data corresponding to the shared expert group and the relevant sub-data corresponding to the exclusive expert group in the preset multi-objective prediction model. The number of exclusive expert groups is the same as the number of targets that need to be predicted by the preset multi-objective prediction model, and each exclusive expert group corresponds to a different relevant sub-data. The exclusive expert group includes multiple exclusive experts, which are used to process the data unique to the corresponding target in the relevant data; the shared expert group includes multiple shared experts, which are used to process the data common to all targets in the relevant data.
以下以客户为企业，对本申请实施例进行说明，对于客户为个人的情况，其具体实施方式与客户为企业的各实施例基本相同，在此不再赘述。需要说明的是，其区别在于相关数据不同，相应地，预设多目标预测模型针对该相关数据也存在区别，但模型架构相同。The following is an explanation of the embodiments of the present application with the client being an enterprise. For the case where the client is an individual, the specific implementation method is basically the same as the embodiments where the client is an enterprise, and will not be repeated here. It should be noted that the difference lies in the different relevant data, and accordingly, the preset multi-objective prediction model is also different for the relevant data, but the model architecture is the same.
在本实施例中，目标包括两个，分别为保险预测和信贷额度预测，产品包括保险和信贷，产品的需求即为保险产品客单价和信贷额度。可以理解，保险预测任务和信贷额度预测任务之间存在一定的关联，而预设多目标预测模型是要求各目标之间存在一定关联的。In this embodiment, there are two objectives, namely, insurance prediction and credit line prediction, the products include insurance and credit, and the product demand is the insurance product customer unit price and credit line. It can be understood that there is a certain correlation between the insurance prediction task and the credit line prediction task, and the preset multi-objective prediction model requires a certain correlation between the objectives.
示例性的，共享专家包括但不限于企业资质专家、企业司法专家、企业信用专家，这些专家用于处理企业的基本数据，即基本数据包括但不限于上述与企业资质、企业司法、企业信用相关的数据。之所以将该基本数据用于供共享专家处理，是因为企业在具备好的企业资质、好的企业司法以及好的企业信用等的情况下，其具有更高的信贷额度和保险产品客单价潜力。即共享专家组既能用于预测信贷额度，也能用于预测保险产品客单价。Exemplarily, shared experts include but are not limited to enterprise qualification experts, enterprise legal experts, and enterprise credit experts. These experts are used to process the basic data of the enterprise, that is, the basic data includes but is not limited to the above data related to enterprise qualification, enterprise legal, and enterprise credit. The reason why the basic data is used for processing by shared experts is that when an enterprise has good enterprise qualification, good enterprise legal, and good enterprise credit, it has a higher credit limit and insurance product customer unit price potential. That is, the shared expert group can be used to predict both the credit limit and the insurance product customer unit price.
示例性的，独占专家包括保险预测任务对应的独占专家和信贷额度预测任务对应的独占专家。其中，保险预测任务对应的独占专家包括但不限于该领域内的保险平台行为专家、外部购保数据专家；信贷额度预测任务对应的独占专家包括但不限于该领域内的信贷平台行为专家和贷款即时上下文专家。其中，保险平台行为专家用于处理企业的相关人员在平台上对保险相关页面访问的情况对应的相关数据；外部购保数据专家用于处理企业的相关人员在外部购保对应的相关数据；信贷平台行为专家用于处理企业的相关人员在平台上对信贷相关页面访问的情况对应的相关数据；贷款即时上下文专家用于处理企业的相关人员借款对应的相关数据。Exemplarily, the exclusive experts include exclusive experts corresponding to insurance prediction tasks and exclusive experts corresponding to credit line prediction tasks. Among them, the exclusive experts corresponding to insurance prediction tasks include but are not limited to insurance platform behavior experts and external insurance purchase data experts in the field; the exclusive experts corresponding to credit line prediction tasks include but are not limited to credit platform behavior experts and loan instant context experts in the field. Among them, the insurance platform behavior experts are used to process the relevant data corresponding to the visits of relevant personnel of the enterprise to insurance-related pages on the platform; the external insurance purchase data experts are used to process the relevant data corresponding to the external insurance purchases of relevant personnel of the enterprise; the credit platform behavior experts are used to process the relevant data corresponding to the visits of relevant personnel of the enterprise to credit-related pages on the platform; the loan instant context experts are used to process the relevant data corresponding to the borrowings of relevant personnel of the enterprise.
步骤S320，将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果。Step S320, input the customer feature vector into a preset multi-objective resource prediction model, and determine the exclusive expert sub-model output result and the shared expert sub-model output result of each target in the preset multi-objective resource prediction model respectively through the exclusive expert sub-model and the shared expert sub-model of each target.
可以理解，在通过预设多目标预测模型和相关数据对各个目标进行预测时，并不直接将相关数据输入预设多目标预测模型，而是将相关数据处理成客户特征向量后，在将该客户特征向量输入至预设多目标预测模型。其中，独占专家子模型输出结果和共享专家子模型输出结果共同表征客户对产品的需求。It can be understood that when predicting each target through the preset multi-target prediction model and related data, the related data is not directly input into the preset multi-target prediction model, but the related data is processed into a customer feature vector, and then the customer feature vector is input into the preset multi-target prediction model. Among them, the output results of the exclusive expert sub-model and the output results of the shared expert sub-model jointly represent the customer's demand for the product.
示例性的，所述将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果，可以包括：Exemplarily, the step of inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert sub-model output result and the shared expert sub-model output result of each objective in the preset multi-objective resource prediction model respectively through the exclusive expert sub-model and the shared expert sub-model of each objective, may include:
将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；并基于相同的预设权重为所述各个目标的所述独占专家输出结果和所述共享专家输出结果进行加权处理，从而分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果。The customer feature vector is input into a preset multi-objective resource prediction model, and the exclusive expert output result and the shared expert output result of each target are respectively determined by the exclusive experts and the shared experts corresponding to each target in the preset multi-objective resource prediction model; and the exclusive expert output result and the shared expert output result of each target are weighted based on the same preset weight, so as to respectively determine the exclusive expert sub-model output result and the shared expert sub-model output result of each target.
示例性的，所述独占专家子模型包括多个独占专家组，一个所述独占专家组包括多个独占专家，其中，一个所述独占专家组对应一个所述目标；所述共享专家子模型包括一个共享专家组、所述共享专家组包括多个共享专家，其中，各所述目标共享一个所述共享专家组；所述将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果，还可以包括：Exemplarily, the exclusive expert sub-model includes multiple exclusive expert groups, one exclusive expert group includes multiple exclusive experts, wherein one exclusive expert group corresponds to one target; the shared expert sub-model includes one shared expert group, the shared expert group includes multiple shared experts, wherein each target shares one shared expert group; the step of inputting the customer feature vector into a preset multi-target resource prediction model, and respectively determining the output result of the exclusive expert sub-model and the output result of the shared expert sub-model of each target in the preset multi-target resource prediction model through the exclusive expert sub-model and the shared expert sub-model of each target, may also include:
将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；Inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert output result and the shared expert output result of each objective respectively through the exclusive experts and the shared experts corresponding to each objective in the preset multi-objective resource prediction model;
基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果。Based on the third difference weight of each exclusive expert and the fourth weight of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, the exclusive expert sub-model output results and the shared expert sub-model output results of each target are respectively determined.
需要说明的是，独占专家和共享专家的底层网络结构相同，即共享专家和独占专家的框架网络的相同，两者不同之处在于网络参数不同。基于进一步降低跷跷板效应的目的，本申请实施例为所有独占专家组中的各个独占专家设置有第三差异权重，以及为共享专家组中的各个共享专家设置有第四权重，从而通过第三差异权重调节各个目标对应的独占专家的独占专家输出结果，以及通过第四权重调节各个目标对应的共享专家的共享专家输出结果。具体地，第i个共享专家的框架网络的运算通过表示，第j个独占专家的框架网络的运算通过/>表示；第i个共享专家的权重通过/>来表示，第j个独占专家的权重通过/>来表示；相应地，任一目标的独占专家子模型输出结果可以表示为任一目标的共享专家子模型输出结果可以表示为/>It should be noted that the underlying network structure of exclusive experts and shared experts is the same, that is, the framework network of shared experts and exclusive experts is the same, and the difference between the two lies in the different network parameters. For the purpose of further reducing the seesaw effect, the embodiment of the present application sets a third difference weight for each exclusive expert in all exclusive expert groups, and sets a fourth weight for each shared expert in the shared expert group, so as to adjust the exclusive expert output results of the exclusive experts corresponding to each target through the third difference weight, and adjust the shared expert output results of the shared experts corresponding to each target through the fourth weight. Specifically, the operation of the framework network of the i-th shared expert is performed through Indicates that the operation of the framework network of the j-th exclusive expert is performed by/> Indicates; the weight of the i-th shared expert is obtained by/> To express, the weight of the j-th exclusive expert is expressed by/> To represent; accordingly, the output result of the exclusive expert sub-model of any target can be expressed as The output of the shared expert sub-model for any target can be expressed as/>
可以理解，通过对独占专家组的各个独占专家设置不同权重以及对共享专家组的各个共享专家设置不同权重，从而实现对专家组内部进行差异化，进一步降低跷跷板效应。It can be understood that by setting different weights for each exclusive expert in the exclusive expert group and setting different weights for each shared expert in the shared expert group, differentiation can be achieved within the expert group and the seesaw effect can be further reduced.
进一步地，所述独占专家子模型包括多个独占专家组分别对应的第一独占专家门控网络，所述共享专家子模型包括一个共享专家组对应的第一共享专家门控网络；所述基于各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的独占专家输出结果和所述共享专家输出结果，分别确定各个目标的独占专家子模型输出结果和共享专家子模型输出结果，包括：Further, the exclusive expert sub-model includes first exclusive expert gating networks corresponding to multiple exclusive expert groups respectively, and the shared expert sub-model includes a first shared expert gating network corresponding to a shared expert group; the exclusive expert sub-model output results and the shared expert sub-model output results of each target are respectively determined based on the third difference weights for each exclusive expert and the fourth weights for each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, including:
通过所述第一独占专家门控网络，基于各个目标针对各个独占专家的第三差异权重对所述各个目标的独占专家输出结果进行加权处理，分别确定所述各个目标的独占专家子模型输出结果；Through the first exclusive expert gating network, based on the third difference weight of each target for each exclusive expert, the exclusive expert output results of each target are weighted, and the exclusive expert sub-model output results of each target are determined respectively;
通过所述第一共享专家门控网络，基于所述各个目标针对各个共享专家的第三差异权重对所述各个目标的共享专家输出结果进行加权处理，分别确定所述各个目标的共享专家子模型输出结果。The shared expert output results of each target are weighted based on the third difference weights of each target for each shared expert through the first shared expert gating network, and the shared expert sub-model output results of each target are determined respectively.
其中，第一独占专家门控网络和第一共享专家门控网络均为注意力网络。Among them, the first exclusive expert gated network and the first shared expert gated network are both attention networks.
可以理解，本实施例中的第一独占专家门控网络对应上述实施例中的第一共享专家门控网络对应上述实施例中的/>第一共享专家门控网络通过gs(x)i来表示，第一独占专家门控网络通过gc(x)j来表示；相应地，任一目标的独占专家子模型输出结果可以表示为/>任一目标的共享专家子模型输出结果可以表示为It can be understood that the first exclusive expert gating network in this embodiment corresponds to The first shared expert gating network corresponds to the above embodiment /> The first shared expert gating network is represented bygs (x)i , and the first exclusive expert gating network is represented bygc (x)j ; accordingly, the output result of the exclusive expert sub-model of any target can be expressed as/> The output of the shared expert sub-model for any target can be expressed as
步骤S330，基于所述各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述共享专家子模型输出结果和所述独占专家子模型输出结果，确定所述客户的资源表示结果。Step S330, determining the resource representation result of the client based on the first difference weight of the respective objectives for the shared expert sub-model, the second difference weight for the exclusive expert sub-model, the output result of the shared expert sub-model and the output result of the exclusive expert sub-model.
需要说明的是，通过分配权重后的独占专家输出结果和共享专家输出结果计算得到资源表示结果，是通过Tower层来实现的，该Tower层是一个全连接神经网络，可以理解，每一目标对应的模型输出结果对应一个Tower层。可以理解，通过第一差异权重和第二差异权重实现了独占专家组和共享专家组的组外部差异化，从而进一步降低跷跷板效应。It should be noted that the resource representation result is calculated by assigning weights to the exclusive expert output results and the shared expert output results, which is achieved through the Tower layer, which is a fully connected neural network. It can be understood that the model output result corresponding to each target corresponds to a Tower layer. It can be understood that the external differentiation of the exclusive expert group and the shared expert group is achieved through the first difference weight and the second difference weight, thereby further reducing the seesaw effect.
示例性的，所述基于所述各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述共享专家子模型输出结果和所述独占专家子模型输出结果，确定所述客户的资源表示结果，还可以包括：Exemplarily, the determining the resource representation result of the client based on the first difference weight of the shared expert sub-model, the second difference weight of the exclusive expert sub-model, the output result of the shared expert sub-model, and the output result of the exclusive expert sub-model for each target may further include:
分别基于所述各个目标针对所述共享专家子模型的第一差异权重对所述共享专家子模型输出结果进行加权处理，得到第一加权处理结果；Performing weighted processing on the output results of the shared expert sub-model based on the first difference weights of the shared expert sub-model for each of the objectives to obtain a first weighted processing result;
分别基于所述各个目标针对所述独占专家子模型的第二差异权重对所述独享专家子模型输出结果进行加权处理，得到第二加权处理结果；Performing weighted processing on the output results of the exclusive expert sub-model based on the second difference weights of the exclusive expert sub-model for each of the objectives to obtain a second weighted processing result;
分别基于所述第一加权处理结果和所述第二加权处理结果，得到各个目标的资源表示结果，以及基于所述各个目标的资源表示结果，确定所述客户的资源表示结果。Based on the first weighted processing result and the second weighted processing result, resource representation results of each target are obtained respectively, and based on the resource representation results of each target, the resource representation result of the client is determined.
示例性的，类似于上述第三差异权重和第四权重，本申请实施例同样可以通过两种方式来进行表示，其中，第一差异权重可以表示为或G(x)s，第二差异权重可以表示为或G(x)c。For example, similar to the third difference weight and the fourth weight, the embodiment of the present application can also be expressed in two ways, wherein the first difference weight can be expressed as Or G(x)s , the second difference weight can be expressed as or G(x)c .
对于通过G(x)s表示第二共享专家门控网络，G(x)c表示第二独占专家门控网络，所述预设多目标资源预测模型还包括所述各个目标的独占专家子模型对应的第二独占门控网络和所述各个目标的共享专家子模型对应的第二共享专家门控网络，所述基于所述各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述共享专家子模型输出结果和所述独占专家子模型输出结果，确定所述客户的资源表示结果，包括：For G(x)s representing the second shared expert gating network, G(x)c representing the second exclusive expert gating network, the preset multi-objective resource prediction model also includes the second exclusive gating network corresponding to the exclusive expert sub-model of each objective and the second shared expert gating network corresponding to the shared expert sub-model of each objective, and the resource representation result of the client is determined based on the first difference weight of each objective for the shared expert sub-model, the second difference weight for the exclusive expert sub-model, the output result of the shared expert sub-model and the output result of the exclusive expert sub-model, including:
通过所述第二独占门控网络，分别基于所述各个目标针对所述独占专家子模型的第二差异权重对所述独占专家子模型输出结果进行加权处理；以及通过所述第二共享专家门控网络，分别基于所述各个目标针对所述独占专家子模型的第一差异权重对所述共享专家子模型输出结果进行加权处理，以确定所述客户的资源表示结果。The output results of the exclusive expert sub-model are weighted by the second exclusive gating network based on the second difference weights of the exclusive expert sub-model for each of the objectives; and the output results of the shared expert sub-model are weighted by the second shared expert gating network based on the first difference weights of the exclusive expert sub-model for each of the objectives to determine the resource representation result of the client.
其中，第二独占专家门控网络和第二共享专家门控网络均为注意力网络。资源表示结果通过任一目标的模型输出结果(包括独享专家子模型输出结果和共享专家子模型输出结果)进行加权求和或加权求平均的方式计算得到。需要说明的是，通过第二独占专家门控网络和第二共享专家门控网络来分别输出第二差异权重和第一差异权重，提高了权重的准确性，从而提高了模型的准确性。Among them, the second exclusive expert gating network and the second shared expert gating network are both attention networks. The resource representation result is calculated by weighted summing or weighted averaging the model output results of any target (including the output results of the exclusive expert sub-model and the output results of the shared expert sub-model). It should be noted that the second difference weight and the first difference weight are outputted respectively by the second exclusive expert gating network and the second shared expert gating network, which improves the accuracy of the weight and thus improves the accuracy of the model.
需要说明的是，第二共享专家门控网络包括第一差异权重矩阵，第二独占专家门控网络包括第二差异权重矩阵；所述第二差异权重通过所述第二差异权重矩阵和所述客户特征向量计算得到，所述第一差异权重通过所述第一差异权重矩阵和所述客户特征向量计算得到；其中，所述第二差异权重矩阵和所述第一差异权重矩阵在训练得到所述预设多目标资源预测模型的过程中迭代更新。It should be noted that the second shared expert gating network includes a first difference weight matrix, and the second exclusive expert gating network includes a second difference weight matrix; the second difference weight is calculated by the second difference weight matrix and the customer feature vector, and the first difference weight is calculated by the first difference weight matrix and the customer feature vector; wherein, the second difference weight matrix and the first difference weight matrix are iteratively updated during the process of training the preset multi-objective resource prediction model.
在一实施例中，无论是独占专家组中的所有独占专家还是共享专家组中的所有共享专家，对于客户特征向量中每个专家对应的多个特征维度，只取一个特征维度对应的特征参数用于计算第二差异权重和第一差异权重，以此提高模型训练效率。例如多个目标的客户特征向量为[1,3,3,……,4,2,5,……,3,4,1]，其包括若干个特征维度，其中，“1,3,3”是与目标A对应的一个独占专家的3个特征维度对应的特征参数，“4,2,5”是一个共享专家的3个特征维度对应的特征参数，“3,4,1”与目标B对应的一个独占专家的3个特征维度对应的特征参数，在计算目标A对应的第二差异权重和第一差异权重时，可以仅通过“1，3，3”中的“1”和“4，2，5”中的“4”来计算，而不通过“1，3，3”和“4，2，5”来计算。In one embodiment, whether it is all exclusive experts in the exclusive expert group or all shared experts in the shared expert group, for multiple feature dimensions corresponding to each expert in the customer feature vector, only the feature parameters corresponding to one feature dimension are taken to calculate the second difference weight and the first difference weight, so as to improve the model training efficiency. For example, the customer feature vectors of multiple targets are [1,3,3,...,4,2,5,...,3,4,1], which include several feature dimensions, where "1,3,3" are feature parameters corresponding to the three feature dimensions of an exclusive expert corresponding to target A, "4,2,5" are feature parameters corresponding to the three feature dimensions of a shared expert, and "3,4,1" are feature parameters corresponding to the three feature dimensions of an exclusive expert corresponding to target B. When calculating the second difference weight and the first difference weight corresponding to target A, only "1" in "1,3,3" and "4" in "4,2,5" can be used for calculation, but not "1,3,3" and "4,2,5".
在另一实施例中，第二差异权重通过如下公式一计算：In another embodiment, the second difference weight is calculated by the following formula 1:
其中，x为所述客户特征向量，Wc为第二差异权重矩阵，Ws为第一差异权重矩阵；softmax(·)为归一化指数函数；Wherein, x is the customer feature vector, Wc is the second difference weight matrix, Ws is the first difference weight matrix; softmax(·) is the normalized exponential function;
所述第一差异权重通过如下公式二计算：The first difference weight is calculated by the following formula 2:
其中，x为所述客户特征向量，Wc为第二差异权重矩阵，Ws为第一差异权重矩阵；softmax(·)为归一化指数函数。Wherein, x is the customer feature vector,Wc is the second difference weight matrix,Ws is the first difference weight matrix; and softmax(·) is a normalized exponential function.
需要说明的是，在计算一目标对应的第二差异权重时，是不希望其他目标对应的特征参数参与计算的，因此，需要配置好第二差异权重矩阵，以在计算第二差异权重时剔除客户特征向量中不需要的数据。由于计算x·Wc的值为矩阵相乘的过程，因此该剔除的手段可以为将第一权值矩阵中与不需要的数据对应的位置的数值置0。例如客户特征向量为[1，3，3，……，4，2，5，……，3，4，1]，其包括若干个特征维度，其中，“1，3，3”是与目标A对应的一个独占专家的3个特征维度对应的特征参数，“4，2，5”是一个共享专家的3个特征维度对应的特征参数，“3，4，1”与目标B对应的一个独占专家的3个特征维度对应的特征参数。在计算目标A对应的第二差异权重时，可以将目标B对应的特征参数置0，即第二差异权重矩阵为[a，b，c，……，d，e，f，……，0，0，0]，其中，a，b，c，d，e，f均为非0值。It should be noted that when calculating the second difference weight corresponding to a target, it is not desirable for the feature parameters corresponding to other targets to participate in the calculation. Therefore, it is necessary to configure the second difference weight matrix to eliminate unnecessary data in the customer feature vector when calculating the second difference weight. Since calculating the value of x·Wc is a process of matrix multiplication, the elimination method can be to set the values of the positions corresponding to the unnecessary data in the first weight matrix to 0. For example, the customer feature vector is [1, 3, 3, ..., 4, 2, 5, ..., 3, 4, 1], which includes several feature dimensions, among which "1, 3, 3" is the feature parameter corresponding to the three feature dimensions of an exclusive expert corresponding to target A, "4, 2, 5" is the feature parameter corresponding to the three feature dimensions of a shared expert, and "3, 4, 1" is the feature parameter corresponding to the three feature dimensions of an exclusive expert corresponding to target B. When calculating the second difference weight corresponding to target A, the feature parameters corresponding to target B can be set to 0, that is, the second difference weight matrix is [a, b, c, ..., d, e, f, ..., 0, 0, 0], where a, b, c, d, e, f are all non-zero values.
需要说明的是，第二差异权重矩阵和第一差异权重矩阵分别属于第二共享专家门控网络和第二独占专家门控网络中的参数，即能够通过反向传播算法更新。It should be noted that the second difference weight matrix and the first difference weight matrix belong to parameters in the second shared expert gating network and the second exclusive expert gating network respectively, that is, they can be updated by the back propagation algorithm.
以下针对如何得到任一目标的模型输出结果进行说明：The following is an explanation of how to obtain the model output results for any target:
该任一目标的模型输出结果y通过如下公式三计算得到：The model output result y of any target is calculated by the following formula 3:
yk＝hk(f(x)) (三)yk =hk (f(x)) (III)
其中，x为客户特征向量，f(x)为底层网络结构的运算，即共享专家和独占专家的框架网络的运算；h(x)代表Tower层的运算，上标k为目标的代号。Among them, x is the customer feature vector, f(x) is the operation of the underlying network structure, that is, the operation of the framework network of shared experts and exclusive experts; h(x) represents the operation of the Tower layer, and the superscript k is the code of the target.
具体地，任一目标的模型输出结果y通过如下公式四计算得到：Specifically, the model output result y of any target is calculated by the following formula 4:
其中，m为共享专家组中所有共享专家的数量；G(x)s为第一差异权重的运算；fis(x)为第f个共享专家的运算；gs(x)i为第i个共享专家对应的第四权重的运算；n为独占专家组中所有独占专家的数量；G(x)c为第二差异权重的运算；为第j个独占专家的运算；gc(x)j为第j个独占专家对应的第三差异权重的运算。Wherein, m is the number of all shared experts in the shared expert group; G(x)s is the operation ofthe first difference weight;fis (x) is the operation of the fth shared expert;gs (x)i is the operation of the fourth weight corresponding to the i-th shared expert; n is the number of all exclusive experts in the exclusive expert group; G(x)c is the operation of the second difference weight; is the operation of the j-th exclusive expert; gc (x)j is the operation of the third difference weight corresponding to the j-th exclusive expert.
对于第一差异权重表示为第二差异权重表示为/>的实施例，在此不再赘述。For the first difference weight, it is expressed as The second difference weight is expressed as/> The embodiments of the present invention will not be described in detail here.
需要说明的是，对于预设多目标预测模型，所述预设多目标预测模型是通过多组用于衡量客户资源的样本数据以及样本数据对应的各个目标的资源真实表示对待训练预测模型进行迭代训练得到的It should be noted that the preset multi-objective prediction model is obtained by iteratively training the prediction model to be trained using multiple sets of sample data for measuring customer resources and the true representation of resources of each target corresponding to the sample data.
需要说明的是，在迭代训练过程中，每轮训练都是将特征数据对应的样本特征向量输入待训练预测模型，并得到该待训练预测模型输出的预测结果。通过比较得到该预测结果与标签数据之间的误差，并通过该误差和反向传播算法，对待训练预测模型中的参数进行更新，从而得到更新后的待训练预测模型。其中，待训练预测模型中的参数可以包括第一独占专家门控网络中的参数、第一共享专家门控网络中的参数、第二独占专家门控网络中的参数、第二共享专家门控网络中的参数以及各独占专家的框架网络中的参数和各共享专家的框架网络中的参数。在未对待训练预测模型进行训练之前，需要为上述各参数设置一个初始值。It should be noted that in the iterative training process, each round of training is to input the sample feature vector corresponding to the feature data into the prediction model to be trained, and obtain the prediction result output by the prediction model to be trained. The error between the prediction result and the label data is obtained by comparison, and the parameters in the prediction model to be trained are updated through the error and the back propagation algorithm, so as to obtain the updated prediction model to be trained. Among them, the parameters in the prediction model to be trained may include the parameters in the first exclusive expert gated network, the parameters in the first shared expert gated network, the parameters in the second exclusive expert gated network, the parameters in the second shared expert gated network, and the parameters in the framework network of each exclusive expert and the parameters in the framework network of each shared expert. Before the prediction model to be trained is trained, an initial value needs to be set for each of the above parameters.
具体地，根据各个目标的资源真实表示和各个目标的资源预测表示，迭代训练待训练预测模型、第一差异权重和第二差异权重，直到满足预设迭代条件，结束迭代训练，并将最新的更新后的待训练预测模型作为多目标预测模型。其中，资源预测表示为通过待训练预测模型对样本数据进行预测得到。其中，在通过上述误差对待训练预测模型中的参数进行更新时，对于独占专家组和共享专家组，其更新的是相应目标对应的独占专家组中各独占专家的框架网络中的参数，以及共享专家组中各共享专家的框架网络中的参数。Specifically, according to the real resource representation of each target and the resource prediction representation of each target, the prediction model to be trained, the first difference weight and the second difference weight are iteratively trained until the preset iteration condition is met, the iterative training is terminated, and the latest updated prediction model to be trained is used as the multi-target prediction model. The resource prediction representation is obtained by predicting the sample data through the prediction model to be trained. When the parameters in the prediction model to be trained are updated by the above-mentioned error, for the exclusive expert group and the shared expert group, the parameters in the framework network of each exclusive expert in the exclusive expert group corresponding to the corresponding target and the parameters in the framework network of each shared expert in the shared expert group are updated.
示例性的，框架网络包括全连接神经网络和激活层。其中，激活层包括但不限于sigmoid、tanh、relu。Exemplarily, the framework network includes a fully connected neural network and an activation layer, wherein the activation layer includes but is not limited to sigmoid, tanh, and relu.
以下以一可选实施例对上述通过样本数据对待训练预测模型进行迭代训练的过程进行简要阐述：The following is a brief description of the process of iteratively training the prediction model to be trained using sample data using an optional embodiment:
参照图4，图4为待训练预测模型的一种模型框架示意图。在该模型框架中，目标包括两个，分别为目标A和目标B。首先，对样本数据进行数据特征化处理得到样本特征向量，通过“input”将该样本特征向量输入至GATE_A_S(目标A对应的第一门控网络)、GATE_B_S(目标B对应的第一门控网络)、GATE_A(目标A对应的第一门控子网络)、Ea-1……Ea-m(目标A对应的多个独占专家)、GATE_S(目标A或目标B对应的第二门控子网络)、GATE_B(目标B对应的第一门控子网络)、Es-1……Es-m(多个共享专家)以及Eb-1……Eb-m(目标B对应的多个独占专家)。通过相应参数和该样本特征向量计算得到第二差异权重、第一差异权重、第三差异权重、第四权重以及各框架网络的输出结果。Referring to Fig. 4, Fig. 4 is a schematic diagram of a model framework of a prediction model to be trained. In this model framework, there are two targets, namely target A and target B. First, the sample data is subjected to data characterization processing to obtain a sample feature vector, and the sample feature vector is input into GATE_A_S (the first gated network corresponding to target A), GATE_B_S (the first gated network corresponding to target B), GATE_A (the first gated subnetwork corresponding to target A), Ea-1 ... Eam (multiple exclusive experts corresponding to target A), GATE_S (the second gated subnetwork corresponding to target A or target B), GATE_B (the first gated subnetwork corresponding to target B), Es-1 ...Esm (multiple shared experts) and Eb-1 ... Ebm (multiple exclusive experts corresponding to target B) through "input". The second difference weight, the first difference weight, the third difference weight, the fourth weight and the output results of each framework network are calculated by the corresponding parameters and the sample feature vector.
在计算目标A的模型输出结果时，通过Ea-1……Ea-m计算得到多个独占专家输出结果，并通过该GATE_A为该多个独占专家输出结果分配第三差异权重，得到独占专家子模型输出结果；通过Es-1……Es-m计算得到多个共享专家输出结果，并通过该目标A对应的GATE_S为该多个共享专家输出结果分配第四权重，得到共享专家子模型输出结果；之后通过GATE_A_S分别为独占专家子模型输出结果分配第二差异权重，以及为共享专家子模型输出结果分配第一差异权重。通过TOWER_A对分配权重后的独占专家子模型输出结果和共享专家子模型输出结果进行计算得到目标A的模型输出结果，并通过outputA输出该模型输出结果。When calculating the model output result of target A, multiple exclusive expert output results are calculated by Ea-1 ...Eam , and the third difference weight is assigned to the multiple exclusive expert output results by GATE_A to obtain the exclusive expert sub-model output result; multiple shared expert output results are calculated byEs-1 ...Esm , and the fourth weight is assigned to the multiple shared expert output results by GATE_S corresponding to target A to obtain the shared expert sub-model output result; then, the second difference weight is assigned to the exclusive expert sub-model output result and the first difference weight is assigned to the shared expert sub-model output result by GATE_A_S. The model output result of target A is calculated by TOWER_A for the exclusive expert sub-model output result and the shared expert sub-model output result after the weights are assigned, and the model output result is output by outputA.
可以理解，目标B的模型输出结果的计算过程与目标A的模型输出结果的计算过程基本相同，在此不再赘述。It can be understood that the calculation process of the model output result of target B is basically the same as the calculation process of the model output result of target A, and will not be repeated here.
以下以一可选实施例对上述使用预设多目标预测模型进行预测的过程进行简要阐述：The following is a brief description of the process of using the preset multi-objective prediction model to perform prediction using an optional embodiment:
参照图5，图5为预设多目标预测模型的一种模型框架示意图。在该模型框架中，目标包括两个，分别为目标A和目标B。预测样本特征x为客户特征向量，将其分别输入至GATE1(目标A对应的第二独占门控网络)、GATEA(目标A对应的第一独占专家门控网络和第一共享专家门控网络)和GATE2(目标A对应的第二共享专家门控网络)，分别得到第三差异权重(w1和w2)、第二差异权重、第一差异权重以及第四权重(w3和w4)。其中，501为目标A对应的其他计算过程，在501之后得到目标A对应的模型输出结果；502为目标A对应的其他计算过程，在502之后得到目标B对应的模型输出结果。需要说明的是，目标B对应的第三差异权重、第二差异权重、第一差异权重和第四权重的计算过程与目标A对应的第三差异权重、第二差异权重、第一差异权重和第四权重的计算过程基本相同，在此不再赘述。Referring to Figure 5, Figure 5 is a schematic diagram of a model framework of a preset multi-objective prediction model. In this model framework, there are two targets, namely target A and target B. The predicted sample feature x is a customer feature vector, which is input into GATE1 (the second exclusive gating network corresponding to target A), GATEA (the first exclusive expert gating network and the first shared expert gating network corresponding to target A) and GATE2 (the second shared expert gating network corresponding to target A), respectively, to obtain the third difference weight (w1 and w2), the second difference weight, the first difference weight and the fourth weight (w3 and w4). Among them, 501 is the other calculation process corresponding to target A, and the model output result corresponding to target A is obtained after 501; 502 is the other calculation process corresponding to target A, and the model output result corresponding to target B is obtained after 502. It should be noted that the calculation process of the third difference weight, the second difference weight, the first difference weight and the fourth weight corresponding to target B is basically the same as the calculation process of the third difference weight, the second difference weight, the first difference weight and the fourth weight corresponding to target A, which will not be repeated here.
示例性的，本申请还提供一种模型训练方法，所述模型训练方法包括以下步骤：Exemplarily, the present application also provides a model training method, the model training method comprising the following steps:
获取待训练预测模型；所述待训练预测模型包括待训练共享专家子模型和待训练独占专家子模型，所述待训练共享专家子模型和所述待训练独占专家子模型分别对应多个目标；Obtaining a prediction model to be trained; the prediction model to be trained includes a shared expert sub-model to be trained and an exclusive expert sub-model to be trained, and the shared expert sub-model to be trained and the exclusive expert sub-model to be trained correspond to multiple targets respectively;
基于不同的所述目标，分别为所述待训练共享专家子模型和所述待训练独占专家子模型分配初始的第一差异权重和初始的第二差异权重；Based on the different objectives, respectively assigning an initial first difference weight and an initial second difference weight to the shared expert sub-model to be trained and the exclusive expert sub-model to be trained;
基于所述初始的第一差异权重、所述初始的第二差异权重、所述待训练共享专家子模型和所述待训练独占专家子模型进行迭代训练，得到多目标资源预测模型。Iterative training is performed based on the initial first difference weight, the initial second difference weight, the shared expert sub-model to be trained and the exclusive expert sub-model to be trained to obtain a multi-objective resource prediction model.
示例性的，所述基于所述初始的第一差异权重、所述初始的第二差异权重、所述待训练共享专家子模型和所述待训练独占专家子模型进行迭代训练，得到多目标资源预测模型，包括：Exemplarily, the iterative training based on the initial first difference weight, the initial second difference weight, the shared expert sub-model to be trained, and the exclusive expert sub-model to be trained to obtain the multi-objective resource prediction model includes:
获取样本数据和所述样本数据对应的各个目标的资源真实表示；Acquire sample data and resource real representations of each target corresponding to the sample data;
将所述样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中各个目标的所述待训练独占专家子模型和所述待训练共享专家子模型分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果；Inputting the sample feature vector corresponding to the sample data into the prediction model to be trained, and determining the output result of the exclusive expert sub-model to be trained and the output result of the shared expert sub-model to be trained of each target in the prediction model to be trained respectively through the exclusive expert sub-model to be trained and the shared expert sub-model to be trained of each target;
分别基于所述各个目标针对所述待训练共享专家子模型的初始的第一差异权重对所述共享专家子模型输出结果进行加权处理，得到第三加权处理结果；Performing weighted processing on the output results of the shared expert sub-model based on the initial first difference weights of the shared expert sub-model to be trained based on the various objectives, to obtain a third weighted processing result;
分别基于所述各个目标针对所述待训练独占专家子模型的初始的第二差异权重对所述独享专家子模型输出结果进行加权处理，得到第四加权处理结果；Performing weighted processing on the output results of the exclusive expert sub-model based on the initial second difference weights of the exclusive expert sub-model to be trained based on the respective objectives to obtain a fourth weighted processing result;
分别基于所述第三加权处理结果和所述第四加权处理结果，得到各个目标的资源预测表示；Obtaining resource forecast representations of respective targets based on the third weighted processing result and the fourth weighted processing result respectively;
根据所述各个目标的资源真实表示和所述各个目标的资源预测表示，迭代训练所述待训练预测模型，所述第一差异权重和所述第二差异权重，直到满足预设迭代条件，得到所述多目标资源预测模型。According to the real resource representation of each target and the predicted resource representation of each target, the prediction model to be trained, the first difference weight and the second difference weight are iteratively trained until a preset iteration condition is met to obtain the multi-target resource prediction model.
示例性的，所述待训练独占专家子模型包括多个独占专家组，一个所述独占专家组包括多个独占专家，其中，一个所述独占专家组对应一个所述目标；所述待训练共享专家子模型包括一个共享专家组、所述共享专家组包括多个共享专家，其中，各所述目标共享一个所述共享专家组；所述将所述样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中各个目标的所述待训练独占专家子模型和所述待训练共享专家子模型分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果，包括：Exemplarily, the exclusive expert sub-model to be trained includes multiple exclusive expert groups, one of the exclusive expert groups includes multiple exclusive experts, wherein one exclusive expert group corresponds to one target; the shared expert sub-model to be trained includes one shared expert group, the shared expert group includes multiple shared experts, wherein each target shares one shared expert group; the sample feature vector corresponding to the sample data is input into the prediction model to be trained, and the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained of each target in the prediction model to be trained are respectively determined, including:
将样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；Inputting the sample feature vector corresponding to the sample data into the prediction model to be trained, and determining the exclusive expert output result and the shared expert output result of each target respectively through the exclusive expert and the shared expert corresponding to each target in the prediction model to be trained;
基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果。Based on the third difference weight of each exclusive expert and the fourth weight of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained of each target are determined respectively.
示例性的，所述独占专家子模型包括多个独占专家组分别对应的第一独占专家门控网络，所述共享专家子模型包括一个共享专家组对应的第一共享专家门控网络；所述基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果，包括：Exemplarily, the exclusive expert sub-model includes first exclusive expert gating networks corresponding to multiple exclusive expert groups respectively, and the shared expert sub-model includes a first shared expert gating network corresponding to a shared expert group; the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained for each target are determined based on the third difference weights of each exclusive expert and the fourth weights of each shared expert for each target, and the output results of the exclusive experts and the shared experts for each target, respectively, including:
通过所述第一独占专家门控网络，基于各个目标针对各个独占专家的第三差异权重对所述各个目标的独占专家输出结果进行加权处理，分别确定所述各个目标的待训练独占专家子模型输出结果；Through the first exclusive expert gating network, weighted processing is performed on the exclusive expert output results of each target based on the third difference weight of each target for each exclusive expert, and the output results of the exclusive expert sub-model to be trained for each target are determined respectively;
通过所述第一共享专家门控网络，基于所述各个目标针对各个共享专家的第三差异权重对所述各个目标的共享专家输出结果进行加权处理，分别确定所述各个目标的待训练共享专家子模型输出结果。Through the first shared expert gating network, the shared expert output results of each target are weighted based on the third difference weight of each target for each shared expert, and the output results of the shared expert sub-model to be trained for each target are determined respectively.
本申请模型训练方法具体实施方式与上述客户资源确定方法各实施例基本相同，在此不再赘述。The specific implementation method of the model training method of the present application is basically the same as the various embodiments of the above-mentioned customer resource determination method, and will not be repeated here.
此外，本申请还提供一种客户资源确定装置，该客户资源确定装置包括：In addition, the present application also provides a client resource determination device, the client resource determination device comprising:
向量确定模块，用于当接收到客户资源确定请求时，基于所述客户资源确定请求包含的客户标识，确定客户的客户特征向量；A vector determination module, configured to determine a customer feature vector of a customer based on a customer identifier included in the customer resource determination request when receiving the customer resource determination request;
向量输入模块，用于将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中各个目标的独占专家子模型和共享专家子模型分别确定各个目标的独占专家输出结果和共享专家输出结果；A vector input module, used for inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert output result and the shared expert output result of each target respectively through the exclusive expert sub-model and the shared expert sub-model of each target in the preset multi-objective resource prediction model;
结果确定模块，用于基于各个目标针对所述共享专家子模型的第一差异权重、针对所述独占专家子模型的第二差异权重、所述独占专家输出结果和所述共享专家输出结果，确定所述客户的资源表示结果；A result determination module, configured to determine a resource representation result of the client based on a first difference weight of each target for the shared expert sub-model, a second difference weight for the exclusive expert sub-model, the exclusive expert output result, and the shared expert output result;
示例性的，所述结果确定模块具体用于：Exemplarily, the result determination module is specifically used for:
分别基于所述各个目标针对所述共享专家子模型的第一差异权重对所述共享专家子模型输出结果进行加权处理，得到第一加权处理结果；Performing weighted processing on the output results of the shared expert sub-model based on the first difference weights of the shared expert sub-model for each of the objectives to obtain a first weighted processing result;
分别基于所述各个目标针对所述独占专家子模型的第二差异权重对所述独享专家子模型输出结果进行加权处理，得到第二加权处理结果；Performing weighted processing on the output results of the exclusive expert sub-model based on the second difference weights of the exclusive expert sub-model for each of the objectives to obtain a second weighted processing result;
分别基于所述第一加权处理结果和所述第二加权处理结果，得到各个目标的资源表示结果，以及基于所述各个目标的资源表示结果，确定所述客户的资源表示结果。Based on the first weighted processing result and the second weighted processing result, resource representation results of each target are obtained respectively, and based on the resource representation results of each target, the resource representation result of the client is determined.
示例性的，所述独占专家子模型包括多个独占专家组，一个所述独占专家组包括多个独占专家，其中，一个所述独占专家组对应一个所述目标；所述共享专家子模型包括一个共享专家组、所述共享专家组包括多个共享专家，其中，各所述目标共享一个所述共享专家组；所述向量输入模块具体用于：Exemplarily, the exclusive expert sub-model includes multiple exclusive expert groups, one exclusive expert group includes multiple exclusive experts, wherein one exclusive expert group corresponds to one target; the shared expert sub-model includes one shared expert group, the shared expert group includes multiple shared experts, wherein each target shares one shared expert group; the vector input module is specifically used for:
将所述客户特征向量输入到预设多目标资源预测模型，通过所述预设多目标资源预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；Inputting the customer feature vector into a preset multi-objective resource prediction model, and determining the exclusive expert output result and the shared expert output result of each objective respectively through the exclusive experts and the shared experts corresponding to each objective in the preset multi-objective resource prediction model;
基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的独占专家子模型输出结果和共享专家子模型输出结果。Based on the third difference weight of each exclusive expert and the fourth weight of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, the exclusive expert sub-model output results and the shared expert sub-model output results of each target are determined respectively.
示例性的，所述独占专家子模型包括多个独占专家组分别对应的第一独占专家门控网络，所述共享专家子模型包括一个共享专家组对应的第一共享专家门控网络；所述向量输入模块还用于：Exemplarily, the exclusive expert sub-model includes first exclusive expert gating networks corresponding to a plurality of exclusive expert groups respectively, and the shared expert sub-model includes a first shared expert gating network corresponding to a shared expert group; the vector input module is further used for:
通过所述第一独占专家门控网络，基于各个目标针对各个独占专家的第三差异权重对所述各个目标的独占专家输出结果进行加权处理，分别确定所述各个目标的独占专家子模型输出结果；Through the first exclusive expert gating network, based on the third difference weight of each target for each exclusive expert, the exclusive expert output results of each target are weighted, and the exclusive expert sub-model output results of each target are determined respectively;
通过所述第一共享专家门控网络，基于所述各个目标针对各个共享专家的第四权重对所述各个目标的共享专家输出结果进行加权处理，分别确定所述各个目标的共享专家子模型输出结果。The shared expert output results of each target are weighted based on the fourth weights of each target for each shared expert through the first shared expert gating network, and the shared expert sub-model output results of each target are determined respectively.
示例性的，所述预设多目标资源预测模型还包括所述各个目标的独占专家子模型对应的第二独占门控网络和所述各个目标的共享专家子模型对应的第二共享专家门控网络，所述结果确定模块具体用于：Exemplarily, the preset multi-objective resource prediction model further includes a second exclusive gating network corresponding to the exclusive expert sub-model of each objective and a second shared expert gating network corresponding to the shared expert sub-model of each objective, and the result determination module is specifically used for:
通过所述第二独占门控网络，分别基于所述各个目标针对所述独占专家子模型的第二差异权重对所述独占专家子模型输出结果进行加权处理；以及通过所述第二共享专家门控网络，分别基于所述各个目标针对所述独占专家子模型的第一差异权重对所述共享专家子模型输出结果进行加权处理，以确定所述客户的资源表示结果The output results of the exclusive expert sub-model are weighted by the second exclusive gating network based on the second difference weights of the exclusive expert sub-model for each of the objectives; and the output results of the shared expert sub-model are weighted by the second shared expert gating network based on the first difference weights of the exclusive expert sub-model for each of the objectives to determine the resource representation result of the client.
本申请客户资源确定装置具体实施方式与上述客户资源确定方法各实施例基本相同，在此不再赘述。The specific implementation of the customer resource determination device of the present application is basically the same as the various embodiments of the customer resource determination method described above, and will not be repeated here.
此外，本申请还提供一种模型训练装置，该模型训练装置包括：In addition, the present application also provides a model training device, which includes:
模型获取模块，用于获取待训练预测模型；所述待训练预测模型包括待训练共享专家子模型和待训练独占专家子模型，所述待训练共享专家子模型和所述待训练独占专家子模型分别对应多个目标；A model acquisition module, used to acquire a prediction model to be trained; the prediction model to be trained includes a shared expert sub-model to be trained and an exclusive expert sub-model to be trained, and the shared expert sub-model to be trained and the exclusive expert sub-model to be trained correspond to multiple targets respectively;
权重分配模块，用于基于不同的所述目标，分别为所述待训练共享专家子模型和所述待训练独占专家子模型分配初始的第一差异权重和初始的第二差异权重；A weight allocation module, used for respectively allocating an initial first difference weight and an initial second difference weight to the shared expert sub-model to be trained and the exclusive expert sub-model to be trained based on different objectives;
迭代训练模块，用于基于所述初始的第一差异权重、所述初始的第二差异权重、所述待训练共享专家子模型和所述待训练独占专家子模型进行迭代训练，得到多目标资源预测模型。The iterative training module is used to perform iterative training based on the initial first difference weight, the initial second difference weight, the shared expert sub-model to be trained and the exclusive expert sub-model to be trained to obtain a multi-objective resource prediction model.
示例性的，所述迭代训练模块具体用于：Exemplarily, the iterative training module is specifically used for:
获取样本数据和所述样本数据对应的各个目标的资源真实表示；Acquire sample data and resource real representations of each target corresponding to the sample data;
将所述样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中各个目标的所述待训练独占专家子模型和所述待训练共享专家子模型分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果；Inputting the sample feature vector corresponding to the sample data into the prediction model to be trained, and determining the output result of the exclusive expert sub-model to be trained and the output result of the shared expert sub-model to be trained of each target in the prediction model to be trained respectively through the exclusive expert sub-model to be trained and the shared expert sub-model to be trained of each target;
分别基于所述各个目标针对所述待训练共享专家子模型的初始的第一差异权重对所述共享专家子模型输出结果进行加权处理，得到第三加权处理结果；Performing weighted processing on the output results of the shared expert sub-model based on the initial first difference weights of the shared expert sub-model to be trained based on the various objectives, to obtain a third weighted processing result;
分别基于所述各个目标针对所述待训练独占专家子模型的初始的第二差异权重对所述独享专家子模型输出结果进行加权处理，得到第四加权处理结果；Performing weighted processing on the output results of the exclusive expert sub-model based on the initial second difference weights of the exclusive expert sub-model to be trained based on the respective objectives to obtain a fourth weighted processing result;
分别基于所述第三加权处理结果和所述第四加权处理结果，得到各个目标的资源预测表示；Obtaining resource forecast representations of respective targets based on the third weighted processing result and the fourth weighted processing result respectively;
根据所述各个目标的资源真实表示和所述各个目标的资源预测表示，迭代训练所述待训练预测模型，所述第一差异权重和所述第二差异权重，直到满足预设迭代条件，得到所述多目标资源预测模型。According to the real resource representation of each target and the predicted resource representation of each target, the prediction model to be trained, the first difference weight and the second difference weight are iteratively trained until a preset iteration condition is met to obtain the multi-target resource prediction model.
示例性的，所述待训练独占专家子模型包括多个独占专家组，一个所述独占专家组包括多个独占专家，其中，一个所述独占专家组对应一个所述目标；所述待训练共享专家子模型包括一个共享专家组、所述共享专家组包括多个共享专家，其中，各所述目标共享一个所述共享专家组；所述迭代训练模块还用于：Exemplarily, the exclusive expert sub-model to be trained includes multiple exclusive expert groups, one of the exclusive expert groups includes multiple exclusive experts, wherein one exclusive expert group corresponds to one target; the shared expert sub-model to be trained includes one shared expert group, the shared expert group includes multiple shared experts, wherein each target shares one shared expert group; the iterative training module is further used for:
将样本数据对应的样本特征向量输入到所述待训练预测模型，通过所述待训练预测模型中所述各个目标分别对应的所述独占专家和所述共享专家分别确定所述各个目标的独占专家输出结果和共享专家输出结果；Inputting the sample feature vector corresponding to the sample data into the prediction model to be trained, and determining the exclusive expert output result and the shared expert output result of each target respectively through the exclusive expert and the shared expert corresponding to each target in the prediction model to be trained;
基于所述各个目标针对各个独占专家的第三差异权重和各个共享专家的第四权重，和所述各个目标的所述独占专家输出结果和所述共享专家输出结果，分别确定所述各个目标的待训练独占专家子模型输出结果和待训练共享专家子模型输出结果。Based on the third difference weight of each exclusive expert and the fourth weight of each shared expert for each target, and the exclusive expert output results and the shared expert output results of each target, the output results of the exclusive expert sub-model to be trained and the output results of the shared expert sub-model to be trained of each target are determined respectively.
示例性的，所述独占专家子模型包括多个独占专家组分别对应的第一独占专家门控网络，所述共享专家子模型包括一个共享专家组对应的第一共享专家门控网络；所述迭代训练模块还用于：Exemplarily, the exclusive expert sub-model includes first exclusive expert gating networks corresponding to a plurality of exclusive expert groups respectively, and the shared expert sub-model includes a first shared expert gating network corresponding to a shared expert group; the iterative training module is further used for:
通过所述第一独占专家门控网络，基于各个目标针对各个独占专家的第三差异权重对所述各个目标的独占专家输出结果进行加权处理，分别确定所述各个目标的待训练独占专家子模型输出结果；Through the first exclusive expert gating network, weighted processing is performed on the exclusive expert output results of each target based on the third difference weight of each target for each exclusive expert, and the output results of the exclusive expert sub-model to be trained for each target are determined respectively;
通过所述第一共享专家门控网络，基于所述各个目标针对各个共享专家的第三差异权重对所述各个目标的共享专家输出结果进行加权处理，分别确定所述各个目标的待训练共享专家子模型输出结果。Through the first shared expert gating network, the shared expert output results of each target are weighted based on the third difference weight of each target for each shared expert, and the output results of the shared expert sub-model to be trained for each target are determined respectively.
本申请模型训练装置具体实施方式与上述模型训练方法各实施例基本相同，在此不再赘述。The specific implementation of the model training device of the present application is basically the same as the embodiments of the above-mentioned model training method, and will not be repeated here.
此外，本申请还提供一种电子设备。如图6所示，图6是本申请实施例方案涉及的硬件运行环境的结构示意图。In addition, the present application also provides an electronic device. As shown in FIG6 , FIG6 is a schematic diagram of the structure of the hardware operating environment involved in the embodiment of the present application.
示例性的，图6即可为电子设备的硬件运行环境的结构示意图。For example, FIG6 is a schematic diagram of the structure of the hardware operating environment of the electronic device.
如图6所示，该电子设备可以包括处理器601、通信接口602、存储器603和通信总线604，其中，处理器601、通信接口602和存储器603通过通信总线604完成相互间的通信，存储器603，用于存放计算机程序；处理器601，用于执行存储器603上所存放的程序时，实现客户资源确定方法的步骤。As shown in Figure 6, the electronic device may include a processor 601, a communication interface 602, a memory 603 and a communication bus 604, wherein the processor 601, the communication interface 602 and the memory 603 communicate with each other through the communication bus 604, and the memory 603 is used to store computer programs; the processor 601 is used to implement the steps of the customer resource determination method when executing the program stored in the memory 603.
上述电子设备提到的通信总线604可以是外设部件互连标准(PeripheralComponent Interconnect,PCI)总线或扩展工业标准结构(Extended Industry StandardArchitecture,EISA)总线等。该通信总线604可以分为地址总线、数据总线和控制总线等。为便于表示，图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。The communication bus 604 mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus 604 can be divided into an address bus, a data bus, and a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
通信接口602用于上述电子设备与其他设备之间的通信。The communication interface 602 is used for communication between the above electronic device and other devices.
存储器603可以包括随机存取存储器(Random Access Memory,RMD),也可以包括非易失性存储器(Non-Volatile Memory,NM)，例如至少一个磁盘存储器。可选的，存储器603还可以是至少一个位于远离前述处理器601的存储装置。The memory 603 may include a random access memory (RMD) or a non-volatile memory (NM), such as at least one disk memory. Optionally, the memory 603 may also be at least one storage device located away from the processor 601.
上述的处理器601可以是通用处理器，包括中央处理器(Central ProcessingUnit,CPU)、网络处理器(Network Processor,NP)等；还可以是数字信号处理器(DigitalSignal Processor,DSP)、专用集成电路(Application Specific IntegratedCircuit,ASIC)、现场可编程门阵列(Field-Programmable Gate Array,FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。The above-mentioned processor 601 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
本申请电子设备具体实施方式与上述客户资源确定方法各实施例基本相同，在此不再赘述。The specific implementation of the electronic device of the present application is basically the same as the above-mentioned embodiments of the customer resource determination method, and will not be repeated here.
此外，本申请实施例还提出一种计算机可读存储介质，所述计算机可读存储介质上存储有客户资源确定程序，所述客户资源确定程序被处理器执行时实现如上所述的客户资源确定方法的步骤。In addition, an embodiment of the present application further proposes a computer-readable storage medium, on which a customer resource determination program is stored. When the customer resource determination program is executed by a processor, the steps of the customer resource determination method described above are implemented.
本申请计算机可读存储介质具体实施方式与上述客户资源确定方法各实施例基本相同，在此不再赘述。The specific implementation of the computer-readable storage medium of the present application is basically the same as the various embodiments of the above-mentioned customer resource determination method, and will not be repeated here.
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者系统不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者系统所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者系统中还存在另外的相同要素。It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。The serial numbers of the embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.
另外，在本申请说明书和所附权利要求书的描述中，术语“第一”、“第二”、“第三”等仅用于区分描述，而不能理解为指示或暗示相对重要性。还应理解的是，虽然术语“第一”、“第二”等在文本中在一些本申请实施例中用来描述各种元素，但是这些元素不应该受到这些术语的限制。这些术语只是用来将一个元素与另一元素区分开。例如，第一表格可以被命名为第二表格，并且类似地，第二表格可以被命名为第一表格，而不背离各种所描述的实施例的范围。第一表格和第二表格都是表格，但是它们不是同一表格。In addition, in the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish descriptions, and cannot be understood as indicating or suggesting relative importance. It should also be understood that although the terms "first", "second", etc. are used to describe various elements in some embodiments of the present application in the text, these elements should not be limited by these terms. These terms are only used to distinguish one element from another element. For example, the first table can be named as the second table, and similarly, the second table can be named as the first table without departing from the scope of the various described embodiments. The first table and the second table are both tables, but they are not the same table.
以上仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311871449.0ACN118096229A (en) | 2023-12-29 | 2023-12-29 | Customer resource determination method, model training method, electronic device and storage medium |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311871449.0ACN118096229A (en) | 2023-12-29 | 2023-12-29 | Customer resource determination method, model training method, electronic device and storage medium |
| Publication Number | Publication Date |
|---|---|
| CN118096229Atrue CN118096229A (en) | 2024-05-28 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311871449.0APendingCN118096229A (en) | 2023-12-29 | 2023-12-29 | Customer resource determination method, model training method, electronic device and storage medium |
| Country | Link |
|---|---|
| CN (1) | CN118096229A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN120234286A (en)* | 2025-05-30 | 2025-07-01 | 苏州元脑智能科技有限公司 | Data processing method |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN120234286A (en)* | 2025-05-30 | 2025-07-01 | 苏州元脑智能科技有限公司 | Data processing method |
| Publication | Publication Date | Title |
|---|---|---|
| US11227217B1 (en) | Entity transaction attribute determination method and apparatus | |
| CN111340244B (en) | Prediction method, training method, device, server and medium | |
| CN112163963B (en) | Service recommendation method, device, computer equipment and storage medium | |
| CN112580952A (en) | User behavior risk prediction method and device, electronic equipment and storage medium | |
| CN113407854A (en) | Application recommendation method, device and equipment and computer readable storage medium | |
| CN111951008A (en) | A risk prediction method, apparatus, electronic device and readable storage medium | |
| CN112446764A (en) | Game commodity recommendation method and device and electronic equipment | |
| JP2019040529A (en) | Correction device, correction method, and correction program | |
| CN111210255A (en) | Advertisement pushing method and device and electronic equipment | |
| CN118096229A (en) | Customer resource determination method, model training method, electronic device and storage medium | |
| CN113919921B (en) | A product recommendation method based on multi-task learning model and related equipment | |
| WO2025171818A1 (en) | Data processing method, apparatus and device, and computer-readable storage medium | |
| CN113034168A (en) | Content item delivery method and device, computer equipment and storage medium | |
| CN113743906A (en) | Method and device for determining service processing strategy | |
| CN112446763A (en) | Service recommendation method and device and electronic equipment | |
| CN111626789A (en) | House price prediction method, device, equipment and storage medium | |
| CN117853180A (en) | Data pricing method, device, equipment and medium based on reinforcement learning | |
| CN113822455A (en) | A time prediction method, device, server and storage medium | |
| CN113159877A (en) | Data processing method, device, system and computer readable storage medium | |
| CN116910373A (en) | House source recommendation method and device, electronic equipment and storage medium | |
| CN117035846A (en) | Information prediction method and device and related equipment | |
| KR102383509B1 (en) | System for matching space and company | |
| CN115759283A (en) | Model interpretation method and device, electronic equipment and storage medium | |
| CN115063145A (en) | Prediction method, device, electronic device and storage medium of transaction risk factor | |
| CN114971871A (en) | Method, device, apparatus, medium and program product for calculating a creditable amount |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |