Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
For easy understanding, the following describes the basic technical concept related to the embodiment of the present application:
artificial intelligence (Artificial Intelligence, AI): artificial intelligence is the theory, method, technique and application system that uses a digital computer or a digital computer-controlled machine to simulate, extend and expand human intelligence, sense the environment, acquire knowledge and use the knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning, automatic driving, intelligent traffic and other directions.
The business data processing scheme provided by the embodiment of the application particularly relates to an artificial intelligence Machine Learning (ML) technology. Machine learning is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and the like. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.
On-line advertising: may also be referred to as internet advertising, which refers to advertising placed on an advertising spot (e.g., social public number, news website, etc.) on an internet platform.
An advertiser: the entity for displaying the self advertisement through the advertisement space of the internet platform; the advertiser may be a legal person, other economic organization, or individual who promotes or services items, designs, makes, or distributes advertisements by itself or entrusts others.
An advertisement transaction platform: refers to an entity that connects the media owner and the advertiser, and the ad exchange may place the advertiser's ad on the ad spot provided by the media owner. In order to accurately deliver the advertiser's advertisement to the target object, the advertisement transaction platform generally collects information of the object, and thus can accurately deliver the advertisement according to the interest, geographic position or other data of the object.
Referring to fig. 1, fig. 1 is a schematic structural diagram of a network architecture provided in an embodiment of the present application, where the network architecture may include a server 10d and a terminal cluster, and the terminal cluster may include one or more terminal devices, where the number of terminal devices included in the terminal cluster is not limited. As shown in fig. 1, the terminal cluster may specifically include a terminal device 10a, a terminal device 10b, a terminal device 10c, and the like; all terminal devices in the terminal cluster (which may include, for example, terminal device 10a, terminal device 10b, and terminal device 10c, etc.) may be in network connection with the server 10d, so that each terminal device may interact with the server 10d through the network connection.
The terminal devices of the terminal cluster may include, but are not limited to: the application relates to electronic devices such as smart phones, tablet computers, notebook computers, palm computers, mobile internet devices (mobile internet device, MID), wearable devices (such as smart watches, smart bracelets and the like), intelligent voice interaction devices, intelligent household appliances (such as smart televisions and the like), vehicle-mounted devices, aircrafts and the like, and the type of terminal device is not limited. It will be appreciated that each terminal device in the terminal cluster shown in fig. 1 may be provided with an application client (internet platform), and when the application client runs in each terminal device, the application client may interact with the server 10d shown in fig. 1. The application client running in each terminal device may be an independent client, or may be an embedded sub-client integrated in a certain client, which is not limited in the present application.
The application client may specifically include, but is not limited to: a browser, a vehicle-mounted client, a smart home client, an entertainment client (e.g., a game client), a multimedia client (e.g., a video client), a social client, an information client (e.g., a news client), and the like. If the terminal device included in the terminal cluster is a vehicle-mounted device, the vehicle-mounted device may be an intelligent terminal in an intelligent traffic scene, and an application client running in the vehicle-mounted device may be referred to as a vehicle-mounted client.
The server 10d may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, a content delivery network (Content Delivery Network, CDN), basic cloud computing services such as big data and an artificial intelligence platform, and the type of the server is not limited in the present application.
It will be appreciated that the server 10d shown in fig. 1 may push service data for an application client running in the terminal device, or a data module (e.g., public number) integrated in the application client; in other words, the server 10d may filter out appropriate service data and send the service data to the user corresponding to the application client, that is, the user corresponding to the terminal device running the application client. The business data related to the embodiment of the application can be advertisement data (namely advertisement), wherein the advertisement data can be video data, audio data, image-text data, client data and the like, or can be game advertisement, commodity advertisement, application program advertisement, electronic reading advertisement, film advertisement and the like. For example, an advertiser may submit all of the advertising data that needs to be placed to an advertising exchange (e.g., which may be run in server 10 d) that may place the advertiser's advertising data onto an ad spot provided by a media owner (which may be referred to as an entity that provides an ad spot of an internet platform) in an application client. The service data pushed by the server 10d to the media host on the ad slots provided in the application client may be online advertising.
For ease of understanding, it is assumed that an application client running by a terminal device 10a in the terminal cluster (for ease of understanding, a user who will use the terminal device 10a is the object a) may display the service data B pushed by the server 10d in the application client after receiving the service data B. When the object a browses the service data B in the application client operated by the terminal device 10a, it means that the service data B is exposure service data, that is, a presentation operation of the service data B in the application client may be referred to as advertisement exposure.
The triggering operation performed by the object a on the service data B displayed in the application client (which may also be referred to as feedback information of the object a on the service data B) may include, but is not limited to, the following types: the first type can be advertisement clicking, and the advertisement clicking can represent clicking operation of the object A on the service data B; the second type may be advertisement shallow conversion, where the advertisement shallow conversion may represent shallow conversion operation on the service data B after the object a clicks, and the shallow conversion operation herein may be downloading, activating, registering, etc.; the third type is advertisement deep conversion, which may represent deep conversion operations on business data B after the click of object a, where the deep conversion operations may be payment, retention the next day, etc.
The shallow conversion operation is based on the click operation, namely the click operation is performed before the shallow conversion operation; the deep layer transformation operation is based on the shallow layer transformation operation, namely the shallow layer transformation operation is followed by the deep layer transformation operation. It should be understood that the embodiment of the present application may measure the above three types of triggering operations through different indicators. The ratio from advertisement exposure to advertisement click can be represented by CTR (Click Through Rate, click through rate, simply click rate), that is, the actual click-through amount of an advertisement divided by the display amount of the advertisement; the ratio from advertisement click-through to advertisement Conversion, i.e., the actual Conversion of an advertisement divided by the click-through of the advertisement, can be expressed by CVR (Conversion Rate). Wherein, the CVR can be divided into CVR1 (i.e. shallow conversion rate) and CVR2 (i.e. deep conversion rate), the ratio from advertisement click to advertisement shallow conversion can be expressed by CVR1, i.e. the actual shallow conversion of advertisement divided by the click of advertisement, and the ratio from advertisement click to advertisement deep conversion can be expressed by CVR2, i.e. the deep conversion of advertisement divided by the click of advertisement.
The service data processing scheme provided by the embodiment of the application particularly relates to index estimation of service data, such as click rate estimation, conversion rate estimation and the like. In the field of internet search service systems, when an object a inputs a query keyword in a content search page, a query result corresponding to the query keyword and service data pushed for the object a may be displayed in the content search page. The object A can browse and click the query result (can be called as a search natural result domain) and browse and click service data (can be called as a search advertisement result domain) in the content search page, and the index prediction model is trained jointly by introducing the operation data (feedback information) of the object A on the query result into the index modeling optimization of the advertisement system, namely the operation data of the object A on the query result and the operation data of the object A on the service data, so as to obtain the trained index prediction model; for ease of understanding, the index estimation model in the training stage may be referred to as an initial estimation model, and the trained index estimation model may be referred to as a target estimation model. The feedback index predicted value of each object to the service data can be predicted through the target prediction model which is completed through training, so that the prediction precision of the feedback index predicted value can be improved; in other words, the trained target estimation model only estimates the index of the service data.
The network structure of the initial pre-estimation model can comprise a service pre-estimation component and a result pre-estimation component. The service estimation component acts on the search advertisement result domain, namely the service estimation component can be constructed by utilizing the query keywords (sample query data), the object data (the object data in the training stage can be called sample object data), the service data pushed for the sample object (sample service data) and the feedback information of the sample object on the sample service data; the result pre-estimation component acts on the search natural result domain, namely the result pre-estimation component can be constructed by utilizing the query keywords (sample query data), the sample object data, the query results of the query keywords (sample query results) and the feedback information of the sample object on the sample query results. In other words, the sample query data and sample object data may be shared data of the business prediction component and the result prediction component.
It may be understood that the training and application processes of the index estimation model may be performed by a computer device, that is, the service data processing method provided in the embodiment of the present application may be performed by a computer device, where the computer device may be a server 10d in the network architecture shown in fig. 1, or any one of terminal devices in the terminal cluster, or may be a computer program (including program code, for example, an application client integrated by the terminal device), and the embodiment of the present application is not limited to this.
Referring to fig. 2, fig. 2 is a schematic diagram of a service data interaction scenario provided in an embodiment of the present application. The server 20a shown in fig. 2 may be the server 10d in the embodiment corresponding to fig. 1, and the terminal device 20b shown in fig. 2 may be any one of the terminal devices in the terminal cluster in the embodiment corresponding to fig. 1, and the terminal device 20b may install an application client. If the initial pre-estimation model is trained in the server 20a, the trained initial pre-estimation model may be referred to as a target pre-estimation model, and the target pre-estimation model may be published in an advertisement transaction platform integrated by the server 20a, so as to apply the target pre-estimation model in the advertisement recommendation field, where the service scenario in the advertisement recommendation field may specifically include, but is not limited to: content distribution scenes, content search scenes, content viewing scenes, etc., the specific business scenes will not be exemplified here one by one. It will be appreciated that the target object (user using the terminal device 20 b) may view the own target service data of interest pushed by the server 20a in the terminal device 20 b.
For example, in a content search scenario, when the target object performs a search operation on the application client, the application client may send a data acquisition request to the server 20a, so that the server 20a may return the query result and, at the same time, may also return the target service data associated with the query result, so that the application client may display the query result and the target service data at the same time. For another example, in the content distribution scenario, when the target object performs a startup operation on the application client, the application client may send a data acquisition request to the server 20a, so that the server 20a returns the target service data associated with the distribution content while returning the distribution content, and thus the application client may display the distribution content and the target service data at the same time. For another example, in the content viewing scenario, when the target object performs the information flow viewing operation on the application client, the application client may send a data acquisition request to the server 20a, so that the server 20a returns the target service data associated with the information flow content while returning the information flow content, and thus the application client may display the information flow content and the target service data at the same time.
For easy understanding, the embodiment of the present application is described by taking a content search scene as an example, and specific implementation processes of a content distribution scene, a content viewing scene, etc. may refer to the description of the content search scene, and no further description is given here.
As shown in fig. 2, when the target object inputs query data in the application client installed in the terminal device 20b, the terminal device 20b may acquire the query data input in the application client by the target object and transmit a data query request to the server 20 a. After receiving the data query request sent by the terminal device 20b, the server 20a may obtain a service data set (for example, the service data set 20d may include a plurality of service data, such as service data 21a, service data 21b, … …, service data 21n, and the like) from the service database 20c of the server 20 a. The service database 20c may be configured to store all service data (e.g., all advertisement data that the advertiser wants to deliver), and the server 20a may push, for the target object corresponding to the terminal device 20b, the target service data from the service data set 20d based on the target object and the query data input by the target object.
The server 20a may perform the same process for all traffic data in the traffic data set 20 d. The server 20a may obtain target object data corresponding to the target object and query data input by the target object, and take the query data, the target object data and the service data 21a as input data of a target prediction model, so as to convert the input data into service input features. The service input feature may be input to a trained target prediction model, through which a feedback indicator prediction value corresponding to the service input feature may be predicted, where the feedback indicator prediction value may be at least one of a trigger indicator prediction value (e.g., a predicted click rate) and a conversion indicator prediction value (e.g., a predicted conversion rate).
Each service data in the service data set 20d may be used as input data of a target prediction model together with query data input by a target object and target object data corresponding to the target object, so that a feedback index predicted value of the target object for each service data in the service data set 20d may be output through the target prediction model. For example, the feedback index pre-estimation of the target object for the service data 21a in the service data set 20d may be denoted as the feedback index pre-estimation 22a, the feedback index pre-estimation of the target object for the service data 21b may be denoted as the feedback index pre-estimation 22b, the feedback index pre-estimation of the target object for the service data 21c may be denoted as the feedback index pre-estimation 22c, … …, and the feedback index pre-estimation of the target object for the service data 21n may be denoted as the feedback index pre-estimation 22n. Further, the service data in the service data set 20d may be ordered according to the order of the feedback indicator predicted values from large to small, so that the previous service data or the previous service data may be used as target service data, and the target service data may be returned to the terminal device 20b. It will be appreciated that the server 20a may also query the target query result associated with the query data input by the target object, where the target query result may be one or more pieces of query document information, and the target query result may also be returned to the terminal device 20b, that is, the target service data and the target query result may be simultaneously displayed in the application client of the terminal device 20b.
It should be noted that, before the target pre-estimation model is applied to the advertisement recommendation field, the initial pre-estimation model needs to be trained, and the trained initial pre-estimation model can be used as the target pre-estimation model. In the field of advertisement recommendation, after an object inputs a search keyword (query data), in a content search page, the object may perform not only a corresponding operation (e.g., exposure, click on a query result) corresponding to the search keyword, but also a corresponding operation (e.g., exposure, click on business data) associated with the search keyword (e.g., advertisement data); therefore, in the training stage of the initial pre-estimation model, the operation data based on the service data and the operation data of the query result can be jointly modeled, and the search keywords, the object data and the like are used as shared features, so that the target pre-estimation model can be obtained through training in a joint modeling mode, and the training process of the initial pre-estimation model is described below with reference to fig. 3 and 4.
Referring to fig. 3, fig. 3 is a flowchart illustrating a service data processing method according to an embodiment of the present application; it will be appreciated that the service data processing method may be performed by a computer device, which may be a server, or may be a terminal device, which is not limited in this regard. As shown in fig. 3, the service data processing method may include the following steps S101 to S104:
Step S101, sample object data and sample display data associated with sample query data are acquired; the sample presentation data comprises sample query results and sample service data, and carries sample tag information.
In the embodiment of the application, in the training stage of the initial pre-estimation model, the computer equipment can construct a training data set (or may be called a training data stream) of the initial pre-estimation model, and each training sample in the training data set can be composed of one sample query data, sample object data corresponding to the sample query data, sample service data and sample query results corresponding to the sample query data, service data labels corresponding to the sample service data and query result labels corresponding to the sample query results; the business data label and the query result label can be used as sample label information corresponding to the training samples. In other words, for each search query, a training sample of the initial pre-estimation model can be constructed by combining operation data (feedback information) of the user on the display data in the content search page, and sample label information corresponding to the training sample is used for representing the feedback information of the user on the display data; the presentation data in the content search page may include results of the user search query (query results) as well as pushed business data.
For example, a search keyword (e.g., "peony") input in searching a query may be referred to as sample query data, a user inputting the search keyword may be referred to as a sample object, a search result corresponding to the search keyword may be referred to as a sample query result, service data associated with the search keyword may be referred to as sample service data, feedback information of the sample object for the sample query result may be used to label a query result label, and feedback information of the sample object for the sample service data may be used to label a service data label. The sample service data may refer to advertisement data pushed by the advertisement recommendation system for a sample object, and the sample query result may refer to a query result retrieved by the sample query data.
In a content search scenario, the process of constructing the training samples of the initial predictive model may include, but is not limited to: sample query data input by the sample object in the content search page can be obtained, sample query results and sample service data corresponding to the sample query data are obtained in the content search page, and the sample query data and the sample service data are determined to be sample presentation data corresponding to the sample query data. Further, according to the first feedback information of the sample object to the sample query result (i.e. the interactive operation data between the sample object and the sample query result) and the second feedback information of the sample object to the sample service data (i.e. the interactive operation data between the sample object and the sample service data), the sample label information corresponding to the sample display data can be determined; and acquiring the basic attribute and the operation attribute corresponding to the sample object, and determining the basic attribute and the operation attribute as sample object data corresponding to the sample object.
It is to be appreciated that sample business data can be used to represent advertisements displayed in a content search page, which sample business data can include, but is not limited to, business data content (e.g., business data text, business data images, etc.), business identifications (e.g., advertisement identifications), business data object identifications (e.g., advertiser identifications), business data categories (e.g., application advertisements), and so forth; the basic attributes may be used to represent personal basic information corresponding to the sample object, and the operational attributes may be used to represent interest preferences, behavior operations, etc. of the sample object.
The sample service data may be service data in which the sample object has a trigger behavior (clicking operation) in a target period after the search keyword is input, or may be service data in which the sample object does not have a trigger behavior in a target period after the search keyword is input; alternatively, the sample service data may be service data in which the sample object has a forwarding behavior in a target period after the search keyword is input, or may be service data in which the sample object does not have a forwarding behavior in a target period after the search keyword is input. The sample query result may be a query result of the sample object having a trigger behavior in a target period after the search keyword is input, or may be a query result of the sample object not having a trigger behavior in a target period after the search keyword is input; alternatively, the sample query result may be a query result in which the sample object has a forwarding behavior in a target period after the search keyword is input, or may be a query result in which the sample object does not have a forwarding behavior in a target period after the search keyword is input. The target period may be any period of time after the sample object inputs the search keyword and performs the search operation, for example, 5 minutes, 10 minutes, half an hour, etc., which is not limited in the present application.
In one or more embodiments, based on feedback information (operation data) of sample service data and sample query results associated with sample query data by a sample object, a training sample of an initial pre-estimation model may be labeled, where the labeled training sample may be at least divided into the following cases:
(1) if the sample object executes the trigger action on the sample service data and the query result corresponding to the sample query data, any query result of the sample object executing the trigger action may be used as the sample query result, the sample service data and the sample tag information corresponding to the sample query result at this time may be labeled as (1, 1), and the sample query result and the sample service data with the sample tag information of (1, 1) may be returned to the training data stream (for training the initial pre-estimation model, such as the training data set described above).
(2) If the sample object performs a triggering action on the sample service data corresponding to the sample query data and does not perform a triggering action on the query result corresponding to the sample query data, the first query result displayed in the content search page may be used as the sample query result, at this time, the sample service data and sample tag information corresponding to the sample query result may be marked as (1, 0), and the sample query result and sample service data with sample tag information of (1, 0) may be transmitted back to the training data stream.
(3) If the sample object does not execute the triggering action on the sample service data corresponding to the sample query data and executes the triggering action on the query result corresponding to the sample query data, any one query result of the sample object executing the triggering action is taken as a sample query result, at this time, the sample service data and sample label information corresponding to the sample query result can be marked as (0, 1), and the sample query result and sample service data with the sample label information of (0, 1) can be transmitted back to the training data stream.
(4) If the sample object does not execute the triggering action on the sample service data and the sample query data corresponding to the sample query data, the first query result displayed in the content search page is taken as a sample query result, sample label information corresponding to the sample service data and the sample query result at the moment can be marked as (0, 0), and the sample query result and the sample service data with the sample label information of (0, 0) can be transmitted back to the training data stream.
It should be noted that, a value of "1" in the sample tag information indicates that the sample object performs the trigger behavior, and a value of "0" in the sample tag information indicates that the sample object does not perform the trigger behavior. That is, the sample label information corresponding to the training sample of the initial predictive model may be divided into the above four types, that is, four sample label information such as (1, 1), (1, 0), (0, 1) and (0, 0), so that the target predictive model may be obtained by training and used for predicting the trigger index evaluation value (for example, the predictive click rate) corresponding to the service data. Optionally, in the embodiment of the present application, sample tag information may be labeled for the sample query result and the sample service data based on whether the sample object performs a transformation action on the sample query result and the sample service data, where the sample tag information may be equally divided into the four types, and thus, the target prediction model obtained by training may be used to predict the transformation index evaluation value corresponding to the service data.
Step S102, determining a first sample input feature according to the sample query data, the sample object data and the sample service data, and determining a second sample input feature according to the sample query data, the sample object data and the sample query result.
Specifically, the initial estimation model may include a service estimation component and a result estimation component, where the service estimation component may be obtained by modeling based on sample object data, sample query data, sample service data, and a service data tag corresponding to the sample service data, and the result estimation component may be obtained by modeling based on sample object data, sample query result, and a query result tag corresponding to the sample query result. In other words, the sample object data, the sample query data may be shared by the service estimation component and the result estimation component, that is, in a training stage of the initial estimation model, the sample object data, the sample query data, and the sample service data may be input data of the service estimation component, and the sample object data, the sample query data, and the sample query result may be input data of the result estimation component.
Further, the input data of the business pre-estimation component can be converted into a first sample input feature, and the input data of the result pre-estimation component can be converted into a second sample input feature. The process of obtaining the first sample input feature corresponding to the service estimating component may include: sample query vectors corresponding to the sample query data and sample object vectors corresponding to the sample object data can be obtained; and then the sample query vector and the sample object vector can be combined into a sample joint vector, and the sample joint vector is subjected to pooling operation to obtain the component sharing characteristic. The sample query data may be a query text, or may be a query image, etc.; the sample object data may include information such as a base attribute and an operation attribute corresponding to the sample object.
Optionally, if the sample query data is a query text, the query text may be divided into a plurality of unit characters, and each unit character may be further converted into a unit word vector, where the unit word vector may be a sparse vector obtained by one-hot (one-hot) encoding, that is, only one non-zero vector, and the other elements are vectors of zero elements; the text sparse vector corresponding to the query text can be obtained by splicing the unit word vectors corresponding to the plurality of unit characters, the text sparse vector can be input into an Embedding (Embedding) layer, the input text sparse vector can be converted into the text dense vector through the Embedding layer, the dimension of the text dense vector is far smaller than that of the text sparse vector, and the text dense vector at the moment can be used as a sample query vector. If the sample query data is a query image, image preprocessing (for example, graying processing, size adjustment, etc.) may be performed on the query image to obtain an image input matrix corresponding to the query image, where the image input matrix may be used as a sample query vector.
For sample object data corresponding to the sample object, information such as basic attributes, operation attributes and the like contained in the sample object data can be encoded into an object sparse vector, and further, the text sparse vector can be converted into an object dense vector through an embedding layer, and the object dense vector can be used as a sample object vector. And then the sample query vector and the sample object vector can be added (or spliced) to obtain a sample joint vector; and inputting the sample joint vector into a pooling layer, and carrying out pooling operation on the sample joint vector through the pooling layer to obtain the component sharing characteristic. The pooling operation may be an average pooling operation, or may be a maximum pooling operation, which is not limited in the present application.
For sample service data corresponding to the sample object, the process of obtaining a sample service vector corresponding to the sample service data may include: the service attribute and the service data content (for example, text content or image data) corresponding to the sample service data can be obtained, the service attribute is converted into the service attribute characteristic, and the service data content is converted into the service content characteristic; and further, the service attribute features and the service content features can be combined into initial feature vectors, and the dimension reduction processing is carried out on the initial feature vectors to obtain sample service vectors corresponding to the sample service data. The initial feature vector may be a sparse vector, the sample service vector may be a dense vector obtained by performing dimension reduction processing on the sparse vector through the embedding layer, and the sample service vector obtaining process may refer to the sample query vector obtaining process, which is not described herein. Further, the sample service vector and the component sharing feature may be spliced to a first sample input feature corresponding to the service pre-estimation component.
For a result pre-estimation component in the initial pre-estimation model, the computer equipment can acquire a query result vector corresponding to a sample query result, the query result vector can be a search document feature, and the acquisition process of the query result vector can refer to the acquisition process of the sample query vector and is not repeated here; and then the query result vector and the component sharing feature can be spliced into a second sample input feature corresponding to the result pre-estimating component.
Step S103, outputting a first index predicted value corresponding to the first sample input feature through a service pre-estimation component in the initial pre-estimation model, and outputting a second index predicted value corresponding to the second sample input feature through a result pre-estimation component in the initial pre-estimation model.
Specifically, the first sample input feature can be input to a service estimation component in the initial estimation model, and index prediction is performed on the first sample input feature through the service estimation component, so as to obtain a first index predicted value corresponding to the first sample input feature. Meanwhile, the second sample input features can be input to a result prediction component in the initial prediction model, and index prediction is carried out on the second sample input features through the result prediction component to obtain second index predicted values corresponding to the second sample input features. The first indicator predictive value and the second indicator predictive value may be trigger indicator predictive values, or may be conversion indicator predictive values (the conversion indicator predictive values may include shallow conversion indicator predictive values and deep conversion indicator predictive values). It may be understood that the service estimation component and the result estimation component in the initial estimation model may have the same network structure, that is, the calculation process of the first sample input feature in the service estimation component is the same as the calculation process of the second sample input feature in the result estimation component, for example, the service estimation component and the result estimation component may be network structures such as a multi-layer perceptron (may also be referred to as an artificial neural network), a convolutional neural network, a deep learning network, and the like, and the network structure of the initial estimation model is not limited in the present application.
Wherein the calculating process of the first sample input feature in the service estimating component can comprise: after the first sample input feature is input to a service estimation component in an initial estimation model, full-connection processing can be performed on the first sample input feature through an implicit network layer in the service estimation component to obtain a full-connection index feature (which can be called as a first full-connection index feature for easy understanding); and the first full-connection index feature can be input to an activation network layer in the service estimation component, the activation network layer is used for activating the first full-connection index feature to obtain a service index activation feature, and a first index predicted value corresponding to the first sample input feature is determined according to the service index activation feature. The service estimating component may be composed of a multi-layer perceptron and an activating function, the multi-layer perceptron may be composed of an input layer, an hidden layer and an output layer, the network connection mode among the input layer, the hidden layer and the output layer may be a full connection mode, and the hidden network layer may be referred to as the hidden layer in the multi-layer perceptron.
Wherein the calculating of the second sample input feature in the result pre-estimation component may comprise: after the second sample input feature is input to a result pre-estimation component in the initial pre-estimation model, full-connection processing can be performed on the second sample input feature through an implicit network layer in the result pre-estimation component, so as to obtain a full-connection index feature (which can be called as a second full-connection index feature for easy understanding); and inputting the second full-connection index feature into an activation network layer in the result pre-estimation component, performing activation processing on the second full-connection index feature through the activation network layer to obtain a result index activation feature, and determining a second index pre-estimation value corresponding to the second sample input feature according to the result index activation feature.
Step S104, carrying out parameter adjustment on the initial pre-estimated model according to the first index pre-estimated value, the second index pre-estimated value and the sample label information, and determining a target pre-estimated model based on the initial pre-estimated model after parameter adjustment; the target pre-estimation model is used for predicting a feedback index pre-estimation value of the target object on the service data.
Specifically, the sample tag information may include a service data tag corresponding to sample service data and a query result tag corresponding to a sample query result. According to the first index estimated value and the service data label, a service loss value corresponding to the service estimated component can be determined; the determining method of the service loss value may include: carrying out logarithmic processing on the first index predicted value to obtain an information quantity (for convenience of understanding, the information quantity can be called as a first information quantity) associated with the first index predicted value; and determining a service loss value corresponding to the service estimating component according to the product between the service data label and the first information quantity. The service loss value can be calculated by the following formula (1):
wherein, lossa Representing a traffic loss function, f1 (-) represents a business pre-estimation component, xa Representing a first sample input characteristic of a business pre-estimation component, N representing an initial pre-run Estimating the number of training samples of the model for each training, y1i Representing the business data label corresponding to the ith training sample, e.g. y1i Can take on a value of 0, or can take on a value of 1, log (f1 (xa ))、log(1-f1 (xa ) A) may be expressed as a first amount of information. It should be noted that the number of prediction categories of the service prediction component may be 2: triggered and not triggered (exposure only).
Further, according to the second index predicted value and the query result label, a result loss value corresponding to the result pre-estimating component can be determined; the determining method of the result loss value may include: carrying out logarithmic processing on the second index predicted value to obtain an information quantity (for convenience of understanding, the information quantity can be called as a second information quantity) associated with the second index predicted value; and determining a result loss value corresponding to the result pre-estimation component according to the product between the query result label and the second information quantity. The resulting loss value can be calculated by the following formula (2):
wherein, lossd Representing the resulting loss function, f2 (-) represents the result pre-estimation component, xd Representing the second sample input characteristics of the result estimation component, N represents the number of training samples of the initial estimation model which are trained each time, y2i Representing the query result label corresponding to the ith training sample, e.g. the y2i Can take on a value of 0, or can take on a value of 1, y2i log(f2 (xd ))、y2ui log(f2 (xd ) A) may represent a second amount of information.
Further, the service loss value and the result loss value can be weighted and summed to obtain a model total loss corresponding to the initial pre-estimated model, parameter adjustment is performed on the initial pre-estimated model according to the model total loss, and the service pre-estimated component in the initial pre-estimated model after the parameter adjustment is determined to be the target pre-estimated model. Wherein, the total loss of the model can be shown in the following formula (3):
Loss＝αLossa +βLossd (3)
the Loss represents the total Loss of the model, alpha and beta are parameters, alpha is used for controlling a service Loss value, and beta is used for controlling a result Loss value.
And (3) carrying out parameter adjustment on the initial pre-estimated model through the total model Loss shown in the formula (3), for example, carrying out minimum optimization treatment on the total model Loss, carrying out back propagation on the initial pre-estimated model, continuously training network parameters in the initial pre-estimated model, storing the network parameters when the training times of the initial pre-estimated model reach the preset maximum iteration times, and determining a service pre-estimated assembly containing the current network parameters as a target pre-estimated model after training.
Referring to fig. 4, fig. 4 is a schematic structural diagram of an initial estimation model in a training stage according to an embodiment of the present application. As shown in fig. 4, in the training stage of the initial pre-estimation model, training samples of the initial pre-estimation model may be composed of sample query data, sample object data, sample query results and sample service data, and each training sample may carry sample tag information, where the sample tag information may include a service data tag corresponding to the sample service data and a query result tag corresponding to the sample query results. The initial predictive model may include a business predictive component 30a and a result predictive component 30g.
The computer device may combine the sample query data and the sample object data in the training samples, and convert the combined data into sparse vectors 30m, and further may perform dimension reduction processing on the sparse vectors 30m through the embedding layer to obtain sample joint vectors 30f (the sample joint vectors 30f are dense vectors), and perform pooling operation on the sample joint vectors 30f to obtain component sharing features. In the same way, the computer device may convert the sample traffic data in the training samples into sparse vectors 30b, and may further convert the sparse vectors 30b into sample traffic vectors 30c through the embedding layer, and the component sharing features and the sample traffic vectors 30c may be combined into a first sample input feature and input to the multi-layer perceptron 30d. The business index activation feature 30e may be output by the multi-layer perceptron 30d and activation function, and a first index prediction value may be predicted based on the business index activation feature 30 e.
Similarly, the sample query result in the training sample may be converted into the sparse vector 30h in the same manner, and then the sparse vector 30h may be converted into the query result vector 30i through the embedding layer, and the component sharing feature and the query result vector 30i may be combined into the second sample input feature and input to the multi-layer perceptron 30j. The multi-layer perceptron 30j and the activation function may output a resulting indicator activation feature 30k, based on which a second indicator estimate may be predicted. Further, the total model loss corresponding to the initial pre-estimated model can be determined based on the first index pre-estimated value and the service data label, and the second index pre-estimated value and the query result label, the network parameters of the initial pre-estimated model are subjected to continuous iterative adjustment based on the total model loss, and the trained service pre-estimated assembly 30a can be used as a target pre-estimated model; the sample predictive model may be used to predict feedback index predictors (e.g., predictive click rate, or predictive conversion rate, etc.) of the sample object to the business data.
In the embodiment of the application, sample object data, sample query results and sample service data associated with sample query data can be obtained, and the sample query data, the sample object data and the sample service data can be combined into first sample input data and input into a service prediction component in an initial prediction model to obtain a first index predicted value; the sample query data, the sample object data and the sample query result can be combined into a second sample input result, and the second sample input result is input to a result pre-estimation component in the initial pre-estimation model to obtain a second index pre-estimation value; and further, based on the first index predicted value, the second index predicted value and the sample label information, the network parameters of the initial pre-estimated model can be trained, and the trained service pre-estimated assembly can be used as a target pre-estimated model. In the training process of the initial pre-estimation model, the initial pre-estimation model can be trained by utilizing the interaction between the sample object and the sample query result and the interaction between the sample object and the sample service data, so that the feature expression and the preference feature of the user side can be more accurately represented, the prediction precision of the index pre-estimation value corresponding to the service data can be improved, the accurate delivery of the service data is facilitated, and the benefit of the service data can be improved.
Referring to fig. 5, fig. 5 is a second flow chart of a service data processing method according to an embodiment of the present application; it will be appreciated that the service data processing method may be performed by a computer device, which may be a server, or may be a terminal device, which is not limited in this regard. As shown in fig. 5, the service data processing method may include the following steps S201 to S204:
step S201, acquiring target query data and target object data corresponding to the target object, and acquiring a target joint vector corresponding to the target object according to the target query number and the target object data.
Specifically, after the initial pre-estimation model is trained, the trained service pre-estimation component can be determined to be a target pre-estimation model, and the target pre-estimation model can be deployed in the advertisement recommendation system so as to improve the accuracy of the service system in the advertisement recommendation system. When the target object inputs target query data in the content search page and performs search operation, the target query data input by the target object may be acquired, and the base attribute and the history operation data corresponding to the target object may be referred to as target object data. Further, the target query data may be converted into a target query vector, and the target object data may be converted into a target object vector, and the target query vector and the target object vector may be combined into a target joint vector. The process of obtaining the target joint vector may refer to the process of obtaining the sample joint vector described in step S102, and will not be described herein.
Step S202, obtaining candidate service vectors corresponding to each candidate service data in the candidate service data set, and pairing and combining the target joint vector and the candidate service vectors corresponding to each candidate service data to obtain service input features corresponding to each candidate service data.
Specifically, a candidate service data set (for example, an advertisement queue to be put) associated with the target object may be obtained from the service database, where all candidate service data in the candidate service data set have a certain possibility of being pushed to the service data of the target object; for each candidate service data in the candidate service data set, a corresponding candidate service vector may be acquired, and the process of acquiring the candidate service vector is similar to the process of acquiring the sample service vector described in the foregoing step S102, which is not described herein. Further, the target joint vector and the candidate service vector corresponding to each candidate service data can be paired and combined to obtain the service input characteristics corresponding to each candidate service data respectively. For example, each candidate service data in the candidate service data set may be spliced with the target joint vector separately, so that a service input feature corresponding to each candidate service data may be obtained.
Step S203, the business input features corresponding to the candidate business data are sequentially input into the target pre-estimation model, and the feedback index pre-estimation values corresponding to the candidate business data are output through the target pre-estimation model.
Specifically, the service input features corresponding to each candidate service data may be sequentially input to the target prediction model, and the feedback index predicted value of the target object to each candidate service data may be predicted by the target prediction model. The processing of the target prediction model on the service input features is the same as the processing of the service prediction component in the training stage on the first sample input features, and no description is repeated here.
Referring to fig. 6, fig. 6 is a schematic structural diagram of an objective prediction model at an application stage according to an embodiment of the present application. As shown in fig. 6, in the application process of the target prediction model, the computer device may combine the target query data and the target object data, and convert the combined data into a sparse vector 40g, and further may perform a dimension reduction process on the sparse vector 40g through the embedding layer to obtain a target joint vector 40f (the target joint vector 40f is a dense vector), and perform a pooling operation on the target joint vector 40f, so as to obtain a pooled target joint vector. Similarly, the computer device may convert each candidate service data in the candidate service data set into a sparse vector 40b in the same manner, and may further convert the sparse vector 40b into a target service vector 40c through the embedding layer, where the pooled target joint vector and the target service vector 40c may be combined into a service input feature, and the service input feature may be input to the multi-layer perceptron 40d. The service index activation feature 40e can be output through the multi-layer perceptron 40d and the activation function, and a feedback index predicted value of the target object to the service candidate data can be predicted and obtained based on the service index activation feature 40 e.
Step S204, determining target business data for displaying to the target object in the candidate business data set according to the feedback index predicted value.
Specifically, according to the sequence from the large feedback index predicted value to the small feedback index predicted value, the candidate service data with the largest feedback index predicted value in the candidate service data set can be used as target service data pushed to a target object.
In one or more embodiments, the hypothetical feedback indicator predictive value may include a trigger indicator predictive value (e.g., a predicted click rate) and a conversion indicator predictive value (e.g., a predicted conversion rate); the computer device may obtain a unit conversion value (e.g., advertisement bid, that is, a price of an advertiser for an advertisement bid, which may be generally a converted price) corresponding to each candidate service data in the candidate service data set, and determine a service weight value (e.g., advertisement quality score, where the higher the quality score is, the more likely to be pushed to the target object) corresponding to each candidate service data according to historical feedback information (e.g., historical operation data of the historical object on the candidate service data, social activity of each historical object, etc.) corresponding to each candidate service data; determining recommended evaluation values corresponding to the candidate service data respectively according to the unit conversion value, the service weight value, the trigger index evaluation value and the conversion index evaluation value; and determining the candidate business data corresponding to the maximum recommended evaluation value as target business data for displaying to the target object in the candidate business data set.
Alternatively, the calculation process of the recommendation evaluation value may include, but is not limited to: and determining the comprehensive conversion value corresponding to each candidate service data respectively based on the product among the unit conversion value, the trigger index predicted value and the conversion index predicted value. For example, the integrated conversion value may be expressed as ecpm=bid×pctr×pcvr, and eCPM (Effective Cost Per Mille) may refer to a cost that an advertiser needs to pay after displaying an advertisement to one thousand access users on an internet platform, which is referred to as the integrated conversion value in the embodiment of the present application; bid represents an advertising bid; pCTR (predict click through rate) indicates a predicted click rate, corresponding to CTR, which means that the probability of the advertisement being clicked is predicted after the advertisement is put under a certain condition; pCVR (predict conversion rate) indicates estimated conversion rate, which means the probability of converting an advertisement after the advertisement is clicked in a certain situation. Further, based on the sum of the comprehensive conversion value and the service weight value, determining recommendation evaluation values respectively corresponding to the candidate service data; for example, the recommendation evaluation value may be expressed as: score=ecpm+quality, score indicates a recommendation evaluation value (ranking score), and quality indicates a traffic weight value (advertisement quality score).
It can be understood that after the comprehensive conversion values corresponding to the candidate service data are obtained, the candidate service data in the candidate service data set can be subjected to rough sorting (called rough sorting for short) according to the order of the comprehensive conversion values from large to small, and the previous L (L is an integer greater than 1, for example, L can take the value of 5 or 10, etc.) candidate service data are obtained from the candidate service data set based on the rough sorting; and the sum of the comprehensive conversion value and the service weight value can be used as a recommendation evaluation value (score), the selected L candidate service data are carefully selected and ordered based on the recommendation evaluation value, and the forefront candidate service data are used as target service data pushed to a target object. The rough sorting is to search advertisements conforming to the direction according to the advertisement direction (region, age and the like), calculate pCVR and pCTR values in each advertisement rough sorting model and real-time Bid of the advertisements by combining object information and advertisement information, sort the advertisements according to eCPM of the advertisements, and select the top L advertisements to return to carefully selected advertisements. Carefully chosen ranking refers to determining the ranking of advertisements in an advertisement queue based on the ranking score (recommendation rating value) of the advertisements, wherein the higher the ranking score, the higher the ranking in the carefully chosen advertisement queue.
Referring to fig. 7, fig. 7 is a schematic diagram of a recommendation flow of target service data according to an embodiment of the present application. As shown in fig. 7, when the target object inputs target query data in the content search page, the target query data input by the target object and the target object data corresponding to the target object may be acquired, and further, a candidate service data set 50a associated with the target object may be acquired (as shown in fig. 7, the candidate service data set 50a may include service data 51a, service data 51b, … …, service data 51n, and the like). For each service data in the candidate service data set 50a, input data of the target estimated model after training can be formed by the target query data and the target object data.
The input data is converted into service input characteristics, and the feedback index predicted value of the target object to each service data in the candidate service data set 50a can be output through the target pre-estimation model. If the feedback indicator pre-estimates may include a trigger indicator pre-estimate and a conversion pre-estimate, the trigger indicator pre-estimates corresponding to each of the candidate business data sets 50a may be referred to as a trigger indicator pre-estimate set 50b (e.g., business data 51a corresponding to the trigger pre-estimate 52a, business data 51b corresponding to the trigger pre-estimates 52b, … …, business data 51n corresponding to the trigger pre-estimates 52n, etc.), and the conversion indicator pre-estimates corresponding to each of the business data sets 50a may be referred to as a conversion indicator pre-estimate set 50c (e.g., business data 51a corresponding to the conversion pre-estimates 53a, business data 51b corresponding to the conversion pre-estimates 53b, … …, business data 51n corresponding to the conversion pre-estimates 53n, etc.). Of course, the unit conversion value corresponding to each service data in the candidate service data set 50a may be obtained, and the unit conversion value of each service data may be set as the unit conversion value set 50d (for example, the unit conversion value 54a corresponding to the service data 51a, the unit conversion values 54b and … … corresponding to the service data 51b, the unit conversion value 54n corresponding to the service data 51n, and the like).
According to the trigger index pre-estimation set 50b, the conversion index pre-estimation set 50c and the unit conversion value set 50d, the comprehensive conversion value corresponding to each service data in the candidate service data set 50a is calculated, so that the service weight value corresponding to each service data in the candidate service data set 50a can be obtained, the recommendation evaluation value set 50e is obtained based on the comprehensive conversion value corresponding to each service data and the service weight value corresponding to each service data, and the service data with the maximum recommendation evaluation value in the recommendation evaluation value set 50e is determined to be the target service data for pushing to the target object. Service data is pushed for the target object through the target pre-estimation model, so that the accuracy of service data delivery can be improved.
In the embodiment of the application, in the training process of the initial pre-estimation model, the initial pre-estimation model can be trained by utilizing the interaction between the sample object and the sample query result and the interaction between the sample object and the sample service data, so that the feature expression and the preference feature of the user side can be more accurately represented, the target pre-estimation model obtained by training can be used for predicting the feedback index pre-estimation value of the target object to each candidate service data, thereby completing the accurate delivery of the target service data, improving the prediction precision of the index pre-estimation value corresponding to the service data, and improving the income of the service data.
It will be appreciated that in the specific embodiment of the present application, related data such as content of a user search operation, object data (base attribute, operation attribute) and the like may be involved, and when the above embodiments of the present application are applied to specific products or technologies, permission or consent of the user and the like needs to be obtained, and collection, use and processing of related data need to comply with related laws and regulations and standards of related countries and regions.
Referring to fig. 8, fig. 8 is a schematic structural diagram of a service data processing apparatus according to an embodiment of the present application. As shown in fig. 8, the service data processing apparatus 1 includes: the sample data acquisition module 11, the input feature determination module 12, the sample estimated value output module 13 and the model parameter adjustment module 14;
a sample data acquisition module 11 for acquiring sample object data and sample presentation data associated with sample query data; the sample display data comprises sample query results and sample service data, and carries sample tag information;
an input feature determining module 12, configured to determine a first sample input feature according to the sample query data, the sample object data, and the sample service data, and determine a second sample input feature according to the sample query data, the sample object data, and the sample query result;
The sample pre-estimation value output module 13 is configured to output a first index pre-estimation value corresponding to a first sample input feature through a service pre-estimation component in the initial pre-estimation model, and output a second index pre-estimation value corresponding to a second sample input feature through a result pre-estimation component in the initial pre-estimation model;
the model parameter adjustment module 14 is configured to perform parameter adjustment on the initial pre-estimated model according to the first index pre-estimated value, the second index pre-estimated value and the sample tag information, and determine a target pre-estimated model based on the initial pre-estimated model after parameter adjustment; the target pre-estimation model is used for predicting a feedback index pre-estimation value of the target object on the service data.
The functional implementation manners of the sample data obtaining module 11, the input feature determining module 12, the sample pre-estimation value output module 13, and the model parameter adjusting module 14 may refer to steps S101 to S104 in the embodiment corresponding to fig. 3, and will not be described herein.
In one or more embodiments, the sample data acquisition module 11 includes: a sample presentation data acquisition unit 111, a feedback information acquisition unit 112, a sample object data acquisition unit 113;
a sample presentation data obtaining unit 111, configured to obtain sample query data input by a sample object in a content search page, obtain a sample query result and sample service data corresponding to the sample query data in the content search page, and determine the sample query data and the sample service data as sample presentation data corresponding to the sample query data;
The feedback information obtaining unit 112 is configured to determine sample tag information corresponding to sample display data according to first feedback information of a sample object on a sample query result and second feedback information of the sample object on sample service data;
the sample object data obtaining unit 113 is configured to obtain a basic attribute and an operation attribute corresponding to a sample object, and determine the basic attribute and the operation attribute as sample object data corresponding to the sample object.
The functional implementation manners of the sample presentation data obtaining unit 111, the feedback information obtaining unit 112, and the sample object data obtaining unit 113 may refer to step S101 in the embodiment corresponding to fig. 3, and will not be described herein.
In one or more embodiments, the input feature determination module 12 determines a first sample input feature from the sample query data, the sample object data, and the sample business data, including:
a first vector obtaining unit 121, configured to obtain a sample query vector corresponding to sample query data, and obtain a sample object vector corresponding to sample object data;
the shared feature obtaining unit 122 is configured to combine the sample query vector and the sample object vector into a sample joint vector, and perform a pooling operation on the sample joint vector to obtain a component shared feature;
A second vector obtaining unit 123, configured to obtain a sample service vector corresponding to the sample service data, and combine the sample service vector and the component sharing feature into a first sample input feature.
Optionally, the second vector obtaining unit 123 obtains a sample service vector corresponding to the sample service data, including:
acquiring service attributes and service data contents corresponding to sample service data, converting the service attributes into service attribute characteristics, and converting the service data contents into service content characteristics;
and combining the service attribute characteristics and the service content characteristics into initial characteristic vectors, and performing dimension reduction processing on the initial characteristic vectors to obtain sample service vectors corresponding to the sample service data.
The functional implementation of the first vector obtaining unit 121, the shared feature obtaining unit 122, and the second vector obtaining unit 123 may refer to step S102 in the embodiment corresponding to fig. 3, which is not described herein.
In one or more embodiments, the sample pre-estimation value output module 13 outputs, through a result pre-estimation component in the initial pre-estimation model, a second index pre-estimation value corresponding to a second sample input feature, including:
the index feature obtaining unit 131 is configured to input the second sample input feature to a result prediction component in the initial prediction model, and perform full-connection processing on the second sample input feature through an implicit network layer in the result prediction component to obtain a full-connection index feature;
The index feature activation unit 132 is configured to input the full-connection index feature to an activation network layer in the result prediction component, perform activation processing on the full-connection index feature through the activation network layer, obtain a result index activation feature, and determine a second index predicted value corresponding to the second sample input feature according to the result index activation feature.
The functional implementation of the index feature obtaining unit 131 and the index feature activating unit 132 may refer to step S103 in the embodiment corresponding to fig. 3, and will not be described herein.
In one or more embodiments, the sample tag information includes a service data tag corresponding to sample service data and a query result tag corresponding to a sample query result;
the model parameter adjustment module 14 includes: a first loss value determination unit 141, a second loss value determination unit 142, a model parameter training unit 143;
a first loss value determining unit 141, configured to determine a service loss value corresponding to the service estimating component according to the first index predicted value and the service data tag;
the second loss value determining unit 142 is configured to determine a result loss value corresponding to the result estimating component according to the second instruction predicted value and the query result label;
The model parameter training unit 143 is configured to perform weighted summation on the service loss value and the result loss value to obtain a model total loss corresponding to the initial pre-estimated model, perform parameter adjustment on the initial pre-estimated model according to the model total loss, and determine a service pre-estimated component in the initial pre-estimated model after parameter adjustment as a target pre-estimated model.
Alternatively, the second loss value determining unit 142 is specifically configured to:
carrying out logarithmic processing on the second index estimated value to obtain information quantity related to the second index estimated value;
and determining a result loss value corresponding to the result pre-estimation component according to the product between the query result label and the information quantity.
The functional implementation manners of the first loss value determining unit 141, the second loss value determining unit 142, and the model parameter training unit 143 may refer to step S104 in the embodiment corresponding to fig. 3, and will not be described herein.
In one or more embodiments, the service data processing apparatus 1 may include: the target joint vector acquisition module 15, the service input characteristic acquisition module 16, the feedback predicted value output module 17 and the target service data determination module 18;
the target joint vector acquisition module 15 is configured to acquire target query data and target object data corresponding to a target object, and acquire a target joint vector corresponding to the target object according to the target query number and the target object data;
The service input feature acquiring module 16 is configured to acquire candidate service vectors corresponding to each candidate service data in the candidate service data set, and pair and combine the target joint vector with the candidate service vectors corresponding to each candidate service data to obtain service input features corresponding to each candidate service data respectively;
the feedback predicted value output module 17 is configured to sequentially input service input features corresponding to each candidate service data to the target prediction model, and output feedback indicator predicted values corresponding to each candidate service data respectively through the target prediction model;
the target service data determining module 18 is configured to determine target service data for displaying to the target object in the candidate service data set according to the feedback indicator predicted value.
The functional implementation manners of the target joint vector obtaining module 15, the service input feature obtaining module 16, the feedback predicted value output module 17, and the target service data determining module 18 may refer to steps S201 to S204 in the embodiment corresponding to fig. 5, and will not be described herein.
In one or more embodiments, the feedback indicator predictive value includes a trigger indicator predictive value and a conversion indicator predictive value;
The target traffic data determination module 18 may include: a candidate service data acquisition unit 181, a recommendation evaluation value determination unit 182, and a service data selection unit 183;
a candidate service data obtaining unit 181, configured to obtain unit conversion values corresponding to each candidate service data in the candidate service data set, and determine service weight values corresponding to each candidate service data according to historical feedback information corresponding to each candidate service data;
a recommendation evaluation value determining unit 182, configured to determine recommendation evaluation values corresponding to the candidate service data respectively according to the unit conversion value, the service weight value, the trigger index predicted value and the conversion index predicted value;
the service data selecting unit 183 is configured to determine, from the candidate service data set, candidate service data corresponding to the maximum recommended evaluation value as target service data for displaying to the target object.
Alternatively, the recommendation evaluation value determining unit 182 is specifically configured to:
determining the comprehensive conversion value corresponding to each candidate service data respectively based on the product among the unit conversion value, the trigger index predicted value and the conversion index predicted value;
and determining recommendation evaluation values respectively corresponding to the candidate service data based on the sum of the comprehensive conversion value and the service weight value.
The functional implementation manner of the candidate service data obtaining unit 181, the recommendation evaluation value determining unit 182, and the service data selecting unit 183 may refer to step S204 in the embodiment corresponding to fig. 5, and will not be described herein.
In the embodiment of the application, sample object data, sample query results and sample service data associated with sample query data can be obtained, and the sample query data, the sample object data and the sample service data can be combined into first sample input data and input into a service prediction component in an initial prediction model to obtain a first index predicted value; the sample query data, the sample object data and the sample query result can be combined into a second sample input result, and the second sample input result is input to a result pre-estimation component in the initial pre-estimation model to obtain a second index pre-estimation value; and further, based on the first index predicted value, the second index predicted value and the sample label information, the network parameters of the initial pre-estimated model can be trained, and the trained service pre-estimated assembly can be used as a target pre-estimated model. In the training process of the initial pre-estimation model, the initial pre-estimation model can be jointly trained by utilizing the interaction between the sample object and the sample query result and the interaction between the sample object and the sample service data, so that the target pre-estimation model obtained by training can improve the prediction precision of the index pre-estimation value corresponding to the service data, and is beneficial to the accurate delivery of the service data, thereby improving the benefits of the service data.
Further, referring to fig. 9, fig. 9 is a schematic structural diagram of a computer device according to an embodiment of the present application. As shown in fig. 9, the computer device 1000 may be a terminal device, for example, the terminal device 10a in the embodiment corresponding to fig. 1, or a server, for example, the server 10d in the embodiment corresponding to fig. 1, which is not limited herein. For ease of understanding, the present application takes a computer device as an example of a terminal device, and the computer device 1000 may include: processor 1001, network interface 1004, and memory 1005, in addition, the computer device 1000 may further comprise: a user interface 1003, and at least one communication bus 1002. Wherein the communication bus 1002 is used to enable connected communication between these components. The user interface 1003 may also include a standard wired interface, a wireless interface, among others. The network interface 1004 may optionally include a standard wired interface, a wireless interface (e.g., WI-FI interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as at least one disk memory. The memory 1005 may also optionally be at least one storage device located remotely from the processor 1001. As shown in fig. 9, an operating system, a network communication module, a user interface module, and a device control application may be included in a memory 1005, which is one type of computer-readable storage medium.
The network interface 1004 in the computer device 1000 may also provide network communication functions, and the optional user interface 1003 may also include a Display screen (Display) and a Keyboard (Keyboard). In the computer device 1000 shown in fig. 9, the network interface 1004 may provide network communication functions; while user interface 1003 is primarily used as an interface for providing input to a user; and the processor 1001 may be used to invoke a device control application stored in the memory 1005 to implement:
acquiring sample object data and sample presentation data associated with sample query data; the sample display data comprises sample query results and sample service data, and carries sample tag information;
determining a first sample input feature according to the sample query data, the sample object data and the sample service data, and determining a second sample input feature according to the sample query data, the sample object data and the sample query result;
outputting a first index predicted value corresponding to the first sample input characteristic through a service predicting component in the initial predicting model, and outputting a second index predicted value corresponding to the second sample input characteristic through a result predicting component in the initial predicting model;
According to the first index estimated value, the second index estimated value and the sample label information, carrying out parameter adjustment on the initial estimated model, and determining a target estimated model based on the initial estimated model after parameter adjustment; the target pre-estimation model is used for predicting a feedback index pre-estimation value of the target object on the service data.
It should be understood that the computer device 1000 described in the embodiment of the present application may perform the description of the service data processing method in any of the embodiments of fig. 3 or fig. 5, and may also perform the description of the service data processing apparatus 1 in the embodiment corresponding to fig. 8, which is not repeated herein. In addition, the description of the beneficial effects of the same method is omitted.
Furthermore, it should be noted here that: the embodiment of the present application further provides a computer readable storage medium, in which a computer program executed by the service data processing apparatus 1 mentioned above is stored, and the computer program includes program instructions, when executed by a processor, can execute the description of the service data processing method in any of the embodiments of fig. 3 or fig. 5, and therefore, a detailed description will not be given here. In addition, the description of the beneficial effects of the same method is omitted. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), or the like. For technical details not disclosed in the embodiments of the computer-readable storage medium according to the present application, please refer to the description of the method embodiments of the present application. As an example, program instructions may be deployed to be executed on one computing device or on multiple computing devices at one site or, alternatively, across multiple computing devices distributed across multiple sites and interconnected by a communication network, where the multiple computing devices distributed across multiple sites and interconnected by the communication network may constitute a blockchain system.
In addition, it should be noted that: embodiments of the present application also provide a computer program product or computer program that may include computer instructions that may be stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor may execute the computer instructions, so that the computer device performs the description of the service data processing method in any one of the foregoing embodiments of fig. 3 or fig. 5, and thus, a detailed description will not be given here. In addition, the description of the beneficial effects of the same method is omitted. For technical details not disclosed in the computer program product or the computer program embodiments according to the present application, reference is made to the description of the method embodiments according to the present application.
The terms first, second and the like in the description and in the claims and drawings of embodiments of the application, are used for distinguishing between different media content and not for describing a particular sequential order. Furthermore, the term "include" and any variations thereof is intended to cover a non-exclusive inclusion. For example, a process, method, apparatus, article, or device that comprises a list of steps or elements is not limited to the list of steps or modules but may, in the alternative, include other steps or modules not listed or inherent to such process, method, apparatus, article, or device.
Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein may be embodied in electronic hardware, in computer software, or in a combination of the two, and that the elements and steps of the examples have been generally described in terms of function in the foregoing description to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
The method and related apparatus provided in the embodiments of the present application are described with reference to the flowchart and/or schematic structural diagrams of the method provided in the embodiments of the present application, and each flow and/or block of the flowchart and/or schematic structural diagrams of the method may be implemented by computer program instructions, and combinations of flows and/or blocks in the flowchart and/or block diagrams. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or structural diagram block or blocks. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or structures.
The foregoing disclosure is illustrative of the present application and is not to be construed as limiting the scope of the application, which is defined by the appended claims.