Model training method and device, data processing method and device and object recommendation method and device

Abstract

The embodiment of the invention provides a model training method and device, a data processing method and device and an object recommendation method and device, and the model training method comprises the steps: determining exposure conversion training data according to exposure click training data corresponding to an exposure click rate prediction model and click conversion training data corresponding to a click conversion rate prediction model; extracting exposure conversion characteristics of the exposure conversion training data, and determining a target prediction result by using a click prediction result output by the exposure click rate prediction model based on the exposure conversion characteristics and a conversion prediction result output by the click conversion rate prediction model based on the exposure conversion characteristics; training a click conversion rate estimation model according to the target prediction result and an exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model; by expanding the sample space, the target conversion rate prediction model can be better generalized to the representation space of more samples, the generalization performance of the model is improved, and the method can be applied to wider scenes.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and in particular to model training, data processing, and object recommendation methods and devices. The embodiments of this specification also relate to a computing device, a computer-readable storage medium, and a computer program product. Background Art

[0002] Currently, the CVR model (Click-to-View Rate Model l, which is an important module in the recommendation system and advertising system, mainly used to estimate the probability that a user will continue to convert after clicking) usually uses click data as the training set, where the samples in the training set that are clicked but not converted are negative samples, and the samples that are clicked and converted are positive samples.

[0003] When the CVR model uses positive samples of "click and conversion", the prediction target of the CVR model is "the probability of conversion among the samples that are exposed and clicked"; however, when the number of positive samples is small, the CVR model has the problem of sparse training sample space, and the CVR model can only predict the probability of conversion among the samples that are exposed and clicked, which greatly reduces the generalization ability of the CVR model. Summary of the invention

[0004] In view of this, an embodiment of the present specification provides a model training method. One or more embodiments of the present specification also relate to a model training device, a data processing method, a data processing device, an object recommendation method, an object recommendation device, a model training method applied to the cloud, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects of the CVR model in the prior art that the training sample space is sparse and the generalization ability is poor.

[0005] According to a first aspect of an embodiment of this specification, a model training method is provided, including:

[0006] Determine exposure conversion training data according to the exposure click rate prediction model's exposure click training data and the click conversion training data corresponding to the click conversion rate prediction model;

[0007] Extracting exposure conversion features of the exposure conversion training data, and using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features to determine a target prediction result;

[0008] The click-through conversion rate estimation model is trained according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

[0009] According to a second aspect of an embodiment of this specification, a model training device is provided, comprising:

[0010] A data determination module is configured to determine exposure-conversion training data according to the exposure-click training data corresponding to the exposure-click rate estimation model and the click-conversion training data corresponding to the click-conversion rate estimation model;

[0011] A result prediction module is configured to extract the exposure conversion features of the exposure conversion training data, and determine a target prediction result by using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features;

[0012] The model training module is configured to train the click-through conversion rate estimation model according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

[0013] According to a third aspect of an embodiment of this specification, a data processing method is provided, including:

[0014] Input the target data into the target conversion rate estimation model and the exposure click rate estimation model respectively;

[0015] Using the target conversion rate estimation model, obtaining a conversion prediction result corresponding to the target data;

[0016] Using the exposure click rate prediction model, obtaining a click prediction result corresponding to the target data;

[0017] A target conversion result of the target data is obtained according to the conversion prediction result and the click prediction result.

[0018] According to a fourth aspect of an embodiment of this specification, there is provided a data processing device, including:

[0019] A model input module is configured to input target data into a target conversion rate estimation model and an exposure click rate estimation model respectively;

[0020] A first prediction module is configured to obtain a conversion prediction result corresponding to the target data by using the target conversion rate estimation model;

[0021] A second prediction module is configured to obtain a click prediction result corresponding to the target data by using the exposure click rate prediction model;

[0022] The result obtaining module is configured to obtain the target conversion result of the target data according to the conversion prediction result and the click prediction result.

[0023] According to a fifth aspect of an embodiment of this specification, there is provided an object recommendation method, including:

[0024] In response to an object recommendation request sent by a client, determining user information corresponding to the client, and determining a plurality of candidate objects and context information according to the user information;

[0025] Determine target data according to the user information, the multiple candidate objects and the context information, and input the target data into a target conversion rate estimation model and an exposure click rate estimation model respectively;

[0026] Using the target conversion rate prediction model, obtaining a conversion prediction result corresponding to the target data, and using the exposure click rate prediction model, obtaining a click prediction result corresponding to the target data;

[0027] Obtaining a target conversion result of the target data according to the conversion prediction result and the click prediction result;

[0028] According to the target conversion result, the multiple candidate objects are sorted to obtain an object recommendation result, and the object recommendation result is returned to the client to display the object recommendation result on a user interaction interface of the client.

[0029] According to a sixth aspect of an embodiment of this specification, there is provided an object recommendation device, including:

[0030] A response module, configured to respond to an object recommendation request sent by a client, determine user information corresponding to the client, and determine a plurality of candidate objects and context information according to the user information;

[0031] An input module, configured to determine target data according to the user information, the multiple candidate objects and the context information, and input the target data into a target conversion rate estimation model and an exposure click rate estimation model respectively;

[0032] A prediction module is configured to obtain a conversion prediction result corresponding to the target data by using the target conversion rate prediction model, and to obtain a click prediction result corresponding to the target data by using the exposure click rate prediction model;

[0033] An acquisition module, configured to acquire a target conversion result of the target data according to the conversion prediction result and the click prediction result;

[0034] The return module is configured to sort the multiple candidate objects according to the target conversion result, obtain an object recommendation result, and return the object recommendation result to the client so as to display the object recommendation result on the user interaction interface of the client.

[0035] According to a seventh aspect of the embodiments of this specification, a model training method applied to the cloud is provided, including:

[0036] Receiving a model training request sent by the device end, and determining exposure-conversion training data in response to the model training request and according to the exposure-click training data corresponding to the exposure-click rate estimation model and the click-conversion training data corresponding to the click-conversion rate estimation model;

[0037] Extracting exposure conversion features of the exposure conversion training data, and using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features to determine a target prediction result;

[0038] According to the target prediction result and the exposure conversion label in the exposure conversion training data, the click-through conversion rate estimation model is trained to obtain a target conversion rate estimation model;

[0039] The target conversion rate estimation model is sent to the device end to deploy the target conversion rate estimation model on the device end.

[0040] According to an eighth aspect of the embodiments of this specification, a computing device is provided, including:

[0041] Memory and processor;

[0042] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, the steps of the above-mentioned model training method, data processing method and object recommendation method are implemented.

[0043] According to the ninth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores a computer program / instruction, and when the computer program / instruction is executed by a processor, the steps of the above-mentioned model training method, data processing method and object recommendation method are implemented.

[0044] According to the tenth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned model training method, data processing method and object recommendation method.

[0045] The model training method provided in the embodiments of the present specification determines exposure-conversion training data according to exposure-click training data corresponding to an exposure-click rate estimation network and click-conversion training data corresponding to a click-conversion rate estimation network. When the exposure-conversion training data is determined based on the exposure-click training data and the click-conversion training data, the exposure-conversion training data expands the sample space of the click-conversion rate estimation model compared to the click-conversion training data, thereby solving the problem of sparse training sample space of the click-conversion rate estimation model. The exposure-conversion features of the exposure-conversion training data are extracted, and a click prediction result output by the exposure-click rate estimation model based on the exposure-conversion features and a conversion prediction result output by the click-conversion rate estimation model based on the exposure-conversion features are used to determine a target prediction result. The click-conversion rate estimation model is trained according to the target prediction result and an exposure-conversion label in the exposure-conversion training data to obtain a target conversion rate estimation model. When the target prediction result is determined based on the click prediction result and the conversion prediction result, sample space modeling based on all samples is implemented based on joint probability, thereby improving the generalization performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a schematic diagram of a scenario of an object recommendation method provided by an embodiment of this specification;

[0047] Figure 2 is a flow chart of a model training method provided by an embodiment of this specification;

[0048] Figure 3 is a flow chart of a data processing method provided by an embodiment of this specification;

[0049] Figure 4 is a flow chart of an object recommendation method provided by an embodiment of this specification;

[0050] Figure 5 is a schematic diagram of sample space improvement provided by an embodiment of this specification;

[0051] Figure 6 It is a structural diagram of a multi-task learning framework provided by an embodiment of this specification;

[0052] Figure 7 It is a schematic diagram of evaluation results of a click-through conversion rate prediction model corresponding to different training steps provided by an embodiment of this specification;

[0053] Figure 8 It is a structural schematic diagram of a model training device provided by an embodiment of this specification;

[0054] Fig. 9is a structural schematic diagram of a data processing device provided by an embodiment of this specification;

[0055] Fig.10 is a structural diagram of an object recommendation device provided by an embodiment of this specification;

[0056] Fig.11 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION

[0057] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.

[0058] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0059] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0060] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0061] First, the terms involved in one or more embodiments of this specification are explained.

[0062] CVR model: Click-to-View Rate Model l, click-to-view rate model, is an important module in recommendation systems and advertising systems. It is mainly used to estimate the probability that a user will continue to convert after clicking.

[0063] CTR model: Click-Through Rate model, used to estimate the probability of a user clicking on an ad.

[0064] Conversion: The act of completing the intended action (such as purchasing a product, registering an account, etc.) after clicking on an ad.

[0065] In this specification, a model training method is provided. This specification also involves a model training device, a data processing method, a data processing device, an object recommendation method, an object recommendation device, a model training method applied to the cloud, a computing device, a computer-readable storage medium and a computer program product, which are described in detail one by one in the following embodiments.

[0066] See also Figure 1 , Figure 1 A schematic diagram of a scenario of an object recommendation method provided according to an embodiment of the present specification is shown.

[0067] Specifically, the object recommendation method is applied to the server 104, and the terminal device 102 is used to send an object recommendation request to the server 104; the target conversion rate estimation model and the exposure click rate estimation model are trained in the server 104, and when the server 104 receives the object recommendation request sent by the terminal device 102, the user information corresponding to the terminal device 102 is determined according to the object recommendation request, and multiple candidate objects and context information of each candidate object are determined according to the user information; target data is determined according to the user information, the multiple candidate objects and the context information of each candidate object, and the target data is respectively input into the target conversion rate estimation model and the exposure click rate estimation model; using the target The conversion rate prediction model is used to obtain the conversion prediction result corresponding to the target data, and the click prediction result corresponding to the target data is obtained by using the exposure click rate prediction model; the target conversion result of the target data is obtained according to the conversion prediction result and the click prediction result; the multiple candidate objects are sorted according to the target conversion result to obtain the object recommendation result (for example, the object recommendation result is object 1, object 2, object 3, object 4, object 5, object 6...), and the object recommendation result is returned to the terminal device 102 to display the object recommendation result on the user interaction interface of the terminal device 102. Specifically, when displaying the object recommendation result, it can be displayed according to a preset arrangement method, for example Figure 1As shown, they are displayed in two columns from left to right and from top to bottom.

[0068] Among them, the target conversion rate estimation model is trained by the following method:

[0069] Determine exposure conversion training data according to the exposure click rate prediction model's exposure click training data and the click conversion training data corresponding to the click conversion rate prediction model;

[0070] Extracting exposure conversion features of the exposure conversion training data, and using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features to determine a target prediction result;

[0071] The click-through conversion rate estimation model is trained according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

[0072] The end-side device 102 may include a browser, an APP (Application), or a web application such as an H5 (Hyper Text Markup Language 5, version 5 of Hypertext Markup Language) application, or a light application (also known as a mini-program, a lightweight application) or a cloud application, etc. The end-side device may be based on the software development kit (SDK, Software Development Kit) of the corresponding service provided by the server, such as based on the real-time communication (RTC, Real Time Communication) SDK development and acquisition, etc. The end-side device may be deployed in an electronic device and needs to rely on the device to run or some APPs in the device to run, etc. The electronic device may have a display screen and support information browsing, etc., such as a personal mobile terminal such as a mobile phone, a tablet computer, a personal computer, etc. Various other types of applications may also be configured in the electronic device, such as human-computer dialogue applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0073] Server 104 can be understood as a server that provides various services, including physical servers and cloud servers, such as a server that provides communication services to multiple clients, a server for background training that provides support for models used on clients, and a server that processes data sent by clients. It should be noted that server 104 can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. Server 104 can also be a server for a distributed system, or a server combined with a blockchain. Server 104 can also be a cloud server for basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (CDN, Content De l ivery Network), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0074] It is worth noting that the object recommendation method provided in the embodiments of this specification can be executed by the server 104. In other embodiments of this specification, the target conversion rate prediction model and the exposure click rate prediction model can be deployed in the terminal device 102, so that the terminal device 102 can also have similar functions to the server 104, thereby executing the object recommendation method provided in the embodiments of this specification; in other embodiments, the object recommendation method provided in the embodiments of this specification can also be executed jointly by the terminal device 102 and the server 104.

[0075] The object recommendation method provided in the embodiments of this specification improves the accuracy of conversion prediction results on the basis of improving the generalization performance of the target conversion rate estimation model through an expanded sample space and sample space modeling based on the full amount of samples. Therefore, by comprehensively considering the conversion prediction results and the click prediction results, it is possible to more accurately understand the user's preference for candidate objects, which helps to optimize the personalized recommendation algorithm and provide users with recommended content that is more in line with their interests and needs.

[0076] See also Figure 2 , Figure 2 A flow chart of a model training method provided by an embodiment of the present specification is shown, which specifically includes the following steps.

[0077] Step 202: Determine exposure-conversion training data according to the exposure-click rate prediction model's exposure-click rate training data and the click-conversion training data corresponding to the click-conversion rate prediction model.

[0078] Among them, the exposure-click rate prediction model can be understood as a model used to estimate the probability of a user clicking on an exposed object (advertisement, product, etc.), such as the exposure-click rate prediction model is a CTR model; the positive samples in the exposure-click training data corresponding to the exposure-click rate prediction model are the samples that are clicked in the exposure samples, and the negative samples in the exposure-click training data are the samples that are not clicked in the exposure samples.

[0079] The click-through conversion rate prediction model can be understood as a model used to predict the probability that a user will continue to convert after clicking on an object. For example, the click-through conversion rate prediction model is a CVR model. The positive samples in the click-through conversion training data corresponding to the click-through conversion rate prediction model are the converted samples in the click samples, and the negative samples in the click-through conversion training data are the unconverted samples in the click samples.

[0080] Specifically, according to different application scenarios, different sample objects in the training data, and different conversion behaviors; for example, in the product recommendation scenario, the exposure sample is the product pushed to the user interaction interface so that the user can see it, and the click sample is the product clicked by the user. The user's conversion behavior can be understood as purchasing the product; in the live broadcast scenario, the exposure sample is the live broadcast pushed to the user interaction interface so that the user can see it, and the click sample is the live broadcast clicked and entered by the user. The user's conversion behavior can be understood as the user's behavior of changing from a certain initial state (such as browsing, watching) to a more valuable state (such as purchasing, paying, etc.). For example, conversion behaviors include but are not limited to sending barrage (interactive behavior), sending live broadcast gifts (payment behavior), and purchasing items from the shopping link in the live broadcast room (purchase behavior).

[0081] In one or more embodiments of the present specification, before determining the exposure-conversion training data, first determine the exposure-click training data corresponding to the exposure-click rate estimation model and the click-conversion training data corresponding to the click-conversion rate estimation model. Specifically, determine the positive and negative samples corresponding to the exposure-click rate estimation model and the positive and negative samples corresponding to the click-conversion rate estimation model from the initial data set; construct the exposure-click training data based on the positive and negative samples corresponding to the exposure-click rate estimation model, and construct the click-conversion training data based on the positive and negative samples corresponding to the click-conversion rate estimation model. The specific implementation method is as follows:

[0082] Before determining the exposure-conversion training data according to the exposure-click rate prediction model's exposure-click rate training data and the click-conversion training data corresponding to the click-conversion rate prediction model, the method further includes:

[0083] Determine the initial data set;

[0084] Determine the samples that are clicked and converted in the initial data set as positive samples of the exposure-to-click rate estimation model and the click-to-conversion rate estimation model, determine the samples that are clicked but not converted in the initial data set as positive samples of the exposure-to-click rate estimation model and negative samples of the click-to-conversion rate estimation model, and determine the samples that are exposed but not clicked in the initial data set as negative samples of the exposure-to-click rate estimation model;

[0085] Determining exposure-click training data corresponding to the exposure-click rate prediction model according to the positive samples and negative samples of the exposure-click rate prediction model;

[0086] According to the positive samples and negative samples of the click-through conversion rate prediction model, click-through conversion training data corresponding to the click-through conversion rate prediction model is determined.

[0087] Specifically, the initial data set includes exposure samples, click samples and conversion samples, among which click samples are a subset of exposure samples, that is, samples can only be clicked when exposed, and conversion samples are a subset of conversion samples, that is, samples can only be converted when clicked.

[0088] Since the exposure-to-click rate prediction model is used to predict the click rate, the positive samples of the exposure-to-click rate prediction model can be determined based on the click samples in the initial data set (the click samples include samples that were clicked and converted, and samples that were clicked but not converted), and the negative samples of the exposure-to-click rate prediction model can be determined based on the samples that were exposed but not clicked in the initial data set. Based on the positive and negative samples corresponding to the exposure-to-click rate prediction model, the exposure-to-click training data of the exposure-to-click rate prediction model is determined.

[0089] Since the click-through conversion rate prediction model is used to predict the conversion rate, the positive samples of the click-through conversion rate prediction model can be determined based on the conversion samples (i.e., samples that are clicked and converted) in the initial data set, and the negative samples of the click-through conversion rate prediction model can be determined based on the unconverted samples in the initial data set (unconverted samples include samples that are clicked but not converted, and samples that are exposed but not clicked). Based on the positive and negative samples of the click-through conversion rate prediction model, the click-through conversion training data of the click-through conversion rate prediction model is determined.

[0090] In practical applications, negative samples of exposure-conversion training data are determined based on samples that are clicked among exposure samples, samples that are not clicked among exposure samples, and samples that are not converted among click samples (these three types of samples are exposed but not converted samples); positive samples of exposure-conversion training data are determined based on samples that are converted among click samples (i.e., samples that are exposed and converted).

[0091] Subsequently, when using the exposure-conversion training data to train the click-through rate prediction model, the sample space was expanded. That is, on the basis of originally using the click-through training data to train the CVR model, the use of exposure-click training data was added. The click-through training data used by the CVR model is several million, of which the positive samples include hundreds of thousands. However, the CTR model training uses exposure-click training data of up to 70 million. By determining the exposure-conversion training data based on the exposure-click training data and the click-conversion training data, the sample space deviation is effectively repaired, solving the problem of sparse sample space for CVR model training.

[0092] Step 204: extract the exposure conversion features of the exposure conversion training data, and use the click prediction results output by the exposure click-through rate prediction model based on the exposure conversion features, and the conversion prediction results output by the click-through rate prediction model based on the exposure conversion features to determine the target prediction results.

[0093] Among them, a multi-task learning framework can be constructed based on the exposure-click rate prediction model and the click-through rate prediction model. In the multi-task learning framework, the exposure-click rate prediction model is used to implement the click prediction task, and the click-through rate prediction model is used to implement the conversion prediction task.

[0094] The click prediction result can be understood as the probability value of a certain exposure sample (such as an advertisement, product recommendation page, etc.) being clicked by the user based on the user's historical behavior, current context information (such as time, location, device type, etc.) and the characteristics of the advertisement or product. The click prediction result can be expressed by the following formula: (商卡被点击|曝光的商卡) .

[0095] The conversion prediction result can be understood as the probability value of the click-through conversion rate estimation model further predicting that the user will complete a certain conversion behavior (such as purchasing a product, registering an account, filling out a form, etc.) after determining that the user has clicked on a certain exposure sample. The conversion prediction result can be expressed using the following formula: (商卡转化|商卡被点击，曝光的商卡) .

[0096] Specifically, the target prediction result is determined by combining the click prediction result and the conversion prediction result. The target prediction result can be understood as the joint probability value of the two events of click and conversion. That is, the target prediction result is obtained by the following formula: P (商卡转化，商卡被点击|曝光的商卡) =P (商卡被点击|曝光的商卡) ×P (商卡转化|商卡被点击，曝光的商卡) .

[0097] In one or more embodiments of the present specification, the exposure conversion features of the exposure conversion training data are extracted through a shared feature extraction network, and the exposure conversion features are respectively input into the exposure click rate estimation model and the click conversion rate estimation model of different learning tasks. The specific implementation is as follows:

[0098] The step of extracting the exposure conversion features of the exposure conversion training data includes:

[0099] Inputting the exposure conversion training data into a shared feature extraction network, and using the shared feature extraction network to perform feature extraction on the exposure conversion training data to obtain exposure conversion features of the exposure conversion training data;

[0100] The step of determining a target prediction result by using the click prediction result output by the exposure-to-click rate prediction model based on the exposure-to-conversion feature and the conversion prediction result output by the click-to-conversion rate prediction model based on the exposure-to-conversion feature includes:

[0101] Inputting the exposure conversion feature into the exposure click rate prediction model to obtain the click prediction result output by the exposure click rate prediction network;

[0102] Inputting the exposure conversion feature into the click-through conversion rate estimation model to obtain the conversion prediction result output by the click-through conversion rate estimation network;

[0103] The target prediction result is obtained according to the click prediction result and the conversion prediction result.

[0104] Among them, the shared feature extraction network is used to extract useful features from the input data, and these features will be used for subsequent click-through rate estimation and conversion rate estimation; it should be noted that the shared feature extraction network includes an embedding layer and a network layer.

[0105] Specifically, the exposure conversion training data is input into a shared feature extraction network, and useful features, namely, exposure conversion features, are extracted from the exposure conversion training data using the shared feature extraction network. The exposure conversion features are respectively input into an exposure click rate prediction model and a click conversion rate prediction model to obtain click prediction results and conversion prediction results. Based on these two prediction results, a target prediction result is calculated.

[0106] In actual applications, taking the product recommendation scenario of an e-commerce platform as an example, a user browses a product page (exposure), then clicks on the product (click), and finally buys the product (conversion). This is a complete exposure-conversion chain. By training a model to predict the overall probability of a user seeing the product page and purchasing the product, we can better recommend products that users are more interested in and have a greater probability of purchasing on the product page.

[0107] The exposure conversion training data is a data set that records users' browsing, clicking, and purchasing behaviors. The exposure conversion training data is input into a shared feature extraction network to extract product features (such as price, category, brand, etc.) in the product page and user features (such as historical purchasing behavior, browsing preferences, etc.). Of course, it can also include some contextual information about the products in the product page.

[0108] The features extracted by the shared feature extraction network are input into the exposure click rate prediction model to obtain the probability of a user clicking on a product (click prediction result), and the features extracted by the shared feature extraction network are input into the click conversion rate prediction model to obtain the probability of a user purchasing a product (conversion prediction result). By multiplying these two probabilities, the overall probability of a user seeing the product page and purchasing the product is obtained, that is, the probability of exposure conversion (target prediction result).

[0109] The model training method provided in the embodiments of this specification can extract key features that affect both click and conversion behaviors from exposure-conversion training data through a shared feature extraction network. These features are respectively input into the exposure-click rate prediction model and the click-conversion rate prediction model, so that both models can make full use of the information in the data and improve the accuracy of the prediction. The target prediction results also provide data support for subsequent model adjustments.

[0110] Step 206: training the click-through conversion rate estimation model according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

[0111] Among them, the exposure conversion label can be understood as the actual result of the sample in the exposure conversion training data, that is, whether the user actually clicked or converted. The exposure conversion label is used to compare with the target prediction result during the training process to train the model and adjust the model parameters.

[0112] The target conversion rate prediction model can be understood as a model that, after training and optimization, can more accurately predict the user's behavior from exposure to conversion.

[0113] Specifically, the exposure conversion training data is used to train the click-through conversion rate estimation model. When the target prediction result is obtained by the prediction results output by the exposure-to-click rate estimation model and the click-to-conversion rate estimation model, the target prediction result is compared with the actual exposure-to-conversion label in the exposure conversion training data. Based on the comparison result, the model parameters of the click-through conversion rate estimation model are continuously adjusted and optimized, and the target conversion rate estimation model is obtained based on the click-through conversion rate estimation model after the model parameters are adjusted.

[0114] Every day, e-commerce websites, social media platforms, etc. introduce a large number of new products and users. These new elements bring brand new information, such as product descriptions, prices, categories, and user behavior history, preferences, etc. In machine learning models, especially deep learning models, this information will be converted into numerical form, namely features. For products and users, embedding technology (embedding technology, a technology that maps high-dimensional data to low-dimensional vector space) is usually used to convert various attribute information such as product descriptions and user historical behaviors into vectors in high-dimensional space. These vectors are embedding features.

[0115] Since new products and new users are added every day, their corresponding embedding features will change the distribution of the feature space as a whole; in the daily training process of the model, as the data of a day is gradually trained, the model will try to adapt to the data distribution of that day, that is, the model will converge to the embedding distribution change of that day. However, due to the addition of new elements, the data distribution of this day is different from that of the previous day. If the model relies too much on the knowledge learned the day before (that is, the embedding distribution of the previous day), then it cannot adapt well to the new elements, resulting in a decrease in the generalization ability of the model. Generalization ability refers to the ability of a model to make accurate predictions on data it has not seen. In the product recommendation scenario, if the model cannot handle new products and new users well, then its product recommendation performance will deteriorate when facing new users or recommending new products.

[0116] In one or more embodiments of the present specification, in order to solve the above problem, the target conversion rate estimation model is obtained by using the first target model parameters determined in the middle training step and the second target model parameters determined in the last training step. The specific implementation is as follows:

[0117] Specifically, first determine the total number of training steps for model training, and determine multiple initial training steps for deriving model parameters from the total training steps. The specific implementation is as follows:

[0118] Determine multiple initial training steps, including:

[0119] Determining the total training steps for the click-through conversion rate estimation model according to the data volume of the exposure-to-conversion training data and the preset data volume for each round of training;

[0120] According to a preset step number interval, a plurality of initial training steps are determined from the total training steps.

[0121] The preset amount of training data per round can be understood as the amount of data used in each round of iteration during model training. This value is usually preset to ensure the stability of the model training process and prevent instability in training due to excessive or insufficient data. The total training step can be understood as the number of iterations (or training steps) that the model needs to perform during the entire model training process. This value is calculated based on the total amount of exposure conversion training data and the preset amount of training data per round. The model parameters will be adjusted once each iteration of the model (or each training step of the model).

[0122] The preset step interval can be understood as a pre-set step interval. According to the preset step interval, during the model training process, it is determined how many steps are used to derive the model parameters, thereby determining multiple initial training steps for deriving the model parameters, that is, through the preset step interval, multiple initial training steps can be determined from the total training steps; the initial training step can be understood as a training step for deriving model parameters. For example, the total training steps are 100 steps, the preset step interval is 10 steps, and the initial training steps are the 10th step, the 20th step, the 30th step... the 90th step, the 100th step. At these initial training steps, the model parameters corresponding to the training steps are derived, wherein the initial training steps include the last training step, that is, the model parameters of the last training step need to be saved.

[0123] For example, there is a dataset containing 1 million exposure conversion training data, and the preset amount of training data per round is 100,000. Then, during the entire training process, the click-through conversion rate estimation model needs to be iterated 10 times (that is, the total training steps are 10). If the preset step interval is set to 2, then 5 initial training steps can be determined from the total training steps, namely step 2, step 4, step 6, step 8, and step 10.

[0124] Specifically, the model parameters of the click-through conversion rate estimation model corresponding to the multiple initial training steps are determined, and then the first target model parameters and the second target model parameters are determined from these model parameters to obtain the final target conversion rate estimation model. The specific implementation method is as follows:

[0125] The step of training the click-through conversion rate estimation model according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain the target conversion rate estimation model includes:

[0126] Determining a plurality of initial training steps, exposure conversion training data corresponding to each of the plurality of initial training steps, and a target prediction result of the exposure conversion training data corresponding to each of the initial training steps;

[0127] Obtaining model parameters of the click-through conversion rate estimation model corresponding to each initial training step according to the target prediction results corresponding to each initial training step and the exposure conversion labels in the exposure conversion training data;

[0128] A first target model parameter and a second target model parameter are determined from the model parameters corresponding to each initial training step, and the target conversion rate estimation model is obtained according to the first target model parameter and the second target model parameter.

[0129] Among them, the first target model parameter can be understood as a model parameter whose model performance is better than the model performance corresponding to other model parameters; the second target model parameter can be understood as a model parameter obtained by adjusting the click-through conversion rate prediction model according to more learned training data.

[0130] In practical applications, in each training step, the model is trained using the exposure-conversion training data corresponding to the training step, and the model parameters are adjusted once. When multiple initial training steps are determined, the target prediction result of the exposure-conversion training data corresponding to each initial training step can be determined by determining the exposure-conversion training data corresponding to each initial training step. By comparing the target prediction result corresponding to each initial training step with the exposure-conversion label in the exposure-conversion training data, the model parameters of the initial conversion rate estimation model corresponding to each initial training step can be determined.

[0131] From multiple model parameters, determine a first target model parameter whose model performance is better than other model parameters, and a second target model parameter obtained by training with more training data; and determine a final target conversion rate estimation model based on the first target model parameter and the second target model parameter.

[0132] Specifically, according to the model parameters corresponding to different initial training steps, the click-through conversion rate estimation models of different initial training steps can be determined, and the click-through conversion rate estimation models of different initial training steps are evaluated, and the click-through conversion rate estimation model with better performance can be determined from the click-through conversion rate estimation models of different initial training steps according to the evaluation results, and the initial training step corresponding to the click-through conversion rate estimation model with better performance is determined as the first target training step, and its corresponding model parameters are the first target model parameters; in the training step at the end of the model training iteration, the click-through conversion rate estimation model has learned all the data in the training data, and the training step at the end of the model training iteration is determined as the second target training step, and the model parameters corresponding to the training step at the end of the model training iteration are the second target model parameters. The specific implementation method is described as follows:

[0133] The step of determining a first target model parameter and a second target model parameter from the model parameters corresponding to each of the initial training steps, and obtaining the target conversion rate estimation model according to the first target model parameter and the second target model parameter, comprises:

[0134] Determine, according to the model parameters corresponding to each initial training step, a click-through conversion rate estimation model corresponding to each initial training step, wherein the click-through conversion rate estimation model corresponding to each initial training step is obtained by training according to the target prediction results corresponding to each initial training step and the exposure conversion labels in the exposure conversion training data;

[0135] Evaluate the click-through conversion rate estimation models corresponding to the initial training steps to obtain evaluation results, and determine the click-through conversion rate estimation model corresponding to the first target training step from the click-through conversion rate estimation models corresponding to the multiple initial training steps according to the evaluation results, wherein the evaluation result of the click-through conversion rate estimation model corresponding to the first target training step is greater than the evaluation results of the click-through conversion rate estimation models corresponding to other training steps;

[0136] Determine the first target model parameters according to the click-through conversion rate estimation model corresponding to the first target training step, and determine the second target model parameters according to the click-through conversion rate estimation model corresponding to the second target training step, wherein the second target training step is the training step at which the model training iteration ends;

[0137] The target conversion rate estimation model is obtained according to the first target model parameters and the second target model parameters.

[0138] Among them, the evaluation result can be understood as the result of quantitative evaluation of the model performance through some evaluation indicators (such as accuracy, recall rate, AUC, etc.); the first target training step can be understood as the training step in which the evaluation result of the click-through conversion rate estimation model is better than the evaluation results of other initial training steps among multiple initial training steps; the second target training step can be understood as the training step at the end of the model training iteration, which usually corresponds to the last step of the entire training process.

[0139] In practical applications, the click-through conversion rate estimation model corresponding to each initial training step can be derived using the model parameters corresponding to each initial training step. These derived click-through conversion rate estimation models correspond to different training stages. When evaluating each derived click-through conversion rate estimation model, the click-through conversion rate estimation model can be evaluated using indicators such as AUC (Area Under the Curve, which is a commonly used indicator for measuring the performance of a binary classification model, especially when evaluating the ability of a model to distinguish between positive and negative samples. AUC refers to the area under the ROC curve (Receiver Operating Characteristic Curve), and its value range is between 0 and 1. The larger the AUC value, the better the performance of the model, and the better it can distinguish between positive and negative samples), accuracy, recall rate, etc., to obtain the evaluation results of the click-through conversion rate estimation model corresponding to each initial training step, thereby determining the click-through conversion rate estimation model with better performance corresponding to the first target training step from multiple click-through conversion rate estimation models based on the evaluation results.

[0140] The model parameters corresponding to the last training step are determined as the second target model parameters, so as to obtain the target conversion rate estimation model according to the first target model parameters and the second target model parameters.

[0141] Specifically, when the click-through conversion rate prediction model includes a network layer and an embedding layer, the network parameters corresponding to the network layer and the embedding layer can be determined respectively, for example, the network parameters corresponding to the network layer are the first target model parameters, and the network parameters corresponding to the embedding layer are the second target model parameters. According to the network layer with the first target model parameters and the embedding layer with the second target model parameters, the target conversion rate prediction model is obtained to ensure that the target conversion rate prediction model will not have the above-mentioned problem of decreased model generalization ability due to new products and new users. The specific implementation method is as follows:

[0142] The step of determining the first target model parameters according to the click-through conversion rate estimation model corresponding to the first target training step, and determining the second target model parameters according to the click-through conversion rate estimation model corresponding to the second target training step, includes:

[0143] Determine the network parameters of the network layer in the click-through conversion rate estimation model corresponding to the first target training step as the first target model parameters;

[0144] Determine the network parameters of the embedding layer in the click-through conversion rate estimation model corresponding to the second target training step as the second target model parameters;

[0145] The step of obtaining the target conversion rate estimation model according to the first target model parameter and the second target model parameter includes:

[0146] The target conversion rate estimation model is obtained according to the network layer having the first target model parameters and the embedding layer having the second target model parameters.

[0147] The network layer can be understood as a Dense DNN (Dense Deep Neura l Network) layer, which refers to a deep neural network composed of multiple fully connected layers. It can process continuous or embedded dense features. This network can perform complex nonlinear transformations to extract higher-level abstract features from the original features.

[0148] The embedding layer can be understood as a Spares Embedding layer, which is usually used to process high-dimensional and sparse category features (such as user ID, product ID, etc.). It reduces the feature dimension by mapping the ID of each category to a low-dimensional vector space and captures the potential relationship between categories.

[0149] In practical applications, the network layer derived from the intermediate step (the first target training step) is used together with the embedding layer that is finally trained to form an online inference model (i.e., the target conversion rate estimation model).

[0150] Specifically, the features extracted by the network layer are Dense features, which are usually continuous values ​​or discrete features with a small range, such as the user's age, rating, etc. These features are relatively stable in the data set and are not easily affected by new elements (new products, new users). Therefore, using the network layer in the click-through conversion rate prediction model with better performance can provide a certain degree of stability, so that the target conversion rate prediction model will not fluctuate too much when facing new data; and Dense features are easier to learn and generalize by the model because of their limited and continuous range, which means that the model can make reasonable predictions on unseen data based on these features, thereby improving the generalization ability of the model.

[0151] The features proposed by the embedding layer are Sparse features. Sparse features are used to capture personalization and diversity. That is, Sparse features are usually features with a large number of possible values ​​but few actual usage values, such as product identification and user historical behavior. These features can capture personalized information about users and products. Through Sparse features, the target conversion rate prediction model can accurately recommend products that meet user interests. With the addition of new products and new users, Sparse features will continue to expand their value range, providing more diversity for the target conversion rate prediction model, which helps the model better adapt to changes in data distribution and reduce overfitting of new elements.

[0152] Dense features and Sparse features each have their own advantages in the recommendation system and complement each other. Dense features provide stability and generalization capabilities, while Sparse features capture personalization and diversity. Combining the two can give full play to their respective advantages and improve the performance of the target conversion rate prediction model.

[0153] The Dense features derived from the intermediate step can capture the basic attributes and behavioral trends of users, while the Sparse features finally trained can reflect the latest user interests and product information. This combination enables the model to adapt to changes in new data while maintaining a certain degree of stability and accuracy.

[0154] By introducing the Dense features derived from the intermediate step, the target conversion rate estimation model can achieve a smooth transition when new elements appear. The stability provided by the Dense features helps the model establish connections between new and old data, reducing performance fluctuations caused by the addition of new elements; the Sparse features that are finally trained can capture the latest changes in data distribution, allowing the model to quickly adapt to the appearance of new elements, thereby making the target conversion rate estimation model more adaptable and robust when facing new data.

[0155] The model training method provided in the embodiments of this specification uses the network layer derived from the intermediate step (first target training step) and the embedding layer that is finally trained (second target training step) to form an online inference model. It can give full play to the stability and generalization ability of the network layer corresponding to the Dense feature and the personalization and diversity advantages of the embedding layer corresponding to the Sparse feature, thereby reducing the impact of new elements on the model and improving the performance and generalization of the model.

[0156] In one or more embodiments of the present specification, the target loss function is calculated based on the target prediction result and the exposure conversion label, so as to adjust the model parameters of the exposure click rate estimation model and the click conversion rate estimation model according to the target loss function, and determine the target conversion rate estimation model according to the adjusted click conversion rate estimation model. The specific implementation is as follows:

[0157] The step of training the click-through conversion rate estimation model according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain the target conversion rate estimation model includes:

[0158] Determining a target loss function according to the target prediction result and the exposure conversion label in the exposure conversion training data;

[0159] According to the target loss function, adjusting the model parameters of the click-through conversion rate estimation model;

[0160] The target conversion rate estimation model is determined according to the click-through conversion rate estimation model after the model parameters are adjusted.

[0161] Among them, the target loss function can be understood as an indicator that measures the difference between the model's predicted value and the actual value. During the training process, the goal of the model is to minimize this loss function to improve the accuracy of the prediction.

[0162] Specifically, based on the click-through rate prediction model of the multi-task learning framework, the exposure-to-click rate prediction model is used as an auxiliary training task. In practical applications, the existing CVR model parameters are used to initialize the shared feature extraction network and the CVR prediction task network. The exposure-to-click rate prediction model is determined as the CTR prediction task network in the multi-task learning framework. The shared feature extraction network includes an embedding layer and a network layer. The exposure conversion training data is input into the shared feature extraction network to obtain exposure conversion features. The exposure conversion features are respectively input into the CVR prediction task network and the CTR prediction task network to obtain click prediction results and conversion prediction results, and the target prediction results are determined.

[0163] According to the target prediction result and the actual exposure conversion label, a target loss function is calculated, wherein the target loss function can be a total loss function determined by weighted summation of the loss functions of the two tasks. By minimizing the target loss function, the network parameters of each network in the multi-task learning framework are continuously adjusted to improve the prediction accuracy. Based on the shared feature extraction network and the CVR estimation task network after the network parameters are adjusted, an optimized click-through conversion rate estimation model is obtained, and the target conversion rate estimation model is determined according to the optimized click-through conversion rate estimation model; it should be noted that the shared feature extraction network includes an embedding layer and a network layer. When the parameters of the shared feature extraction network are also adjusted based on the target prediction result and the exposure conversion label, based on the method described in the above embodiment, according to the network layer with the first target model parameters and the embedding layer with the second target model parameters, the shared feature extraction network after the network parameters are adjusted is determined, thereby realizing the shared feature extraction network and the CVR estimation task network after the network parameters are adjusted, obtaining the click-through conversion rate estimation model after the model parameters are adjusted, and obtaining the target conversion rate estimation model.

[0164] The model training method provided in the embodiments of this specification can reuse the "exposed but not clicked" samples based on the exposure conversion training data to effectively repair the sample space deviation, and use the click prediction results output by the exposure click rate prediction model and the conversion prediction results output by the click conversion rate prediction model to determine the target prediction results, and regard the conversion rate as the joint probability of the two events of click and conversion, thereby expanding the sample space and solving the problem of sparse training sample space in a simple and efficient way; construct a CTR model as an auxiliary training task, and use a multi-task learning framework and joint probability to achieve sample space modeling based on the full amount of exposure samples, thereby improving the generalization performance of the model.

[0165] The embodiments of the present specification also provide a model training method applied to the cloud, which can receive a model training request sent by a device end, and determine exposure-conversion training data in response to the model training request, based on the exposure-click training data corresponding to the exposure-click rate prediction model, and the click-conversion training data corresponding to the click-conversion rate prediction model; extract the exposure-conversion features of the exposure-conversion training data, and use the click prediction results output by the exposure-click rate prediction model based on the exposure-conversion features, and the conversion prediction results output by the click-conversion rate prediction model based on the exposure-conversion features, to determine the target prediction results; train the click-conversion rate prediction model based on the target prediction results and the exposure-conversion labels in the exposure-conversion training data to obtain a target conversion rate prediction model; and send the target conversion rate prediction model to the device end to deploy the target conversion rate prediction model on the device end.

[0166] Specifically, through the model training method applied to the cloud, the target conversion rate estimation model can be deployed on the device side, so as to achieve the prediction of the target conversion result by using the target conversion rate estimation model on the device side.

[0167] The specific implementation method can be found in the above embodiments, which will not be described in detail here.

[0168] See also Figure 3 , Figure 3 A flow chart of a data processing method provided by an embodiment of the present specification is shown, which specifically includes the following steps.

[0169] Step 302: inputting the target data into the target conversion rate estimation model and the exposure click rate estimation model respectively;

[0170] Step 304: using the target conversion rate estimation model to obtain a conversion prediction result corresponding to the target data;

[0171] Step 306: using the exposure click rate prediction model to obtain click prediction results corresponding to the target data;

[0172] Step 308: Obtain a target conversion result of the target data according to the conversion prediction result and the click prediction result.

[0173] In one or more embodiments of the present specification, the user information corresponding to the client can be first determined according to the object recommendation request, and multiple candidate objects can be screened to obtain, thereby reducing the amount of data input to the model when calculating the click probability and conversion probability for the multiple candidate objects, thereby improving the efficiency of data processing. The specific implementation is as follows:

[0174] Before inputting the target data into the target conversion rate estimation model and the exposure click rate estimation model, the method further includes:

[0175] In response to an object recommendation request sent by a client, determining user information corresponding to the client, and determining a plurality of candidate objects and context information of each candidate object according to the user information;

[0176] The target data is determined according to the user information, the multiple candidate objects, and context information of each candidate object.

[0177] Among them, the object recommendation request can be understood as a request sent by the client (such as an application on the user's device) to request the system to recommend a series of objects (such as goods, content, advertisements, etc.) to the user. User information can be understood as information related to the user corresponding to the client that sends the object recommendation request, which may include the user's profile, historical behavior, preference settings, etc. Candidate objects can be understood as a series of objects that may be recommended to the user based on the user information. The candidate objects have a certain relevance or potential interest to the user.

[0178] Contextual information can be understood as additional information related to the candidate object, such as the attributes of the object itself (such as price, category, brand, etc.), the historical interaction records between the user and the object, the current time and place, etc. Contextual information helps to more accurately assess the user's interest in the candidate object.

[0179] The target data can be understood as data used to be input into the target conversion rate estimation model and the exposure click rate estimation model, which usually includes a combination or representation of user information, candidate objects and their context information.

[0180] In actual applications, when a client sends an object recommendation request, it first determines multiple candidate objects and the context information of these candidate objects based on the user information carried in the request, and constructs target data based on this information. These target data will be input into the target conversion rate prediction model and the exposure click rate prediction model to predict the user's click and conversion probability for the candidate objects.

[0181] Taking the product recommendation scenario of an e-commerce platform as an example, the application on the user's device sends a product recommendation request (object recommendation request) to the server. After receiving the request, the server first parses the user information carried in the request, such as the user's ID (identification), historical purchase records, browsing records, etc. Based on the user information, the server selects multiple candidate products related to the user's interests from the product library. For each candidate product, the server can also obtain relevant contextual information, such as the product's price, discount information, and the user's historical evaluation of the product. The server combines or represents the user information, candidate products, and their contextual information to construct target data for input into the model.

[0182] The server inputs the target data into the exposure click rate prediction model and the target conversion rate prediction model to predict the user's click probability (i.e., click prediction result) and conversion probability (i.e., conversion prediction result) for the candidate products, and recommends more suitable products to the user based on the click probability and the target conversion result determined by the conversion probability.

[0183] The data processing method provided in the embodiments of this specification can provide users with more personalized and accurate recommendation services through such a process, thereby improving user satisfaction and the conversion rate of the platform.

[0184] In one or more embodiments of the present specification, the step of using the target conversion rate estimation model to obtain a conversion prediction result corresponding to the target data includes:

[0185] Obtaining target features of the target data using a shared feature extraction network of the target conversion rate estimation model;

[0186] The target features are input into the click-through conversion rate estimation network of the target conversion rate estimation model to obtain the conversion prediction result corresponding to the target data.

[0187] Specifically, in the target conversion rate prediction model, the role of the shared feature extraction network is to extract features related to the target data that are useful for predicting the conversion probability. These features may include basic information of the user, historical behavior, attributes of the candidate object, etc. The click-through conversion rate prediction network is a network that predicts the click-through conversion rate based on the extracted features. The click-through conversion rate prediction network receives the target features from the shared feature extraction network as input and outputs a predicted value, which represents the conversion probability corresponding to the target data.

[0188] In actual applications, when the target data is input into the target conversion rate prediction model, it will first be sent to the shared feature extraction network for processing to extract features related to the target data, namely the target features; the extracted target features are then input into the click-through conversion rate prediction network, which uses the target features to calculate the conversion probability corresponding to the target data and output a conversion prediction result.

[0189] The data processing method provided in the embodiments of this specification, the shared feature extraction network can capture key information in the target data, which is crucial for accurately predicting the user's conversion behavior. When the target features are accurately extracted and input into the click-through conversion rate prediction network, the model can more accurately predict whether the user will convert, thereby improving the accuracy and reliability of the prediction.

[0190] In one or more embodiments of the present specification, after obtaining the target conversion result of the target data according to the conversion prediction result and the click prediction result, the method further includes:

[0191] According to the target conversion result of the target data, the multiple candidate objects are sorted to obtain an object recommendation result, and the object recommendation result is returned to the client to be displayed on a user interaction interface of the client.

[0192] Multiple candidate objects are sorted by target conversion results, so that the candidate objects most likely to be converted by users are placed in front to improve the recommendation effect and user satisfaction. After the sorting is completed, a certain number of candidate objects are selected as object recommendation results based on the sorting results, and the object recommendation results are returned to the client. After receiving the object recommendation results, the client will display these recommended objects on the user interaction interface for users to view and select.

[0193] It should be noted that when the target conversion rate estimation model is deployed in a production environment and real-time recommendations are performed, a model monitoring mechanism can be established to regularly check the performance and stability of the target conversion rate estimation model to ensure that the target conversion rate estimation model can continue to provide high-quality object recommendation results. At predetermined intervals (for example, every day), the model parameters are adjusted or the model is retrained based on feedback to adapt to new data and changes in user behavior.

[0194] The data processing method provided in the embodiments of this specification can more accurately understand the user's preferences for candidate objects by comprehensively considering the conversion prediction results and click prediction results, which helps to optimize the personalized recommendation algorithm and provide users with recommended content that is more in line with their interests and needs.

[0195] See also Figure 4 , Figure 4A flowchart of an object recommendation method provided by an embodiment of the present specification is shown, which specifically includes the following steps.

[0196] Step 402: In response to an object recommendation request sent by a client, determining user information corresponding to the client, and determining a plurality of candidate objects and context information according to the user information;

[0197] Step 404: determining target data according to the user information, the multiple candidate objects and the context information, and inputting the target data into a target conversion rate estimation model and an exposure click rate estimation model respectively;

[0198] Step 406: using the target conversion rate estimation model to obtain a conversion prediction result corresponding to the target data, and using the exposure click rate estimation model to obtain a click prediction result corresponding to the target data;

[0199] Step 408: Obtaining a target conversion result of the target data according to the conversion prediction result and the click prediction result;

[0200] Step 410: sort the multiple candidate objects according to the target conversion result to obtain an object recommendation result, and return the object recommendation result to the client to display the object recommendation result on the user interaction interface of the client.

[0201] The specific implementation method can be found in the above embodiments, which will not be described in detail here.

[0202] The object recommendation method provided in the embodiments of this specification can more accurately understand the user's preferences for candidate objects by comprehensively considering the conversion prediction results and click prediction results, which helps to optimize the personalized recommendation algorithm and provide users with recommended content that is more in line with their interests and needs. By displaying the object recommendation results on the user interaction interface of the client, it is also convenient for users to view and select.

[0203] See also Figure 5 , Figure 5 A schematic diagram of sample space improvement provided by an embodiment of the present specification is shown.

[0204] Specifically, the CVR model uses the "click and convert" samples as positive samples, which will make the prediction target of the CVR model "the probability of conversion among the samples that are exposed and clicked, that is, P (商卡转化|商卡被点击，曝光的商卡) However, it is impossible to predict the probability of conversion if the samples that were exposed but not clicked were clicked, that is, P (商卡转化|商卡未被点击，曝光的商卡) ” Such training data ignores the distribution of samples that have not been clicked, resulting in a large deviation in the model’s training sample space and greatly reducing the model’s generalization ability.

[0205] Obtain click training data corresponding to the task of estimating click rate (i.e., the exposure-click training data in the above embodiment) and conversion training data corresponding to the task of estimating conversion rate (i.e., the click-to-conversion training data in the above embodiment); wherein, the click training data includes: CTR positive samples and CTR negative samples, and the conversion training data includes: CVR positive samples and CVR negative samples (each sample consists of user features, product features, context features, and positive and negative labels), and adjust the training data based on the click training data and the conversion training data.

[0206] If a sample is a clicked and converted sample, it is a positive sample of CTR and a positive sample of CVR; if a sample is a clicked but not converted sample, it is a positive sample of CTR and a negative sample of CVR; if a sample is an exposed but not clicked sample, it is a negative sample of CTR.

[0207] By reusing the "exposed but not clicked" samples, the sample space deviation can be effectively repaired. Specifically, the probability formula is used to train and derive the CVR model and derive the conversion probability:

[0208] P (商卡转化，商卡被点击|曝光的商卡) =P (商卡被点击|曝光的商卡) *P (商卡转化|商卡被点击，曝光的商卡)

[0209] In the above formula, P (商卡被点击|曝光的商卡) is the modeling prediction result of the CTR model, P (商卡转化|商卡被点击，曝光的商卡) It is the modeling prediction result of the CVR model; the conversion rate is regarded as the joint probability of the two events of click and conversion, and the sample space is expanded; that is, by constructing CTR as an auxiliary training task, based on the joint probability, the modeling of the sample space based on the full exposure of the business card is realized, and the training positive samples of the click-to-conversion rate estimation model are determined according to the samples of click and conversion (that is, exposure and conversion); the training negative samples of the initial conversion rate estimation model are determined according to the samples of exposure but not click and the samples of click but not conversion (that is, exposure but not conversion).

[0210] When the original CVR model was trained, the training data only included samples of clicks and conversions (positive samples) and samples of clicks but not converted (negative samples). By adding samples of exposure but not converted (negative samples) and expanding the sample space, the problem of sparse training sample space was solved. Through cross-space sample multi-task learning and the use of multi-task learning training methods, the cross-space samples were effectively utilized, thereby improving the generalization of the model.

[0211] See also Figure 6 , Figure 6 A schematic diagram of the structure of a multi-task learning framework provided by an embodiment of the present specification is shown.

[0212] Among them, the sparse embedding layer (Spares Embedding, i.e. the embedding layer of the above embodiment) and the dense neural network layer (Dense DNN Network, i.e. the network layer of the above embodiment) together constitute a shared feature extraction network. The sparse embedding layer is usually used to process high-dimensional and sparse category features (such as user information, product information, etc.). It reduces the feature dimension by mapping the information of each category to a low-dimensional vector space and captures the potential relationship between categories; the dense neural network layer refers to a deep neural network composed of multiple fully connected layers, which can process continuous or embedded dense features. This network is capable of performing complex nonlinear transformations to extract higher-level abstract features from the original features.

[0213] A multi-task learning network is constructed based on the shared feature extraction network, the CTR estimation layer (i.e., the CTR estimation task network of the above embodiment) and the CVR estimation layer (i.e., the CVR estimation task network of the above embodiment), and multi-task joint training is performed. Based on the shared feature extraction network and the CVR estimation task network after network parameter adjustment, an optimized CVR model is obtained, and the optimized CVR model is determined as the target conversion rate estimation model.

[0214] During the training process, the dense features (dense neural network layer) derived from the intermediate steps are combined with the sparse features (sparse embedding layer) that are finally trained to form an online inference model. The impact of different export steps on model performance (such as AUC) is observed, and the dense features derived at a specific number of steps are selected to optimize the model performance.

[0215] The most suitable dense derivation step can be determined experimentally, which can be achieved by training for multiple days and printing out the AUC change graph, so as to select the most suitable dense derivation step for each scenario. Figure 7 As shown, Figure 7 A schematic diagram showing evaluation results of a click-through conversion rate prediction model corresponding to different training steps provided in an embodiment of the present specification is shown.

[0216] When evaluating the trained model, the evaluation indicators include but are not limited to AUC, accuracy, recall and other indicators to ensure the effectiveness of the model in practical applications. Based on the evaluation results, the model with better performance is selected.

[0217] Start training the model according to the set training parameters (such as batch size, number of training steps, etc.), and export the model parameters corresponding to the current training steps after each specific number of steps (such as every 10 steps, every 50 steps, etc.); after all data samples are trained, use the trained model (the model of the last step) to obtain the sparse embedding layer; use the validation set or new test data to perform performance evaluation on the models at different export stages (i.e., the click-through conversion rate estimation model corresponding to the initial training step), such as calculating indicators such as AUC and accuracy; determine the evaluation results of the model by comparing the accuracy of the prediction results of the models corresponding to different export steps; select the best performing model based on the evaluation results, for example, when the training steps are 200, the model parameters corresponding to the 100th step may perform best; determine the dense neural network layer based on the determined model with the best performance; determine the target conversion rate estimation model based on the dense neural network layer and the sparse embedding layer of the model corresponding to the last step.

[0218] By using the probability formula, without changing the original CVR model and CTR model training method, the full amount of exposure business card sample space is simply and efficiently introduced to the CVR model, solving the problem of sample selection bias in traditional CVR model training. This means that the target conversion rate estimation model obtained by training using the model training method provided in the embodiment of this specification can be better generalized to the representation space of all samples, not just the business card representation after the click event occurs. A new training method is adopted to increase the model's generalization ability for new materials, and internal experiments have shown that the target conversion rate estimation model is significantly better than existing methods in CVR tasks, proving its effectiveness in practical applications.

[0219] Corresponding to the above method embodiment, this specification also provides a model training device embodiment, Figure 8 FIG. 1 shows a schematic diagram of a model training device provided by an embodiment of the present specification. Figure 8 As shown, the device comprises:

[0220] The data determination module 802 is configured to determine the exposure-conversion training data according to the exposure-click training data corresponding to the exposure-click rate estimation model and the click-conversion training data corresponding to the click-conversion rate estimation model;

[0221] The result prediction module 804 is configured to extract the exposure conversion features of the exposure conversion training data, and determine the target prediction result by using the click prediction result output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction result output by the click conversion rate prediction model based on the exposure conversion features;

[0222] The model training module 806 is configured to train the click-through conversion rate estimation model according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

[0223] Optionally, the model training module 806 is further configured to:

[0224] Determining a plurality of initial training steps, exposure conversion training data corresponding to each of the plurality of initial training steps, and a target prediction result of the exposure conversion training data corresponding to each of the initial training steps;

[0225] Obtaining model parameters of the click-through conversion rate estimation model corresponding to each initial training step according to the target prediction results corresponding to each initial training step and the exposure conversion labels in the exposure conversion training data;

[0226] A first target model parameter and a second target model parameter are determined from the model parameters corresponding to each initial training step, and the target conversion rate estimation model is obtained according to the first target model parameter and the second target model parameter.

[0227] Optionally, the model training module 806 is further configured to:

[0228] Determining the total training steps for the click-through conversion rate estimation model according to the data volume of the exposure-to-conversion training data and the preset data volume for each round of training;

[0229] According to a preset step number interval, a plurality of initial training steps are determined from the total training steps.

[0230] Optionally, the model training module 806 is further configured to:

[0231] Determine, according to the model parameters corresponding to each initial training step, a click-through conversion rate estimation model corresponding to each initial training step, wherein the click-through conversion rate estimation model corresponding to each initial training step is obtained by training according to the target prediction results corresponding to each initial training step and the exposure conversion labels in the exposure conversion training data;

[0232] Evaluate the click-through conversion rate estimation models corresponding to the initial training steps to obtain evaluation results, and determine the click-through conversion rate estimation model corresponding to the first target training step from the click-through conversion rate estimation models corresponding to the multiple initial training steps according to the evaluation results, wherein the evaluation result of the click-through conversion rate estimation model corresponding to the first target training step is greater than the evaluation results of the click-through conversion rate estimation models corresponding to other training steps;

[0233] Determine the first target model parameters according to the click-through conversion rate estimation model corresponding to the first target training step, and determine the second target model parameters according to the click-through conversion rate estimation model corresponding to the second target training step, wherein the second target training step is the training step at which the model training iteration ends;

[0234] The target conversion rate estimation model is obtained according to the first target model parameters and the second target model parameters.

[0235] Optionally, the model training module 806 is further configured to:

[0236] Determining the network parameters of the network layer in the click-through conversion rate estimation model corresponding to the first target training step as the first target model parameters;

[0237] Determine the network parameters of the embedding layer in the click-through conversion rate estimation model corresponding to the second target training step as the second target model parameters;

[0238] The target conversion rate estimation model is obtained according to the network layer having the first target model parameters and the embedding layer having the second target model parameters.

[0239] Optionally, the result prediction module 804 is further configured to:

[0240] Inputting the exposure conversion training data into a shared feature extraction network, and using the shared feature extraction network to perform feature extraction on the exposure conversion training data to obtain exposure conversion features of the exposure conversion training data;

[0241] Inputting the exposure conversion feature into the exposure click rate prediction model to obtain the click prediction result output by the exposure click rate prediction network;

[0242] Inputting the exposure conversion feature into the click-through conversion rate estimation model to obtain the conversion prediction result output by the click-through conversion rate estimation network;

[0243] The target prediction result is obtained according to the click prediction result and the conversion prediction result.

[0244] Optionally, the model training module 806 is further configured to:

[0245] Determining a target loss function according to the target prediction result and the exposure conversion label in the exposure conversion training data;

[0246] According to the target loss function, adjusting the model parameters of the click-through conversion rate estimation model;

[0247] The target conversion rate estimation model is determined according to the click-through conversion rate estimation model after the model parameters are adjusted.

[0248] The device further comprises:

[0249] The data acquisition module is configured to determine an initial data set; determine the samples that are clicked and converted in the initial data set as positive samples of the exposure-click rate prediction model and the click-to-conversion rate prediction model, determine the samples that are clicked but not converted in the initial data set as positive samples of the exposure-click rate prediction model and negative samples of the click-to-conversion rate prediction model, and determine the samples that are exposed but not clicked in the initial data set as negative samples of the exposure-click rate prediction model; determine the exposure-click training data corresponding to the exposure-click rate prediction model based on the positive samples and negative samples of the exposure-click rate prediction model; and determine the click-to-conversion training data corresponding to the click-to-conversion rate prediction model based on the positive samples and negative samples of the exposure-click rate prediction model.

[0250] The above is a schematic scheme of a model training device of this embodiment. It should be noted that the technical scheme of the model training device and the technical scheme of the above-mentioned model training method belong to the same concept, and the details not described in detail in the technical scheme of the model training device can be referred to the description of the technical scheme of the above-mentioned model training method.

[0251] Corresponding to the above method embodiment, this specification also provides a data processing device embodiment, Fig. 9 FIG. 1 is a schematic diagram showing the structure of a data processing device provided by an embodiment of the present specification. Fig. 9 As shown, the device comprises:

[0252] Model input module 902, configured to input target data into target conversion rate estimation model and exposure click rate estimation model respectively;

[0253] The first prediction module 904 is configured to obtain a conversion prediction result corresponding to the target data by using the target conversion rate estimation model;

[0254] The second prediction module 906 is configured to obtain a click prediction result corresponding to the target data by using the exposure click rate prediction model;

[0255] The result obtaining module 908 is configured to obtain the target conversion result of the target data according to the conversion prediction result and the click prediction result.

[0256] Optionally, the first prediction module 904 is further configured to:

[0257] Obtaining target features of the target data using a shared feature extraction network of the target conversion rate estimation model;

[0258] The target features are input into the click-through conversion rate estimation network of the target conversion rate estimation model to obtain the conversion prediction result corresponding to the target data.

[0259] The device further comprises:

[0260] The response module is configured to respond to an object recommendation request sent by a client, determine the user information corresponding to the client, determine multiple candidate objects and context information of each candidate object based on the user information; and determine the target data based on the user information, the multiple candidate objects and the context information of each candidate object.

[0261] The device further comprises:

[0262] The return module is configured to sort the multiple candidate objects according to the target conversion result of the target data, obtain object recommendation results, and return the object recommendation results to the client to display the object recommendation results on the user interaction interface of the client.

[0263] The above is a schematic scheme of a data processing device of this embodiment. It should be noted that the technical scheme of the data processing device and the technical scheme of the above data processing method belong to the same concept, and the details of the technical scheme of the data processing device that are not described in detail can all be referred to the description of the technical scheme of the above data processing method.

[0264] Corresponding to the above method embodiment, this specification also provides an object recommendation device embodiment, Fig.10 FIG. 1 is a schematic diagram showing the structure of an object recommendation device provided by an embodiment of the present specification. Fig.10 As shown, the device comprises:

[0265] A response module 1002 is configured to respond to an object recommendation request sent by a client, determine user information corresponding to the client, and determine a plurality of candidate objects and context information of each candidate object according to the user information;

[0266] An input module 1004 is configured to determine target data according to the user information, the multiple candidate objects and context information of each candidate object, and input the target data into a target conversion rate estimation model and an exposure click rate estimation model respectively;

[0267] The prediction module 1006 is configured to obtain a conversion prediction result corresponding to the target data by using the target conversion rate prediction model, and to obtain a click prediction result corresponding to the target data by using the exposure click rate prediction model;

[0268] An acquisition module 1008 is configured to acquire a target conversion result of the target data according to the conversion prediction result and the click prediction result;

[0269] The return module 1010 is configured to sort the multiple candidate objects according to the target conversion result, obtain an object recommendation result, and return the object recommendation result to the client to display the object recommendation result on the user interaction interface of the client.

[0270] The above is a schematic scheme of an object recommendation device of this embodiment. It should be noted that the technical scheme of the object recommendation device and the technical scheme of the object recommendation method described above belong to the same concept, and the details not described in detail in the technical scheme of the object recommendation device can be referred to the description of the technical scheme of the object recommendation method described above.

[0271] Fig.11 The block diagram of a computing device 1100 according to one embodiment of the present specification is shown. The components of the computing device 1100 include but are not limited to a memory 1110 and a processor 1120. The processor 1120 is connected to the memory 1110 via a bus 1130, and a database 1150 is used to store data.

[0272] The computing device 1100 also includes an access device 1140 that enables the computing device 1100 to communicate via one or more networks 1160. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1140 may include one or more of any type of network interface, wired or wireless (e.g., a network interface card (NIC)), such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, and a Near Field Communication (NFC).

[0273] In one embodiment of the present specification, the above components of the computing device 1100 and Fig.11 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Fig.11 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0274] The computing device 1100 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 1100 may also be a mobile or stationary server.

[0275] Among them, the processor 1120 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the above-mentioned model training method, data processing method and object recommendation method.

[0276] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the computing device embodiment, since it is basically similar to the model training method, data processing method and object recommendation method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the model training method, data processing method and object recommendation method embodiments.

[0277] An embodiment of the present specification also provides a computer-readable storage medium storing a computer program / instruction, which implements the steps of the above-mentioned model training method, data processing method and object recommendation method when executed by a processor.

[0278] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the computer-readable storage medium embodiment, since it is basically similar to the model training method, data processing method and object recommendation method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the model training method, data processing method and object recommendation method embodiments.

[0279] An embodiment of the present specification also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned model training method, data processing method and object recommendation method.

[0280] The above is a schematic scheme of a computer program product of this embodiment. It should be noted that the technical scheme of the computer program product and the technical schemes of the above-mentioned model training method, data processing method and object recommendation method belong to the same concept, and the details not described in detail in the technical scheme of the computer program product can be referred to the description of the technical schemes of the above-mentioned model training method, data processing method and object recommendation method.

[0281] The above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0282] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0283] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0284] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0285] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. A model training method, comprising: Determine exposure conversion training data according to the exposure click rate prediction model's exposure click training data and the click conversion training data corresponding to the click conversion rate prediction model; Extracting exposure conversion features of the exposure conversion training data, and using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features to determine a target prediction result; The click-through conversion rate estimation model is trained according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

2. According to the model training method of claim 1, the step of training the click-through conversion rate estimation model based on the target prediction result and the exposure conversion label in the exposure conversion training data to obtain the target conversion rate estimation model comprises: Determining a plurality of initial training steps, exposure conversion training data corresponding to each of the plurality of initial training steps, and a target prediction result of the exposure conversion training data corresponding to each of the initial training steps; Obtaining model parameters of the click-through conversion rate estimation model corresponding to each initial training step according to the target prediction results corresponding to each initial training step and the exposure conversion labels in the exposure conversion training data; A first target model parameter and a second target model parameter are determined from the model parameters corresponding to each initial training step, and the target conversion rate estimation model is obtained according to the first target model parameter and the second target model parameter.

3. The model training method according to claim 2, wherein determining a plurality of initial training steps comprises: Determining the total training steps for the click-through conversion rate estimation model according to the data volume of the exposure-to-conversion training data and the preset data volume for each round of training; According to a preset step number interval, a plurality of initial training steps are determined from the total training steps.

4. The model training method according to claim 2, wherein determining a first target model parameter and a second target model parameter from the model parameters corresponding to each initial training step, and obtaining the target conversion rate estimation model according to the first target model parameter and the second target model parameter comprises: Determine, according to the model parameters corresponding to each initial training step, a click-through conversion rate estimation model corresponding to each initial training step, wherein the click-through conversion rate estimation model corresponding to each initial training step is obtained by training according to the target prediction results corresponding to each initial training step and the exposure conversion labels in the exposure conversion training data; Evaluate the click-through conversion rate estimation models corresponding to the initial training steps to obtain evaluation results, and determine the click-through conversion rate estimation model corresponding to the first target training step from the click-through conversion rate estimation models corresponding to the multiple initial training steps according to the evaluation results, wherein the evaluation result of the click-through conversion rate estimation model corresponding to the first target training step is greater than the evaluation results of the click-through conversion rate estimation models corresponding to other training steps; Determine the first target model parameters according to the click-through conversion rate estimation model corresponding to the first target training step, and determine the second target model parameters according to the click-through conversion rate estimation model corresponding to the second target training step, wherein the second target training step is the training step at which the model training iteration ends; The target conversion rate estimation model is obtained according to the first target model parameters and the second target model parameters.

5. According to the model training method of claim 4, the click-through conversion rate prediction model includes a network layer and an embedding layer; The step of determining the first target model parameters according to the click-through conversion rate estimation model corresponding to the first target training step, and determining the second target model parameters according to the click-through conversion rate estimation model corresponding to the second target training step, includes: Determining the network parameters of the network layer in the click-through conversion rate estimation model corresponding to the first target training step as the first target model parameters; Determine the network parameters of the embedding layer in the click-through conversion rate estimation model corresponding to the second target training step as the second target model parameters; The step of obtaining the target conversion rate estimation model according to the first target model parameter and the second target model parameter includes: The target conversion rate estimation model is obtained according to the network layer having the first target model parameters and the embedding layer having the second target model parameters.

6. According to the model training method of claim 1, the step of extracting the exposure conversion features of the exposure conversion training data comprises: Inputting the exposure conversion training data into a shared feature extraction network, and using the shared feature extraction network to perform feature extraction on the exposure conversion training data to obtain exposure conversion features of the exposure conversion training data; The step of determining a target prediction result by using the click prediction result output by the exposure-to-click rate prediction model based on the exposure-to-conversion feature and the conversion prediction result output by the click-to-conversion rate prediction model based on the exposure-to-conversion feature includes: Inputting the exposure conversion feature into the exposure click rate prediction model to obtain the click prediction result output by the exposure click rate prediction network; Inputting the exposure conversion feature into the click-through conversion rate estimation model to obtain the conversion prediction result output by the click-through conversion rate estimation network; The target prediction result is obtained according to the click prediction result and the conversion prediction result.

7. According to the model training method according to any one of claims 1 to 6, the step of training the click-through conversion rate estimation model based on the target prediction result and the exposure conversion label in the exposure conversion training data to obtain the target conversion rate estimation model comprises: Determining a target loss function according to the target prediction result and the exposure conversion label in the exposure conversion training data; According to the target loss function, adjusting the model parameters of the click-through conversion rate estimation model; The target conversion rate estimation model is determined according to the click-through conversion rate estimation model after the model parameters are adjusted.

8. The model training method according to any one of claims 1 to 6, before determining the exposure-conversion training data based on the exposure-click rate prediction model and the click-conversion training data corresponding to the click-conversion rate prediction model, the method further comprises: Determine the initial data set; Determine the samples that are clicked and converted in the initial data set as positive samples of the exposure-to-click rate estimation model and the click-to-conversion rate estimation model, determine the samples that are clicked but not converted in the initial data set as positive samples of the exposure-to-click rate estimation model and negative samples of the click-to-conversion rate estimation model, and determine the samples that are exposed but not clicked in the initial data set as negative samples of the exposure-to-click rate estimation model; Determining exposure-click training data corresponding to the exposure-click rate prediction model according to the positive samples and negative samples of the exposure-click rate prediction model; According to the positive samples and negative samples of the click-through conversion rate prediction model, click-through conversion training data corresponding to the click-through conversion rate prediction model is determined.

9. A data processing method, comprising: Input the target data into the target conversion rate estimation model and the exposure click rate estimation model respectively; Using the target conversion rate estimation model, obtaining a conversion prediction result corresponding to the target data; Using the exposure click rate prediction model, obtaining a click prediction result corresponding to the target data; A target conversion result of the target data is obtained according to the conversion prediction result and the click prediction result.

10. The data processing method according to claim 9, wherein the step of using the target conversion rate estimation model to obtain the conversion prediction result corresponding to the target data comprises: Obtaining target features of the target data using a shared feature extraction network of the target conversion rate estimation model; The target features are input into the click-through conversion rate estimation network of the target conversion rate estimation model to obtain the conversion prediction result corresponding to the target data.

11. The data processing method according to claim 9, before inputting the target data into the target conversion rate estimation model and the exposure click rate estimation model, further comprising: In response to an object recommendation request sent by a client, determining user information corresponding to the client, and determining a plurality of candidate objects and context information of each candidate object according to the user information; Determining the target data according to the user information, the multiple candidate objects, and context information of each candidate object; After obtaining the target conversion result of the target data according to the conversion prediction result and the click prediction result, the method further includes: According to the target conversion result of the target data, the multiple candidate objects are sorted to obtain an object recommendation result, and the object recommendation result is returned to the client to be displayed on a user interaction interface of the client.

12. An object recommendation method, comprising: In response to an object recommendation request sent by a client, determining user information corresponding to the client, and determining a plurality of candidate objects and context information of each candidate object according to the user information; Determine target data according to the user information, the multiple candidate objects, and context information of each candidate object, and input the target data into a target conversion rate estimation model and an exposure click rate estimation model respectively; Using the target conversion rate prediction model, obtaining a conversion prediction result corresponding to the target data, and using the exposure click rate prediction model, obtaining a click prediction result corresponding to the target data; Obtaining a target conversion result of the target data according to the conversion prediction result and the click prediction result; According to the target conversion result, the multiple candidate objects are sorted to obtain an object recommendation result, and the object recommendation result is returned to the client to display the object recommendation result on a user interaction interface of the client.

13. A model training method, applied in the cloud, comprising: Receiving a model training request sent by the device end, and determining exposure-conversion training data in response to the model training request and according to the exposure-click training data corresponding to the exposure-click rate estimation model and the click-conversion training data corresponding to the click-conversion rate estimation model; Extracting exposure conversion features of the exposure conversion training data, and using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features to determine a target prediction result; According to the target prediction result and the exposure conversion label in the exposure conversion training data, the click-through conversion rate estimation model is trained to obtain a target conversion rate estimation model; The target conversion rate estimation model is sent to the device end to deploy the target conversion rate estimation model on the device end.

14. A model training device, comprising: A data determination module is configured to determine exposure-conversion training data according to the exposure-click training data corresponding to the exposure-click rate estimation model and the click-conversion training data corresponding to the click-conversion rate estimation model; A result prediction module is configured to extract the exposure conversion features of the exposure conversion training data, and determine a target prediction result by using the click prediction results output by the exposure click rate prediction model based on the exposure conversion features and the conversion prediction results output by the click conversion rate prediction model based on the exposure conversion features; The model training module is configured to train the click-through conversion rate estimation model according to the target prediction result and the exposure conversion label in the exposure conversion training data to obtain a target conversion rate estimation model.

15. A data processing device, comprising: A model input module is configured to input target data into a target conversion rate estimation model and an exposure click rate estimation model respectively; A first prediction module is configured to obtain a conversion prediction result corresponding to the target data by using the target conversion rate estimation model; A second prediction module is configured to obtain a click prediction result corresponding to the target data by using the exposure click rate prediction model; The result obtaining module is configured to obtain the target conversion result of the target data according to the conversion prediction result and the click prediction result.

16. An object recommendation device, comprising: A response module is configured to respond to an object recommendation request sent by a client, determine user information corresponding to the client, and determine a plurality of candidate objects and context information of each candidate object according to the user information; An input module is configured to determine target data according to the user information, the multiple candidate objects and context information of each candidate object, and input the target data into a target conversion rate estimation model and an exposure click rate estimation model respectively; A prediction module is configured to obtain a conversion prediction result corresponding to the target data by using the target conversion rate prediction model, and to obtain a click prediction result corresponding to the target data by using the exposure click rate prediction model; An acquisition module, configured to acquire a target conversion result of the target data according to the conversion prediction result and the click prediction result; The return module is configured to sort the multiple candidate objects according to the target conversion result, obtain an object recommendation result, and return the object recommendation result to the client so as to display the object recommendation result on the user interaction interface of the client.

17. A computing device comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer program / instructions are executed by the processor, the steps of the method described in any one of claims 1 to 13 are implemented.

18. A computer-readable storage medium storing a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the steps of the method according to any one of claims 1 to 13.

19. A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 13.

Images