Click rate estimation method and device, electronic equipment and storage medium

Abstract

Embodiments of the invention provide a click rate estimation method and apparatus, an electronic device and a storage medium. The method comprises the steps of determining a recall object set for a target user; according to the user feature of the target user and the object feature of each recall object, a first feature sequence set is constructed, and each feature sequence in the first feature sequence set comprises the user feature of the target user and the object feature of one recall object; inputting the first feature sequence set into a click rate estimation model; carrying out feature normalization processing on the first feature sequence set by utilizing at least two normalization networks of a click rate estimation model to obtain at least two normalized feature sequence sets, and carrying out fusion processing on the at least two normalized feature sequence sets by utilizing a feature fusion network of the click rate estimation model to obtain at least two fused feature sequence sets; obtaining a second feature sequence set; and estimating the click rate of each recalled object by the target user based on the second feature sequence set by using a click rate estimation network of the click rate estimation model.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of Internet technology, and in particular to a click rate prediction method, device, electronic device and storage medium. Background Art

[0002] In the context of big data and deep learning, the click-through rate prediction model is a core component of the recommendation system. The effect of click-through rate prediction is often positively correlated with the richness of features, that is, the richer the features, the better the prediction effect of the model. However, in order to achieve real-time response, the recommendation system has strict requirements on the time of model training and inference. In order to solve the problem of increased response time caused by feature richness, normalization technology is widely used. Using normalization technology to process data can improve the response speed of the recommendation system.

[0003] Common normalization techniques include batch normalization (BN) and layer normalization (LN). Batch normalization is to calculate the mean and standard deviation of the feature dimension on each batch of samples, and normalize the features based on the calculated mean and standard deviation. Layer normalization is to calculate the mean and standard deviation of all feature dimensions of a single sample, and normalize the features based on the calculated mean and standard deviation.

[0004] In the process of implementing the present invention, the inventors found that the current click rate prediction model usually adopts a single normalized network, for example, only a batch normalized network or only a layer normalized network. Batch normalization is unstable when the batch size is small. Layer normalization does not depend on the batch size, but it is easy to ignore the relationship between layers or samples. That is, each normalization technology has its advantages and scope of application. Using a single normalized network is difficult to cope with a variety of data distribution and dynamic changes, and the model performance is poor. Summary of the invention

[0005] The embodiments of the present invention provide a click rate prediction method, device, electronic device and storage medium, which can cope with various data distribution and dynamic change scenarios and improve the model prediction effect.

[0006] In a first aspect, a click rate estimation method provided by an embodiment of the present invention includes:

[0007] Determine a recall object set for the target user, wherein the recall object set includes at least two recall objects;

[0008] constructing a first feature sequence set according to the user feature of the target user and the object feature of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user feature of the target user and the object feature of one of the at least two recalled objects;

[0009] Inputting the first feature sequence set into a click rate prediction model, wherein the click rate prediction model includes a hybrid normalization network and a click rate prediction network, wherein the hybrid normalization network includes at least two normalization networks and a feature fusion network;

[0010] Using the at least two normalization networks to perform feature normalization processing on the first feature sequence set respectively to obtain at least two normalized feature sequence sets, and using the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a second feature sequence set;

[0011] The click rate prediction network is used to predict the click rate of the target user for each of the recalled objects based on the second feature sequence set.

[0012] The above-mentioned click rate prediction method constructs a first feature sequence set according to the user features of the target user and the object features of the recalled object, forming a diversified feature space, which can more comprehensively describe the relationship between the user and the object, more accurately capture the user's interest points, and help the model to more comprehensively and deeply understand the user behavior pattern, and improve the accuracy of click rate prediction; the user features of the target user are combined with the object features of different recalled objects to construct a feature sequence set, forming batch features, which provides data support for the subsequent use of normalized networks, so that not only layer normalized networks but also batch normalized networks can be used in the future. The click rate prediction model includes at least two normalized networks and a feature fusion network, and at least two normalized networks are used to perform feature normalization processing on the first feature sequence set respectively. Through the mixed normalized network, the feature sequence is normalized from multiple angles, which enhances the diversity of feature expression and maximizes the retention of useful information of the features, effectively reduces the instability and limitations that may be brought about by a single normalized network, and enhances the overall stability and adaptability of the model. With the help of the feature fusion network, the outputs of multiple normalized networks are optimized and combined, further improving the adaptability and performance of the model. The types of at least two normalized networks can be the same or different. When the types of at least two normalized networks are different, different types of normalized networks adapt to different data distributions, can more widely cover a variety of feature distributions, and improve the adaptability of the model; through multi-class normalization methods, different types of features can be processed more comprehensively and deeply, different types of normalization methods can complement each other, and the use of different types of normalized networks can process features from different angles, thereby improving the generalization ability of the model under multiple data distributions. When the types of at least two normalized networks are the same, since the normalized networks have learnable parameters, after training, the learnable parameters of at least two normalized networks will be different, and for the same input, the outputs of at least two normalized networks will also be different, achieving the effect of capturing different aspects and changes of the input data, and being able to effectively process dynamically changing input data; by processing features in parallel through at least two normalized networks of the same type, the consistency and stability of feature processing can be improved, and in the case of random noise and data anomalies, multiple normalized networks can complement and correct each other, and multiple normalization ensures that feature information is sufficient and complete, so that it can cope with a variety of data distributions and improve the expressiveness of the model. The click rate prediction network is used to estimate the user's click rate for each recalled object based on the fused feature sequence set, so that the click rate prediction network can understand user behavior and interests more comprehensively and accurately, improve the accuracy of click rate prediction, and enhance user satisfaction and usage experience, thereby improving the overall performance of the system and user retention rate.

[0013] In a second aspect, an embodiment of the present invention provides a click rate estimation device, comprising:

[0014] An object determination module, used to determine a recall object set for a target user, wherein the recall object set includes at least two recall objects;

[0015] a feature construction module, configured to construct a first feature sequence set according to the user feature of the target user and the object feature of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user feature of the target user and the object feature of one of the at least two recalled objects;

[0016] An execution module is used to input the first feature sequence set into a click-through rate prediction model, the click-through rate prediction model includes a hybrid normalization network and a click-through rate prediction network, the hybrid normalization network includes at least two normalization networks and a feature fusion network; using the at least two normalization networks to perform feature normalization processing on the first feature sequence set respectively to obtain at least two normalized feature sequence sets, using the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a second feature sequence set; using the click-through rate prediction network to estimate the click rate of the target user for each of the recalled objects based on the second feature sequence set.

[0017] In a third aspect, an electronic device provided by an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, a click rate prediction method as in any embodiment of the present invention is implemented.

[0018] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, a click rate prediction method as in any embodiment of the present invention is implemented.

[0019] The description of the second, third and fourth aspects of the embodiments of the present invention can refer to the detailed description of the first aspect; and the beneficial effects of the description of the second, third and fourth aspects can refer to the beneficial effects analysis of the first aspect, which will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 is a flow chart of a click rate estimation method provided by an embodiment of the present invention;

[0022] Figure 2a is another flow chart of the click rate estimation method provided by an embodiment of the present invention;

[0023] Figure 2b This is a schematic diagram of an application scenario of the click rate prediction method provided by an embodiment of the present invention;

[0024] Figure 2c is a structural diagram of a click rate prediction model provided by an embodiment of the present invention;

[0025] Figure 2d is a structural diagram of a hybrid normalized network provided by an embodiment of the present invention;

[0026] Figure 3 It is a flowchart of a training method for a click rate prediction model provided by an embodiment of the present invention;

[0027] Figure 4 is a structural schematic diagram of a view click rate prediction device provided by an embodiment of the present invention;

[0028] Figure 5 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] Figure 1It is a flowchart of the click rate prediction method provided by an embodiment of the present invention. The click rate prediction method provided by an embodiment of the present invention can be applied to the scenario of estimating the click rate of the current user on multiple commodities, advertisements, articles, videos and other objects. The click rate prediction method can be executed by the click rate prediction device provided by an embodiment of the present invention, and the device can be implemented in software and / or hardware. In a specific embodiment, the device can be integrated into an electronic device, and the electronic device can be a computer, a server, etc. The following embodiment is described by taking the example of integrating the click rate prediction device into an electronic device, refer to Figure 1 The click rate estimation method of this embodiment may include the following steps:

[0032] Step 101: determine a recall object set for a target user, where the recall object set includes at least two recall objects.

[0033] The target user may be the user for whom the recommendation system is currently generating recommendations. The recall object set may be a set consisting of a small number of potential interest objects selected from a large number of possible objects before the recommendation system generates recommendations for the target user. The recall objects may include but are not limited to advertisements, commodities, products, articles, and videos. The recall objects may be different in different application scenarios. For example, on an e-commerce platform, if a user has recently browsed a number of electronic products, the recall objects may be a group of products related to the user's recent browsing history, interest tags, etc. selected from all electronic products; on a video platform, if a user often watches science fiction movies, the recall objects may be a group of science fiction movies or works of related directors and actors selected from all videos; on a news platform, if a user likes to read science and technology news, the recall objects may be a group of the latest and science and technology-related news selected from all news articles.

[0034] Specifically, user features of target users and object features of candidate objects can be collected. User features may include user basic features, user behavior features, etc. User basic features refer to features related to the user's basic personal information, which are usually static or less changing, and most of them are explicit data. User basic features can be obtained from the user's registration data on the platform, or by providing a questionnaire for the user to input. For example, user basic features may include the user's age, gender, occupation, geographic location, membership level, etc. User behavior features are data representations of various types of interactive behaviors of users on the platform, which are usually dynamic and can directly reflect the user's interests and preferences as time changes. For example, user behavior features may include browsing records, click records, purchase records, search records, favorites, likes, comments, etc. Object features may be features used in the recommendation system to describe the basic information or attributes of candidate objects. For example, object features may include object identification information, category information, tag information, function information, summary information, brand information, etc.

[0035] A user profile of the target user can be constructed based on the user's basic characteristics, user behavior characteristics, etc., to understand the user's interests and preferences. The user profile is matched with the object characteristics of the candidate object, and some algorithms and strategies are applied to filter out the recalled objects from the candidate objects, that is, to obtain a set of recalled objects for the target user. The algorithms and strategies here can be collaborative filtering, content-based recommendation, rule-based screening, hybrid recommendation, etc., which are not specifically limited here.

[0036] Step 102: construct a first feature sequence set based on the user features of the target user and the object features of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user features of the target user and the object features of one of the at least two recalled objects.

[0037] The first feature sequence set is a data set, in which each feature sequence or data point consists of the user features of the target user and the object features of a recalled object, that is, each feature sequence is a data set consisting of the user features of the target user and the object features of a recalled object. In a specific implementation, the user features of the target user and the object features of each recalled object can be stored separately, that is, a copy of the user features of the target user can be stored in advance. When needed, the user features of the target user and the object features of each recalled object can be taken out, and the user features of the target user can be copied multiple times. Each copy of the user features of the target user is matched with the object features of a recalled object to form a series of feature sequences. These feature sequences are the first feature sequence set.

[0038] For example, as shown in Table 1 below, assuming that there are 3 recalled objects in the recalled object set, the user features of the target user can be combined with the object features of each of the 3 recalled objects, respectively, to obtain 3 feature sequences. Feature sequence 1 can be formed by combining the user features of the target user and the object features of recalled object 1, feature sequence 2 can be formed by combining the user features of the target user and the object features of recalled object 2, and feature sequence 3 can be formed by combining the user features of the target user and the object features of recalled object 3.

[0039] Feature sequence User characteristics of target users Object characteristics of recalled objects Feature sequence 1 User characteristics of target users Object characteristics of recalled object 1 Feature sequence 2 User characteristics of target users Recall the object characteristics of object 2 Feature sequence 3 User characteristics of target users Object characteristics of recalled object 3

[0040] Table 1

[0041] The user features of the target user are combined with the object features of different recalled objects to construct a feature sequence set, forming batch features, which provides data support for the subsequent use of normalized networks, so that not only layer normalized networks but also batch normalized networks can be used.

[0042] Step 103: input the first feature sequence set into a click rate prediction model, where the click rate prediction model includes a hybrid normalization network and a click rate prediction network, where the hybrid normalization network includes at least two normalization networks and a feature fusion network.

[0043] The click rate prediction model is a model used to predict the probability of users clicking on the recalled object. The click rate prediction model can be trained in advance through a training sample set. The normalization network is used to normalize the input features so that the features have the same scale and distribution, which is convenient for subsequent model processing. The feature fusion network fuses the features processed by different normalization networks to generate a unified feature representation.

[0044] Common normalization networks include batch normalization network, layer normalization network, and group normalization network. Batch normalization is to calculate the mean and standard deviation of the feature dimension on each batch of samples, and normalize the features based on the calculated mean and standard deviation. Layer normalization is to calculate the mean and standard deviation of all feature dimensions of a single sample, and normalize the features based on the calculated mean and standard deviation. Group normalization network selects a number of groups, divides the features into multiple groups, calculates the mean and variance of each group of features, and normalizes the corresponding group features based on the calculated mean and standard deviation.

[0045] In this embodiment, the types of at least two normalized networks may be the same or different. For example, at least two normalized networks may both be batch normalized networks, or one of at least two normalized networks may be a batch normalized network and the other may be a layer normalized network. Each normalized network has adjustable parameters, and the adjustable parameters may be adjusted in the model training stage. Specifically, the adjustable parameters of the normalized network may include scaling parameters and offset parameters. Among them, the scaling parameter is used to adjust the amplitude of the normalized feature, and each feature has a corresponding scaling parameter; the offset parameter is used to adjust the offset of the normalized feature, and each feature has a corresponding offset parameter. The values ​​of the adjustable parameters of two different types of normalized networks are usually different. Two normalized networks of the same type will also obtain adjustable parameters with different values ​​through training. Therefore, whether it is two normalized networks of the same type or different types, there will be different normalized outputs for the same input data. In actual use, the selection and matching of the normalized network can be determined according to actual needs.

[0046] At least two normalized networks are of the same type, and parameter adjustment and problem troubleshooting are all concentrated on the same method, which is convenient for model debugging and optimization; it can maintain the consistency of normalization methods, reduce potential problems caused by inconsistencies, and thus improve the stability of the model. At least two normalized networks are of different types. Different types of normalization methods process data and tasks at different levels, which can help the model have stronger generalization capabilities when facing diverse features and differently distributed data, and can concentrate on the advantages and characteristics of multiple types of normalization methods. Different normalization methods process features in different ways, which can provide more angles of representation for the feature space, making the model capture features more comprehensive and rich, and helping to improve the accuracy of the model.

[0047] Step 104: perform feature normalization processing on the first feature sequence set respectively using at least two normalization networks to obtain at least two normalized feature sequence sets, and perform fusion processing on the at least two normalized feature sequence sets using a feature fusion network to obtain a second feature sequence set.

[0048] By using a normalization network to perform feature normalization processing on the first feature sequence set, a normalized feature sequence set can be obtained. For example, there are two normalization networks, and by using normalization network 1 to perform feature normalization processing on the first feature sequence set, normalized feature sequence set 1 can be obtained, and by using normalization network 2 to perform feature normalization processing on the first feature sequence set, normalized feature sequence set 2 can be obtained. At least two normalization networks can be used to perform feature normalization processing on the first feature sequence set in parallel to improve data processing efficiency.

[0049] The purpose of fusing at least two normalized feature sequence sets is to combine the features processed by multiple normalized networks, retain the information extracted by different normalized networks, and then form a new feature sequence set with greater expressiveness. The method for fusing at least two normalized feature sequence sets may be simple concatenation, averaging, weighted averaging, or a more complex neural network fusion method. If weighted averaging is used, the network weights of different normalized networks can be obtained by model learning, or they can be set randomly or manually. The second feature sequence set combines the information processed by different normalized networks, and has a stronger representation capability and richer information content.

[0050] Step 105: using the click rate estimation network to estimate the click rate of each recalled object by the target user based on the second feature sequence set.

[0051] The click rate prediction network starts from the second feature sequence set and estimates the click probability of each recalled object. The click rate prediction network can use a multi-layer perceptron, a long short-term memory network, etc. The click rate value range can be between [0,1].

[0052] For example, the click rate prediction network may include an input layer, a hidden layer and an output layer. The second feature sequence set is input into the input layer, and the hidden layer is used to extract high-order features from the input. The features are transformed and combined during the information transmission process to explore the potential relationship between the features. The last node of the output layer maps the output to the range of (0,1) through an activation function, which represents the click probability of the target user for each recalled object.

[0053] In this embodiment, a first feature sequence set is constructed according to the user features of the target user and the object features of the recalled object, forming a diversified feature space, which can more comprehensively describe the relationship between the user and the object, more accurately capture the user's interest points, and help the model to more comprehensively and deeply understand the user behavior pattern, and improve the accuracy of click-through rate estimation; the user features of the target user are combined with the object features of different recalled objects to construct a feature sequence set, forming batch features, which provides data support for the subsequent use of a normalized network, so that not only a layer normalized network can be used in the future, but also a batch normalized network can be used. The click-through rate prediction model includes at least two normalized networks and a feature fusion network, and at least two normalized networks are used to perform feature normalization processing on the first feature sequence set respectively. Through the mixed normalized network, the feature sequence is normalized from multiple angles, which enhances the diversity of feature expression and maximizes the retention of useful information of the feature, effectively reduces the instability and limitations that may be brought about by a single normalized network, and enhances the overall stability and adaptability of the model. With the help of the feature fusion network, the outputs of multiple normalized networks are optimized and combined, further improving the adaptability and performance of the model. The types of at least two normalized networks can be the same or different. When the types of at least two normalized networks are different, different types of normalized networks adapt to different data distributions, can more widely cover a variety of feature distributions, and improve the adaptability of the model; through multi-class normalization methods, different types of features can be processed more comprehensively and deeply, different types of normalization methods can complement each other, and the use of different types of normalized networks can process features from different angles, thereby improving the generalization ability of the model under multiple data distributions. When the types of at least two normalized networks are the same, since the normalized networks have learnable parameters, after training, the learnable parameters of at least two normalized networks will be different, and for the same input, the outputs of at least two normalized networks will also be different, achieving the effect of capturing different aspects and changes of the input data, and being able to effectively process dynamically changing input data; by processing features in parallel through at least two normalized networks of the same type, the consistency and stability of feature processing can be improved, and in the case of random noise and data anomalies, multiple normalized networks can complement and correct each other, and multiple normalization ensures that feature information is sufficient and complete, so that it can cope with a variety of data distributions and improve the expressiveness of the model. The click rate prediction network is used to estimate the user's click rate for each recalled object based on the fused feature sequence set, so that the click rate prediction network can understand user behavior and interests more comprehensively and accurately, improve the accuracy of click rate prediction, and enhance user satisfaction and usage experience, thereby improving the overall performance of the system and user retention rate.

[0054] The click rate estimation method provided by the embodiment of the present invention is further described below. Figure 2a As shown, the method of this embodiment includes:

[0055] Step 201: determine a recall object set for a target user, where the recall object set includes at least two recall objects.

[0056] The recall object may include but is not limited to advertisements, commodities, products, articles, and videos. The following takes the case where the recall object is an advertisement of a commodity as an example to introduce the application scenario of the click rate estimation method of the embodiment of the present invention. Figure 2b As shown in the figure, when users browse the recommendation positions on e-commerce websites, social media platforms, etc., they may see multiple advertisements. Users may be interested in certain advertisements, click on them, or even make further conversion behaviors such as purchases and registrations. When users browse advertisements, the advertising service system will obtain relevant features in real time based on the user's behavior and context. These features may include: advertising product features, context features... user behavior sequence features. Among them, advertising product features can be features related to the advertising product itself, such as category, price, etc. Context features can be features related to the current display environment or scene, such as time, geographic location, device type, etc. User behavior sequence features can be features related to the user's historical behavior, including user's click, browse, purchase and other behavior records over the past period of time. The advertising service system inputs the acquired features into the click-through rate prediction model, and the click-through rate prediction model outputs the probability of a user clicking on a specific advertisement, that is, the click-through rate (Predicted Click-Through Rate, pCTR). At the same time, the advertising service system will also use the conversion rate prediction model to predict the conversion probability after the user clicks on the advertisement and generate the conversion rate (Predicted Conversion Rate, pCVR). The smart bidding module comprehensively considers pCTR, pCVR and the advertiser's bid to calculate the comprehensive score or ranking score of the advertisement. Common calculation methods such as pCTR*bid are used to determine the priority and display order of the advertisement.

[0057] After a user clicks on an ad or completes subsequent conversion actions such as purchases and registrations, the system will generate click logs and order logs. These logs will record the user's clicks and conversion details and synchronize them to the downstream analysis and decision-making system. Based on the above log data, the advertising system will generate a delivery effect report and provide it to the merchant so that the merchant can view and analyze the effectiveness of the advertisement. Based on the results of clicks and conversions, the billing system will deduct the corresponding fees from the advertiser's account. This fee calculation can be based on a pricing model based on clicks (Cost Per Click, CPC) or conversions (Cost Per Acquisition, CPA). The billing module will record the cost of each ad display, click, and conversion, and ultimately calculate the total cost that the advertiser should pay.

[0058] It can be seen that the click rate prediction model needs to process a large number of features and needs to adopt a normalized network. However, using a single normalized network is difficult to cope with scenarios with multiple data distributions and dynamic changes, and the model performs poorly. Therefore, this application improves the structure of the click rate prediction model and its method of processing data.

[0059] Step 202: construct a first feature sequence set based on the user features of the target user and the object features of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user features of the target user and the object features of one of the at least two recalled objects.

[0060] The user characteristics of the target user include user basic characteristics, user behavior characteristics, and scene characteristics of the scene in which the target user is located when the user behavior characteristics are generated. Among them, user basic characteristics refer to characteristics related to the user's personal basic information, which are usually static or change less, and most of them are explicit data. For example, user basic characteristics may include the user's age, gender, occupation, geographic location, membership level, etc. User behavior characteristics are data representations of various types of interactive behaviors of users on the platform, which are usually dynamic and can directly reflect the user's interests and preferences as time changes. For example, user behavior characteristics may include browsing records, click records, purchase records, search records, collections, likes, comments, etc. The scene characteristics of the scene in which the target user is located when the user behavior characteristics are generated refer to the specific environment and conditions when the user performs specific behaviors, which may include the time, place and weather conditions when the user performs these behaviors, and the type, configuration, network environment, etc. of the device used by the user when performing these behaviors. Scene characteristics help to understand the user's behavior patterns in different situations and improve the prediction accuracy of the click-through rate prediction model. Object characteristics can be characteristics used in the recommendation system to describe the basic information or attributes of candidate objects. For example, the object characteristics may include identification information, category information, tag information, function information, summary information, brand information, etc. of the object.

[0061] Each feature sequence is a data set consisting of user features of a target user and object features of a recalled object. In a specific implementation, the user features of the target user and the object features of each recalled object can be stored separately. When needed, the user features of the target user and the object features of each recalled object can be retrieved, and the user features of the target user can be copied multiple times. Each user feature of the target user is matched with an object feature of a recalled object to form a series of feature sequences, which are the first feature sequence sets.

[0062] Step 203: input the first feature sequence set into a click rate prediction model.

[0063] like Figure 2cAs shown in FIG. 1 , the click rate prediction model may include N feature processing networks, a feature aggregation network, and a click rate prediction network, where N is an integer greater than 1, i.e., it includes at least two feature processing networks. Each feature processing network includes a feedforward neural network and a hybrid normalization network. The hybrid normalization network is shown in FIG. Figure 2d As shown, it includes at least two normalization networks, a gating network and a feature fusion network.

[0064] Step 204: Use a feedforward neural network to perform feature dimension reduction on the first feature sequence set and capture the relationship between the user features and the object features to obtain an enhanced feature sequence set.

[0065] A feedforward neural network may include an input layer, a hidden layer, and an output layer. The input layer of the feedforward neural network takes the first feature sequence set as input and uses several hidden layers. Each layer captures the relationship between user features and object features through a nonlinear activation function to extract more meaningful high-order interaction features. The output layer may be set with fewer neurons than the input layer to achieve feature dimensionality reduction, and the output feature vector is the enhanced feature sequence set.

[0066] Through the feedforward neural network, the mapping from high-dimensional feature space to low-dimensional feature space is achieved, which reduces feature redundancy and improves computational efficiency. The hidden layer in the feedforward neural network can capture the complex nonlinear relationship between the original features, making the generated enhanced features more representative and discriminative. This lays a data foundation for the subsequent click-through rate prediction task and helps improve the prediction ability of the overall system.

[0067] Step 205: Use at least two normalization networks to perform feature normalization processing on the enhanced feature sequence sets respectively to obtain at least two normalized feature sequence sets.

[0068] By using a normalization network to perform feature normalization on the enhanced feature sequence set, a normalized feature sequence set can be obtained. Assume that X represents the enhanced feature sequence set, a hybrid normalization network has k normalization networks, which are recorded as NormLayer1, NormLayer2,…, NormLayer k , then the output of these normalized networks can be expressed as:

[0069] h j =NormLayer j (X),j=1,2,…,k

[0070]

[0071] Among them, h j represents the output of the jth normalized network for input X, and is the mean and variance of the input data in a certain dimension, depending on the type of normalization network. If the normalization network is a batch normalization network, then and is the mean and variance of the feature dimension calculated on this batch of data; if the normalized network is a layer normalized network, then and is the mean and variance of the feature dimension computed over a single feature sequence. ∈ is a very small positive number set to avoid division by zero errors and numerical instability. j and β j is the learnable parameter of the jth normalized network, which is used to scale and translate the normalized value, i.e., γ j is the scaling parameter of the jth normalized network, β j is the translation parameter of the jth normalized network. For two normalized networks of the same or different types, the values ​​of the scaling parameters and translation parameters obtained after training are generally different.

[0072] Step 206, using a gating network to determine the network weight of each normalized network based on the enhanced feature sequence set.

[0073] The gated network is a specific type of neural network that can be pre-trained. During the training process, it will learn to dynamically adjust the importance of different normalization methods according to the type and statistical distribution of the input features. Therefore, it can dynamically generate weights for weighted fusion for each normalized network through the input features. That is, the gated network will analyze the input enhanced feature sequence set and dynamically generate a set of appropriate weights according to the type and statistical properties of the features. This set of weights is the network weight of this set of normalized networks. The gated network includes a fully connected layer and a mapping function layer. The fully connected layer can be used to generate initial weights for each normalized network based on the enhanced feature sequence set; the mapping function layer is used to map the initial weights of each normalized network to a preset value range to obtain the network weights of each normalized network. The preset range can be (0, 1). The mapping function layer can use the normalized exponential function Softmax. The Softmax function ensures that the output is non-negative and the sum is 1, so that these network weights can be interpreted as the importance of each normalized network.

[0074] For example, X is still used to represent the enhanced feature sequence set. A hybrid normalized network has k normalized networks. The process of using the gating network to determine the network weight of each normalized network based on the enhanced feature sequence set can be as follows:

[0075] g j =Gate(X) j

[0076] Gate(X) j =softmaxj (X·W+b)

[0077] Among them, g j represents the network weight of the jth normalized network, Gate represents the gated network, W and b are the weight and bias parameters of the gated network, which are learnable parameters and can be learned during the model training phase.

[0078] The gating network is used to assign weights to the normalized network. The gating network can generate different weight assignments according to the different attributes of the input features and dynamically adapt to feature changes. That is, the weight of each normalized network can be adjusted according to the actual situation of the current feature and is not fixed. Therefore, it can adapt to different input feature environments, making the model more flexible when dealing with complex and diverse input data.

[0079] Step 207, using a feature fusion network to fuse at least two normalized feature sequence sets according to the network weight of each normalized network in the at least two normalized networks, to obtain a fused feature sequence set of a single feature processing network.

[0080] Specifically, the network weight of each normalized network and the normalized feature sequence set obtained using the corresponding normalized network can be weighted and summed to obtain a fused feature sequence set of a single feature processing network.

[0081] For example, the process of fusing at least two normalized feature sequence sets may be as follows:

[0082]

[0083] Where n represents the input data batch size, y i It represents the fused feature sequence set of data with batch size i, that is, the feature sequence set obtained by normalizing and fusing the data with batch size i by a feature processing network. One feature processing network corresponds to one fused feature sequence set. If there are several feature processing networks, there will be several fused feature sequence sets.

[0084] By weighting the outputs of different normalized networks through weighted fusion, the accuracy and stability of feature representation are improved, thereby improving the performance of the click-through rate prediction model.

[0085] Step 208: A feature aggregation network is used to aggregate the fused feature sequence sets of at least two feature processing networks to obtain a second feature sequence set.

[0086] The aggregation processing here can be simple concatenation, averaging, weighted averaging or more complex neural network fusion methods. If weighted averaging is used, the network weights of different feature processing networks can be learned using the model, or they can be set randomly or manually. The output features of multiple feature processing networks are organically integrated to construct a comprehensive feature representation, which can improve the richness and discrimination of the feature representation and provide more expressive feature input for downstream tasks.

[0087] Step 209 : using the click rate prediction network to predict the click rate of each recalled object by the target user based on the second feature sequence set.

[0088] For example, assuming that there are 3 recall objects in the recall object set, the target user's estimated click-through rates for these three recall objects can be obtained here, for example: the click-through rate for recall object 1 is 0.8, the click-through rate for recall object 2 is 0.5, and the click-through rate for recall object 3 is 0.2.

[0089] In this embodiment, a first feature sequence set is constructed according to the user features of the target user and the object features of the recalled object, forming a diversified feature space, which can more comprehensively describe the relationship between the user and the object, more accurately capture the user's interest points, and help the model to more comprehensively and deeply understand the user behavior pattern, and improve the accuracy of click-through rate estimation; the user features of the target user are combined with the object features of different recalled objects to construct a feature sequence set, forming batch features, which provides data support for the subsequent use of a normalized network, so that not only a layer normalized network can be used in the future, but also a batch normalized network can be used. The click-through rate prediction model includes at least two normalized networks and a feature fusion network, and at least two normalized networks are used to perform feature normalization processing on the first feature sequence set respectively. Through the mixed normalized network, the feature sequence is normalized from multiple angles, which enhances the diversity of feature expression and maximizes the retention of useful information of the feature, effectively reduces the instability and limitations that may be brought about by a single normalized network, and enhances the overall stability and adaptability of the model. With the help of the feature fusion network, the outputs of multiple normalized networks are optimized and combined, further improving the adaptability and performance of the model. The types of at least two normalized networks can be the same or different. When the types of at least two normalized networks are different, different types of normalized networks adapt to different data distributions, can more widely cover a variety of feature distributions, and improve the adaptability of the model; through multi-class normalization methods, different types of features can be processed more comprehensively and deeply, different types of normalization methods can complement each other, and the use of different types of normalized networks can process features from different angles, thereby improving the generalization ability of the model under multiple data distributions. When the types of at least two normalized networks are the same, since the normalized networks have learnable parameters, after training, the learnable parameters of at least two normalized networks will be different, and for the same input, the outputs of at least two normalized networks will also be different, achieving the effect of capturing different aspects and changes of the input data, and being able to effectively process dynamically changing input data; by processing features in parallel through at least two normalized networks of the same type, the consistency and stability of feature processing can be improved, and in the case of random noise and data anomalies, multiple normalized networks can complement and correct each other, and multiple normalization ensures that feature information is sufficient and complete, so that it can cope with a variety of data distributions and improve the expressiveness of the model. The click rate prediction network is used to estimate the user's click rate for each recalled object based on the fused feature sequence set, so that the click rate prediction network can understand user behavior and interests more comprehensively and accurately, improve the accuracy of click rate prediction, and enhance user satisfaction and usage experience, thereby improving the overall performance of the system and user retention rate.

[0090] In addition, by using the gating network to assign weights to the normalized network, the gating network can generate different weight assignments according to the different attributes of the input features and dynamically adapt to feature changes. That is, the weight of each normalized network can be adjusted according to the actual situation of the current feature and is not fixed. Therefore, it can adapt to different input feature environments, making the model more flexible when dealing with complex and diverse input data.

[0091] The following introduces the training method of the click rate prediction model. Figure 3 As shown, the training method of the click rate prediction model provided in this embodiment may include the following steps:

[0092] Step 301: Obtain a training sample set.

[0093] Each sample in the training sample set has sample data and a sample label. The sample data of each sample is a feature sequence composed of a user feature of a sample user and an object feature of a sample object. The sample label of each sample is the actual click situation of the sample user in each sample on the corresponding sample object, that is, the present invention is a supervised training model. The actual click situation of the sample user in each sample on the corresponding sample object can be obtained based on the historical record data. The actual click situation includes click or no click. The actual click situation can be digitized for easy processing. For example, if the sample user clicks the corresponding sample object, the actual click situation is represented by 1. If the sample user does not click the corresponding sample object, the actual click situation is represented by 0.

[0094] In the training phase, the input of the model is a large amount of sample data, that is, a large number of feature sequences consisting of user features of different sample users and object features of different sample objects. That is, the model training phase involves a large number of users and a large number of sample objects, while the model inference phase involves the target user and at least two recall objects.

[0095] For example, the feature sequence in the training phase can be as shown in Table 2 below. Assuming that there are 3 sample objects and 2 sample users, the user features of the two sample users can be combined with the object features of the 3 sample objects, respectively, to obtain 6 feature sequences. Feature sequence 1 can be formed by combining the user features of sample user 1 and the object features of sample object 1, feature sequence 2 can be formed by combining the user features of sample user 1 and the object features of sample object 2, and feature sequence 3 can be formed by combining the user features of sample user 1 and the object features of sample object 3. The other feature sequences are similar and will not be described in detail.

[0096]

[0097]

[0098] Table 2

[0099] Step 302: Input the sample data of each sample into a preset model to estimate the click rate, and obtain the estimated click situation of the sample user in each sample on the corresponding sample object.

[0100] The preset model can be an initialized model for estimating click-through rate, and the structure of the preset model and the click-through rate estimation model obtained by final training can be the same. That is, the preset model also includes at least two feature processing networks, a feature aggregation network, and a click-through rate estimation network. Among them, each feature processing network includes a forward neural network and a hybrid normalization network. The hybrid normalization network includes at least two normalization networks, a gating network, and a feature fusion network.

[0101] The structures or types of the hybrid normalized networks included in different feature processing networks may be the same or different. The types of at least two normalized networks may be the same or different. For example, a hybrid normalized network included in one feature processing network includes three batch normalized networks, and a hybrid normalized network included in another feature processing network includes one batch normalized network and one layer normalized network. Alternatively, the hybrid normalized networks included in multiple feature processing networks are all composed of three batch normalized networks; or, the hybrid normalized networks included in multiple feature processing networks are all composed of three layer normalized networks.

[0102] The process of input data processing by the preset model can refer to the process of input data processing by the click rate prediction model described above, which will not be repeated here.

[0103] Step 303: determine the function value of the training loss function according to the expected click situation and the actual click situation of the sample user in each sample on the corresponding sample object.

[0104] Step 304: Optimize the model parameters of the preset model based on the function value to obtain a click rate prediction model.

[0105] The model parameters here may include scaling parameters and translation parameters of each normalized network, and may also include weight and bias parameters of the gating network. Training may be stopped when the function value is less than a preset value, or when the number of iterative training reaches a preset number.

[0106] In this embodiment, the preset model used to train the click-through rate prediction model includes at least two normalized networks, and the combination of these networks can flexibly adapt to different feature distributions to improve training stability and efficiency; combining user features and object features for model training can capture rich information of different types of features; the outputs of multiple feature processing networks are integrated through a feature aggregation network to improve the integrity and discrimination of feature representation; through feature fusion and aggregation processing, the risk of overfitting is reduced and the generalization ability of the model is improved; actual click situations are used as labels for supervised training, so that the model can effectively use historical data to learn user behavior patterns and improve training efficiency; the model structure of the training stage and the inference stage is consistent, ensuring that the trained model can be directly used in actual application scenarios, simplifying model deployment.

[0107] Figure 4 is a schematic diagram of a structure of a click rate prediction device provided by an embodiment of the present invention, and the device is suitable for executing a click rate prediction method provided by an embodiment of the present invention, such as Figure 4 As shown, the device may specifically include:

[0108] The object determination module 401 is used to determine a recall object set for a target user, wherein the recall object set includes at least two recall objects;

[0109] A feature construction module 402 is used to construct a first feature sequence set according to the user feature of the target user and the object feature of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user feature of the target user and the object feature of one of the at least two recalled objects;

[0110] An execution module 403 is used to input the first feature sequence set into a click-through rate prediction model, wherein the click-through rate prediction model includes a hybrid normalization network and a click-through rate prediction network, wherein the hybrid normalization network includes at least two normalization networks and a feature fusion network; using the at least two normalization networks to perform feature normalization processing on the first feature sequence set respectively to obtain at least two normalized feature sequence sets, using the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a second feature sequence set; using the click-through rate prediction network to estimate the click rate of the target user for each of the recalled objects based on the second feature sequence set.

[0111] In one embodiment, the execution module 403 uses the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a second feature sequence set, including:

[0112] The feature fusion network is used to perform fusion processing on the at least two normalized feature sequence sets according to the network weight of each normalized network in the at least two normalized networks to obtain the second feature sequence set.

[0113] In one embodiment, the hybrid normalization network further includes a gating network. After the first feature sequence set is input into the click rate prediction model, the execution module 403 is further used to:

[0114] The gating network is used to determine a network weight of each normalized network based on the first feature sequence set.

[0115] In one embodiment, the gating network includes a fully connected layer and a mapping function layer, and the execution module 403 uses the gating network to determine the network weight of each normalized network based on the first feature sequence set, including:

[0116] Using the fully connected layer to generate initial weights for each normalized network based on the first feature sequence set;

[0117] The mapping function layer is used to map the initial weight of each normalized network to a preset value range to obtain the network weight of each normalized network.

[0118] In one embodiment, the click rate prediction model further includes a feature aggregation network and at least two feature processing networks, each of the at least two feature processing networks includes the hybrid normalization network, and the execution module 403 uses the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a second feature sequence set, including:

[0119] Using the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a fused feature sequence set of a single feature processing network;

[0120] The feature aggregation network is used to aggregate the fused feature sequence sets of the at least two feature processing networks to obtain the second feature sequence set.

[0121] In one embodiment, each feature processing network further includes a feed-forward neural network. After the first feature sequence set is input into the click rate prediction model, the execution module 403 is further configured to:

[0122] Using the feedforward neural network to perform feature dimension reduction on the first feature sequence set and capture the relationship between user features and object features to obtain an enhanced feature sequence set;

[0123] The using the at least two normalization networks to respectively perform feature normalization processing on the first feature sequence set to obtain at least two normalized feature sequence sets includes:

[0124] The at least two normalization networks are used to perform feature normalization processing on the enhanced feature sequence sets respectively to obtain at least two normalized feature sequence sets.

[0125] In one embodiment, the user characteristics of the target user include basic user characteristics, user behavior characteristics, and scene characteristics of the scene in which the target user is located when the user behavior characteristics are generated.

[0126] In one embodiment, the click rate prediction model is trained by the following method:

[0127] Acquire a training sample set, wherein each sample in the training sample set has sample data and a sample label, wherein the sample data of each sample is a feature sequence consisting of a user feature of a sample user and an object feature of a sample object, and the sample label of each sample is an actual click situation of the sample user in each sample on the corresponding sample object;

[0128] Input the sample data of each sample into a preset model to estimate the click rate, and obtain the expected click situation of the sample user in each sample on the corresponding sample object;

[0129] Determine the function value of the training loss function according to the expected click situation and the actual click situation of the sample user in each sample on the corresponding sample object;

[0130] The model parameters of the preset model are optimized based on the function value to obtain the click rate prediction model.

[0131] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example for illustration. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the functional modules described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0132] The device of the embodiment of the present invention constructs a first feature sequence set according to the user features of the target user and the object features of the recalled object, forming a diversified feature space, which can more comprehensively describe the relationship between the user and the object, more accurately capture the user's interest points, and help the model to more comprehensively and deeply understand the user behavior pattern, and improve the accuracy of click-through rate estimation; the user features of the target user are combined with the object features of different recalled objects to construct a feature sequence set, forming batch features, which provides data support for the subsequent use of a normalized network, so that not only a layer normalized network but also a batch normalized network can be used in the future. The click-through rate prediction model includes at least two normalized networks and a feature fusion network, and at least two normalized networks are used to perform feature normalization processing on the first feature sequence set respectively. Through the mixed normalized network, the feature sequence is normalized from multiple angles, which enhances the diversity of feature expression and maximizes the retention of useful information of the feature, effectively reduces the instability and limitations that may be brought about by a single normalized network, and enhances the overall stability and adaptability of the model. With the help of the feature fusion network, the outputs of multiple normalized networks are optimized and combined, further improving the adaptability and performance of the model. The types of at least two normalized networks can be the same or different. When the types of at least two normalized networks are different, different types of normalized networks adapt to different data distributions, can more widely cover a variety of feature distributions, and improve the adaptability of the model; through multi-class normalization methods, different types of features can be processed more comprehensively and deeply, different types of normalization methods can complement each other, and the use of different types of normalized networks can process features from different angles, thereby improving the generalization ability of the model under multiple data distributions. When the types of at least two normalized networks are the same, since the normalized networks have learnable parameters, after training, the learnable parameters of at least two normalized networks will be different, and for the same input, the outputs of at least two normalized networks will also be different, achieving the effect of capturing different aspects and changes of the input data, and being able to effectively process dynamically changing input data; by processing features in parallel through at least two normalized networks of the same type, the consistency and stability of feature processing can be improved, and in the case of random noise and data anomalies, multiple normalized networks can complement and correct each other, and multiple normalization ensures that feature information is sufficient and complete, so that it can cope with a variety of data distributions and improve the expressiveness of the model. The click rate prediction network is used to estimate the user's click rate for each recalled object based on the fused feature sequence set, so that the click rate prediction network can understand user behavior and interests more comprehensively and accurately, improve the accuracy of click rate prediction, and enhance user satisfaction and usage experience, thereby improving the overall performance of the system and user retention rate.

[0133] An embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the click rate prediction method provided in any of the above embodiments is implemented.

[0134] An embodiment of the present invention further provides a computer-readable medium on which a computer program is stored. When the program is executed by a processor, the click rate prediction method provided by any of the above embodiments is implemented.

[0135] Reference below Figure 5 , which shows a schematic diagram of the structure of a computer system 500 of an electronic device suitable for implementing an embodiment of the present invention. Figure 5 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0136] like Figure 5 As shown, the computer system 500 includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage part 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the computer system 500 are also stored. The CPU 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0137] The following components are connected to the I / O interface 505: an input section 506 including a keyboard, a mouse, etc.; an output section 507 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, a modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I / O interface 505 as needed. A removable medium 511, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 510 as needed, so that a computer program read therefrom is installed into the storage section 508 as needed.

[0138] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 509, and / or installed from the removable medium 511. When the computer program is executed by the central processing unit (CPU) 501, the above-mentioned functions defined in the system of the present invention are executed.

[0139] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0140] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0141] The modules and / or units involved in the embodiments of the present invention may be implemented in software or hardware. The modules and / or units described may also be set in a processor, for example, it may be described as: a processor includes an object determination module, a feature construction module, and an execution module. The names of these modules do not, in some cases, constitute limitations on the modules themselves.

[0142] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes:

[0143] Determine a recall object set for a target user, wherein the recall object set includes at least two recall objects; construct a first feature sequence set based on the user features of the target user and the object features of each of the at least two recall objects, wherein each feature sequence in the first feature sequence set includes the user features of the target user and the object features of one of the at least two recall objects; input the first feature sequence set into a click-through rate prediction model, wherein the click-through rate prediction model includes a hybrid normalization network and a click-through rate prediction network, wherein the hybrid normalization network includes at least two normalization networks and a feature fusion network; perform feature normalization processing on the first feature sequence set respectively using the at least two normalized feature sequence sets, and perform fusion processing on the at least two normalized feature sequence sets using the feature fusion network to obtain a second feature sequence set; and use the click-through rate prediction network to estimate the click rate of the target user for each of the recall objects based on the second feature sequence set.

[0144] The technical solution of the embodiment of the present invention constructs a first feature sequence set according to the user features of the target user and the object features of the recalled object, forming a diversified feature space, which can more comprehensively describe the relationship between the user and the object, more accurately capture the user's interest points, and help the model to more comprehensively and deeply understand the user behavior pattern, and improve the accuracy of click-through rate estimation; the user features of the target user are combined with the object features of different recalled objects to construct a feature sequence set, forming batch features, which provides data support for the subsequent use of a normalized network, so that not only a layer normalized network but also a batch normalized network can be used in the future. The click-through rate prediction model includes at least two normalized networks and a feature fusion network, and at least two normalized networks are used to perform feature normalization processing on the first feature sequence set respectively. Through the mixed normalized network, the feature sequence is normalized from multiple angles, which enhances the diversity of feature expression and maximizes the retention of useful information of the feature, effectively reduces the instability and limitations that may be brought about by a single normalized network, and enhances the overall stability and adaptability of the model. With the help of the feature fusion network, the outputs of multiple normalized networks are optimized and combined, further improving the adaptability and performance of the model. The types of at least two normalized networks can be the same or different. When the types of at least two normalized networks are different, different types of normalized networks adapt to different data distributions, can more widely cover a variety of feature distributions, and improve the adaptability of the model; through multi-class normalization methods, different types of features can be processed more comprehensively and deeply, different types of normalization methods can complement each other, and the use of different types of normalized networks can process features from different angles, thereby improving the generalization ability of the model under multiple data distributions. When the types of at least two normalized networks are the same, since the normalized networks have learnable parameters, after training, the learnable parameters of at least two normalized networks will be different, and for the same input, the outputs of at least two normalized networks will also be different, achieving the effect of capturing different aspects and changes of the input data, and being able to effectively process dynamically changing input data; by processing features in parallel through at least two normalized networks of the same type, the consistency and stability of feature processing can be improved, and in the case of random noise and data anomalies, multiple normalized networks can complement and correct each other, and multiple normalization ensures that feature information is sufficient and complete, so that it can cope with a variety of data distributions and improve the expressiveness of the model. The click rate prediction network is used to estimate the user's click rate for each recalled object based on the fused feature sequence set, so that the click rate prediction network can understand user behavior and interests more comprehensively and accurately, improve the accuracy of click rate prediction, and enhance user satisfaction and usage experience, thereby improving the overall performance of the system and user retention rate.

[0145] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0146] It should be noted that the collection, collection, updating, analysis, processing, use, transmission, storage and other aspects of user personal information involved in the technical solution of this disclosure are in compliance with the provisions of relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain the security of user personal information, network security and national security.

[0147] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A click rate prediction method, characterized in that: include: Determine a recall object set for the target user, wherein the recall object set includes at least two recall objects; constructing a first feature sequence set according to the user feature of the target user and the object feature of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user feature of the target user and the object feature of one of the at least two recalled objects; Inputting the first feature sequence set into a click rate prediction model, wherein the click rate prediction model includes a hybrid normalization network and a click rate prediction network, wherein the hybrid normalization network includes at least two normalization networks and a feature fusion network; Using the at least two normalization networks to perform feature normalization processing on the first feature sequence set respectively to obtain at least two normalized feature sequence sets, and using the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a second feature sequence set; The click rate prediction network is used to predict the click rate of the target user for each of the recalled objects based on the second feature sequence set.

2. The click rate estimation method according to claim 1, characterized in that: The step of fusing the at least two normalized feature sequence sets using the feature fusion network to obtain a second feature sequence set includes: The feature fusion network is used to perform fusion processing on the at least two normalized feature sequence sets according to the network weight of each normalized network in the at least two normalized networks to obtain the second feature sequence set.

3. The click rate estimation method according to claim 2, characterized in that: The hybrid normalization network further includes a gating network, and after inputting the first feature sequence set into the click rate prediction model, further includes: The gating network is used to determine a network weight of each normalized network based on the first feature sequence set.

4. The click rate estimation method according to claim 3, characterized in that: The gating network includes a fully connected layer and a mapping function layer, and the determining the network weight of each normalized network based on the first feature sequence set by using the gating network includes: Using the fully connected layer to generate initial weights for each normalized network based on the first feature sequence set; The mapping function layer is used to map the initial weight of each normalized network to a preset value range to obtain the network weight of each normalized network.

5. The click rate prediction method according to claim 1, characterized in that: The click rate prediction model further includes a feature aggregation network and at least two feature processing networks, each of the at least two feature processing networks includes the hybrid normalization network, and the at least two normalized feature sequence sets are fused using the feature fusion network to obtain a second feature sequence set, including: Using the feature fusion network to perform fusion processing on the at least two normalized feature sequence sets to obtain a fused feature sequence set of a single feature processing network; The feature aggregation network is used to aggregate the fused feature sequence sets of the at least two feature processing networks to obtain the second feature sequence set.

6. The click rate prediction method according to claim 5, characterized in that: Each of the feature processing networks further includes a feed-forward neural network, which, after inputting the first feature sequence set into the click rate prediction model, further includes: Using the feedforward neural network to perform feature dimension reduction on the first feature sequence set and capture the relationship between user features and object features to obtain an enhanced feature sequence set; The using the at least two normalization networks to respectively perform feature normalization processing on the first feature sequence set to obtain at least two normalized feature sequence sets includes: The at least two normalization networks are used to perform feature normalization processing on the enhanced feature sequence sets respectively to obtain at least two normalized feature sequence sets.

7. The click rate estimation method according to any one of claims 1 to 6, characterized in that: The user characteristics of the target user include user basic characteristics, user behavior characteristics, and scene characteristics of the scene in which the target user is located when the user behavior characteristics are generated.

8. The click rate estimation method according to any one of claims 1 to 6, characterized in that: The click rate prediction model is trained by the following method: Acquire a training sample set, wherein each sample in the training sample set has sample data and a sample label, wherein the sample data of each sample is a feature sequence consisting of a user feature of a sample user and an object feature of a sample object, and the sample label of each sample is an actual click situation of the sample user in each sample on the corresponding sample object; Input the sample data of each sample into a preset model to estimate the click rate, and obtain the expected click situation of the sample user in each sample on the corresponding sample object; Determine the function value of the training loss function according to the expected click situation and the actual click situation of the sample user in each sample on the corresponding sample object; The model parameters of the preset model are optimized based on the function value to obtain the click rate prediction model.

9. A click rate estimation device, characterized in that: include: An object determination module, used to determine a recall object set for a target user, wherein the recall object set includes at least two recall objects; a feature construction module, configured to construct a first feature sequence set according to the user feature of the target user and the object feature of each of the at least two recalled objects, wherein each feature sequence in the first feature sequence set includes the user feature of the target user and the object feature of one of the at least two recalled objects; An execution module, configured to input the first feature sequence set into a click rate prediction model, wherein the click rate prediction model includes a hybrid normalization network and a click rate prediction network, wherein the hybrid normalization network includes at least two normalization networks and a feature fusion network; The at least two normalization networks are used to perform feature normalization processing on the first feature sequence set respectively to obtain at least two normalized feature sequence sets; the feature fusion network is used to fuse the at least two normalized feature sequence sets to obtain a second feature sequence set; and the click rate prediction network is used to estimate the click rate of the target user for each recalled object based on the second feature sequence set.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the click rate prediction method according to any one of claims 1 to 8 is implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the click rate prediction method according to any one of claims 1 to 8 is implemented.

Images