Method and device for training a click rate prediction model

Abstract

The present disclosure provides a method and device for training a click rate prediction model, and relates to the technical field of data processing. The click rate prediction model comprises a super network module and a prediction module. The specific implementation scheme is as follows: obtaining sample feature information and label values; the sample feature information comprises sample user feature information and target object feature information; the label values are used to indicate whether the sample user has an interaction behavior with the target object; processing the target object feature information based on the super network module to obtain a plurality of adjacency matrices for feature interaction; obtaining a click rate prediction value of the sample user for the target object by using the prediction module according to the sample feature information and the plurality of adjacency matrices; and training the click rate prediction model according to the label values and the click rate prediction value. The present disclosure can quickly improve the click rate prediction accuracy of the click rate prediction model under the condition of small samples, so as to alleviate the cold start problem in the click rate prediction task to a certain extent.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and more particularly to the fields of artificial intelligence, NLP, deep learning and intelligent search, and especially to a training method and apparatus for a click-through rate prediction model. Background Technology

[0002] Click-through rate (CTR) prediction is a task that involves predicting the probability of a user clicking on an item based on user and item characteristics. Accurate CTR prediction enables the recommendation of the most valuable items to users, improving user experience. However, when the amount of item data available for training is very limited, such as in the early stages of a new item's introduction, traditional CTR prediction models perform poorly in predicting CTR, making it difficult to accurately recommend items to suitable users based on the prediction results—a phenomenon known as the cold start problem. Therefore, finding a way to quickly improve the accuracy of CTR prediction using limited item data to recommend items to the appropriate users has become a pressing need. Summary of the Invention

[0003] This disclosure provides a method and apparatus for training a click-through rate prediction model.

[0004] According to a first aspect of this disclosure, a method for training a click-through rate (CTR) prediction model is provided, the CTR prediction model including a hypernetwork module and a prediction module, the training method comprising:

[0005] Acquire sample feature information and label values; the sample feature information includes sample user feature information and target object feature information; the label values ​​are used to indicate whether the sample user has interactive behavior with the target object.

[0006] The target object feature information is processed based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction;

[0007] Based on the sample feature information and the multiple adjacency matrices, the prediction module is used to obtain the predicted click-through rate of the sample users for the target object;

[0008] The click-through rate prediction model is trained based on the label value and the predicted click-through rate value.

[0009] According to a second aspect of this disclosure, a click-through rate prediction method is provided, comprising:

[0010] Obtain user and target object feature information;

[0011] The user's feature information and the target object's feature information are input into a pre-trained click-through rate prediction model; wherein the click-through rate prediction model is trained using the method described in the first aspect above;

[0012] Obtain the click-through rate prediction value output by the click-through rate prediction model, and determine the click-through rate prediction value as the probability that the user interacts with the target object.

[0013] According to a third aspect of this disclosure, a training apparatus for a click-through rate (CTR) prediction model is provided, wherein the CTR prediction model includes a hypernetwork module and a prediction module, and the training apparatus includes:

[0014] The first acquisition module is used to acquire sample feature information and label values; the sample feature information includes sample user feature information and target object feature information; the label values ​​are used to indicate whether the sample user has interactive behavior with the target object.

[0015] The second acquisition module is used to process the feature information of the target object based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction.

[0016] The click-through rate prediction module is used to obtain the predicted click-through rate value of the sample user for the target object based on the sample feature information and the multiple adjacency matrices.

[0017] The training module is used to train the click-through rate prediction model based on the label value and the click-through rate prediction value.

[0018] According to a fourth aspect of this disclosure, a click-through rate prediction device is provided, comprising:

[0019] The first acquisition module is used to acquire the user's feature information and the target object's feature information;

[0020] The second acquisition module is used to input the user's feature information and the target object's feature information into a pre-trained click-through rate prediction model, acquire the click-through rate prediction value output by the click-through rate prediction model, and determine the click-through rate prediction value as the probability that the user interacts with the target object; wherein the click-through rate prediction model is trained using the method described in the first aspect above.

[0021] According to a fifth aspect of this disclosure, an electronic device is provided, comprising:

[0022] At least one processor; and

[0023] A memory communicatively connected to the at least one processor; wherein,

[0024] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above, or to perform the method described in the second aspect above.

[0025] According to a sixth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect above, or to perform the method described in the second aspect above.

[0026] According to a seventh aspect of this disclosure, a computer program product is provided, comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the method described in the first aspect above, or implements the steps of the method described in the second aspect above.

[0027] According to the training method of the click-through rate prediction model disclosed herein, an adjacency matrix for feature interaction of the target object can be generated based on the hypernetwork module. Click-through rate can be predicted based on the adjacency matrix and the click-through rate prediction model can be trained. This can quickly improve the click-through rate prediction accuracy of the click-through rate prediction model under small sample conditions, thereby alleviating the cold start problem in the click-through rate prediction task to a certain extent.

[0028] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0029] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0030] Figure 1 This is a flowchart illustrating a training method for a click-through rate prediction model according to an embodiment of this disclosure.

[0031] Figure 2 This is a flowchart illustrating a training method for another click-through rate prediction model proposed according to an embodiment of this disclosure;

[0032] Figure 3 This is a schematic diagram of a click-through rate prediction model proposed in an embodiment of this disclosure;

[0033] Figure 4 This is a flowchart illustrating a training method for another click-through rate prediction model proposed according to an embodiment of this disclosure;

[0034] Figure 5 This is a flowchart illustrating a click-through rate prediction method according to an embodiment of this disclosure;

[0035] Figure 6 This is a structural block diagram of a training device for a click-through rate prediction model provided in an embodiment of this disclosure;

[0036] Figure 7This is a structural block diagram of a training device for another click-through rate prediction model provided in this embodiment of the present disclosure;

[0037] Figure 8 This is a structural block diagram of a click-through rate prediction device provided in an embodiment of this disclosure;

[0038] Figure 9 This is a block diagram of an electronic device used to implement the click-through rate prediction model training method or click-through rate prediction method of the embodiments of this disclosure. Detailed Implementation

[0039] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0040] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. The singular forms “a” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0041] It should be understood that although the terms first, second, third, etc., may be used to describe various information in embodiments of this disclosure, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of embodiments of this disclosure, and similarly, second information may also be referred to as first information. Depending on the context, the words “if” and “suppose” as used herein may be interpreted as “when”, “when”, or “in response to a determination”.

[0042] It should be noted that the collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0043] In click-through rate (CTR) prediction tasks, finding suitable feature interactions often requires a large amount of data (feature interactions can fuse information from different features to generate new information, thereby improving CTR prediction accuracy). However, for newly appearing items or when only a small amount of item sample data is available, it is difficult to quickly perform feature interactions in CTR prediction tasks, resulting in poor CTR prediction performance. Therefore, this disclosure provides a method and apparatus for training a CTR prediction model. Specifically, the method and apparatus for training a CTR prediction model according to embodiments of this disclosure are described below with reference to the accompanying drawings.

[0044] Figure 1 This is a flowchart illustrating a training method for a click-through rate (CTR) prediction model according to an embodiment of this disclosure. The CTR prediction model includes a hypernetwork module and a prediction module. It should be noted that the training method for the CTR prediction model of this disclosure can be applied to the training device for the CTR prediction model of this disclosure, which can be configured on an electronic device. Furthermore, the CTR prediction model can be applied to recommendation, search, and other related scenarios to provide users with personalized image and text recommendations, personalized media asset recommendations, and other services. Figure 1 As shown, the training method for this click-through rate prediction model may include, but is not limited to, the following steps.

[0045] Step 101: Obtain sample feature information and label values. This sample feature information includes sample user feature information and target object feature information. The label values ​​indicate whether the sample user interacts with the target object.

[0046] The sample user characteristics information may include user ID, user age, user interests, etc. The target object can be an item, a product, or an advertisement, etc. Target object characteristics may include target object ID, target object type, target object attributes, etc. Tag values. This is used to indicate whether a sample user interacts with the target object. Interactions can include actions such as clicking, purchasing, sharing, or adding to favorites.

[0047] In some embodiments of this disclosure, the sample feature information can be sample feature information obtained by transforming discrete original feature information into dense space through feature embedding processing. in, N represents the feature information of the m-th sample. v N represents the number of feature information of the target object. u This represents the number of user feature information samples.

[0048] Step 102: Process the feature information of the target object based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction.

[0049] It should be noted that the adjacency matrix represents the feature graph, used to characterize feature interaction patterns, where nodes represent features and edges represent the correlations between features. In one possible implementation, the hypernetwork module can generate multiple adjacency matrices for feature interactions through multiple graph generators. These adjacency matrices represent feature interaction patterns specific to the target object, allowing for rapid optimization of feature interactions based on a small amount of sample data, even under conditions of limited sample size.

[0050] Step 103: Based on the sample feature information and multiple adjacency matrices, the prediction module is used to obtain the predicted click-through rate of the sample users for the target object.

[0051] As one possible implementation, sample feature information can be determined by identifying neighbor features based on the adjacency matrix, and then fused with neighbor features to perform feature interaction and obtain higher-order feature information. Subsequently, based on this higher-order feature information, a prediction module is used to predict the click-through rate, obtaining the predicted click-through rate value of the sample users for the target object.

[0052] Step 104: Train the click-through rate prediction model based on the tag value and the predicted click-through rate value.

[0053] Optionally, in some embodiments of this disclosure, a loss function can be calculated based on the label value and the predicted click-through rate (CTR) value. The model parameters of the hypernetwork module and the prediction module are then adjusted based on the loss function. As an example, meta-learning can be used to train the CTR prediction model. The gradient of the loss function is passed to the meta-learner for updating, completing the meta-learning training process. The Adam optimizer is used to optimize the model parameters, and the optimization objective can be as follows:

[0054]

[0055]

[0056] in, For the model parameters of the prediction module, L represents the model parameters of the hypernetwork module. meta The loss function for meta-learning, For the loss function specific to each task, φ iThis represents the target object ID embedding specific to target object i and the adjacency matrix specific to target object i. Optionally, the loss function can be cross-entropy. Thus, when the target object first appears, the hypernetwork module can generate the corresponding adjacency matrix for feature interaction. As the amount of target object data available for training gradually increases, the adjacency matrix will be gradually optimized through gradient descent, eventually generating the optimal feature interaction pattern specific to the target object. Therefore, even with small sample conditions, target object feature information and user feature information can quickly find suitable features to interact based on this feature interaction pattern.

[0057] By implementing the embodiments of this disclosure, an adjacency matrix for feature interaction of a target object can be generated based on a hypernetwork module. This matrix represents the target object's unique feature interaction pattern. Click-through rate (CTR) is predicted based on the adjacency matrix, and the CTR prediction model is trained. This can rapidly improve the CTR prediction accuracy under small sample conditions, thus alleviating the cold start problem in CTR prediction tasks to some extent. Effective recommendations can then be provided to users based on the predicted CTR, improving the user experience.

[0058] Figure 2 This is a flowchart illustrating a training method for another click-through rate prediction model proposed according to an embodiment of this disclosure. Wherein, as... Figure 3 As shown, the hypernetwork module in this click-through rate (CTR) prediction model includes multiple one-hot encoding units and a graph generator corresponding to each one-hot encoding unit. The prediction module in this CTR prediction model includes multiple graph neural network (GNN) units and prediction units. The number of GNN units is the same as the number of adjacency matrices, and each GNN unit corresponds to one adjacency matrix. Figure 2 As shown, the training method for this click-through rate prediction model may include, but is not limited to, the following steps.

[0059] Step 201: Obtain sample feature information and label values. This sample feature information includes sample user feature information and target object feature information. The label values ​​indicate whether the sample user interacts with the target object.

[0060] Step 202: Input the target object feature information into multiple one-hot coding units respectively to obtain the one-hot coded features output by each one-hot coding unit.

[0061] like Figure 3 As shown, v i u represents the original target object feature information of target object i. jThe original sample user feature information of sample user j is represented by the original target object feature information and the original sample user feature information. Feature embedding processing is performed on the original target object feature information and the original sample user feature information to obtain the sample user feature information and the target object feature information. It should be noted that, in the embodiments of this disclosure, a graph generator is defined for each sample feature information, i.e. Figure 3 n = N v +N u , where is the total number of sample feature information. The target object feature information is input into n one-hot coding units respectively, and the one-hot coded features output by each one-hot coding unit are obtained.

[0062] Sample feature information can be obtained by transforming discrete original feature information into dense space through feature embedding processing. in, N represents the feature information of the m-th sample. v N represents the number of feature information of the target object. u This represents the number of user feature information samples.

[0063] Step 203: Input the one-hot encoded features output by each one-hot encoded unit into the corresponding graph generator to obtain the output results of each graph generator.

[0064] It should be noted that the output of the graph generator m is the first adjacency matrix. A row of feature m, i.e., the adjacency matrix The m-th row, m∈[1,2,...,N] v +N u Optionally, the formula for obtaining the output of the MLP graph generator m can be found as follows:

[0065]

[0066] Step 204: Concatenate the outputs of each graph generator to obtain a first adjacency matrix for feature interaction.

[0067] In other words, the output results of n graph generators After concatenation, a unique (N) feature interaction property is obtained, specifically for target object i. v +N u )×(N v +N u The first adjacency matrix of )

[0068] Step 205: Iterate through the first adjacency matrix to determine multiple adjacency matrices from the first adjacency matrix and the adjacency matrix obtained after iteration.

[0069] It should be noted that to model high-order feature interactions, multiple graph neural network (Graph Neural Network) units are required, each of which needs to correspond to an adjacency matrix. Therefore, one implementation provides an iterative formula for the adjacency matrix, which can generate the adjacency matrix corresponding to any layer of the Graph Neural Network unit based on the first adjacency matrix. The design of the adjacency matrix iterative formula can refer to the following process:

[0070] First, for the adjacency matrix Make corrections:

[0071]

[0072] Here, `normalize(·)` performs min-max normalization to ensure that the elements in the adjacency matrix are between 0 and 1. Furthermore, to avoid negative impacts on prediction results from low-correlation feature interactions, `sparsify(·,K)` is used to retain the K largest elements in the adjacency matrix, setting the rest to zero. Since feature interactions satisfy the commutative law, the symmetry of the adjacency matrix can be guaranteed as follows:

[0073]

[0074] Based on this, the iterative formula is proposed as follows:

[0075]

[0076] by Figure 3 Taking the illustrated embodiment as an example, based on the first adjacency matrix The adjacency matrix corresponding to unit 2 of the graph neural network can be obtained through the iterative formula.

[0077] Step 206: Based on the sample feature information and multiple adjacency matrices, multiple graph neural network units are used to obtain the high-order feature information output by each graph neural network unit.

[0078] In some embodiments of this disclosure, multiple graph neural network units can be used to perform feature interaction based on sample feature information and multiple adjacency matrices, update the sample feature information state, and obtain the high-order feature information output by each graph neural network unit.

[0079] As an example, the first graph neural network unit can be used, based on the corresponding first adjacency matrix. Determine the neighbor features of the sample feature information, and fuse the sample feature information and its neighbor features to obtain the high-order feature information output by the first graph neural network unit.

[0080] The second graph neural network unit is used, based on the corresponding adjacency matrix. Determine neighbor features of the sample feature information, and combine the high-order feature information output by the first graph neural network unit with the neighbor features to obtain the high-order feature information output by the second graph neural network unit By that analogy, the l-th graph neural network unit is used to determine the neighbor features of the sample feature information based on the corresponding adjacency matrix, and the high-order feature information output by the (l-1)-th graph neural network unit is fused with the neighbor features to obtain the high-order feature information output by the l-th graph neural network unit where 1<l≤N, and N is the number of graph neural network units. The high-order feature information output by all graph neural network units is obtained until all are obtained. It should be noted that, Figure 3 the illustrated embodiment only takes two graph neural network units as an example, which does not constitute a limitation on the number of graph neural network units in the present disclosure.

[0081] In some embodiments of the present disclosure, the high-order feature information output by the l-th graph neural network unit can refer to the following formula:

[0082]

[0083] wherein, is a learnable parameter.

[0084] Step 207, according to the high-order feature information output by each graph neural network unit, a prediction unit is used to obtain a click-through rate prediction value of the sample user for the target object.

[0085] In some embodiments of the present disclosure, the high-order feature information output by each graph neural network unit can be fused, and based on the final representation of the fused sample feature information, the prediction unit is used to obtain the click-through rate prediction value of the sample user for the target object.

[0086] Step 208, train the click-through rate prediction model according to the label values and the click-through rate prediction values.

[0087] In the embodiments of the present disclosure, step 201 and step 208 may be implemented in any one of the manners in the respective embodiments of the present disclosure respectively, which is not specifically limited by the present disclosure and will not be repeated herein.

[0088] By implementing the embodiments of this disclosure, an adjacency matrix for feature interaction of a target object can be generated based on a hypernetwork module. Multiple graph neural network units are then used to perform high-order feature interaction based on the adjacency matrix. Click-through rate (CTR) is predicted based on this high-order feature information, and the CTR prediction model is trained. This can rapidly improve the CTR prediction accuracy of the model under small sample conditions, further enhancing the accuracy of predicting user click-through rates on target objects, thus alleviating the cold-start problem in CTR prediction tasks to some extent. Effective recommendations can then be provided to users based on the predicted CTR, improving user experience.

[0089] Figure 4 This is a flowchart illustrating a training method for another click-through rate (CTR) prediction model proposed according to an embodiment of this disclosure. The prediction module in this CTR prediction model includes multiple graph neural network (GNN) units, a feature fusion unit, and a prediction unit. The number of GNN units is the same as the number of adjacency matrices, and each GNN unit corresponds to one adjacency matrix. For example... Figure 4 As shown, the training method for this click-through rate prediction model may include, but is not limited to, the following steps.

[0090] Step 401: Obtain sample feature information and label values. This sample feature information includes sample user feature information and target object feature information. The label values ​​indicate whether the sample user interacts with the target object.

[0091] Step 402: Process the feature information of the target object based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction.

[0092] Step 403: Based on the sample feature information and multiple adjacency matrices, multiple graph neural network units are used to obtain the high-order feature information output by each graph neural network unit.

[0093] Step 404: Based on the feature fusion unit, an attention mechanism is used to fuse the high-order feature information output by each graph neural network unit to obtain the feature representation of each feature in the sample feature information.

[0094] Alternatively, the high-order feature information output by each graph neural network unit can be processed using an attention mechanism via the following formula. Perform fusion processing to obtain the feature representation of the m-th feature among the sample feature information.

[0095]

[0096] head h =attention(W q,h H m W k,h H m Wv,h H m )

[0097]

[0098] in N l This represents the number of units in the graph neural network.

[0099] Step 405: Input the feature representations of each feature in the sample feature information into the prediction unit to obtain the predicted click-through rate of the sample users for the target object.

[0100] In one possible implementation, the prediction unit can utilize the parameter W. c,1 First multilayer perceptron Evaluate the weights of the feature representations of each feature in the sample feature information:

[0101]

[0102] Then, by using parameter W c,2 Second multilayer perceptron The feature representation of each feature in the sample feature information is scored and weighted and summed to obtain the predicted click-through rate of the sample users for the target object.

[0103]

[0104] Step 406: Train the click-through rate prediction model based on the tag value and the predicted click-through rate value.

[0105] In the embodiments of this disclosure, steps 401-403 and step 406 can be implemented in any of the embodiments of this disclosure. This disclosure does not specifically limit them and will not elaborate further.

[0106] By implementing the embodiments of this disclosure, an adjacency matrix for feature interaction of a target object can be generated based on a hypernetwork module. This matrix represents the target object's unique feature interaction pattern. Multiple graph neural network units are then used to perform high-order feature interaction based on the adjacency matrix. An attention mechanism is employed to fuse the high-order feature information output by these multiple graph neural network units, thereby predicting the click-through rate (CTR) and training the CTR prediction model. This approach can rapidly improve the CTR prediction accuracy under small sample conditions, further enhancing the accuracy of predicting user click-through rates for the target object, thus alleviating the cold-start problem in CTR prediction tasks to some extent. Furthermore, effective recommendations can be provided to users based on the predicted CTR, improving the user experience.

[0107] This disclosure also proposes a click-through rate prediction method. Figure 5This is a flowchart illustrating a click-through rate (CTR) prediction method according to an embodiment of this disclosure. It should be noted that the CTR prediction method of this disclosure can be applied to the CTR prediction device of this disclosure, which can be configured on an electronic device. Figure 5 As shown, the click-through rate prediction method may include, but is not limited to, the following steps.

[0108] Step 501: Obtain the user's feature information and the target object's feature information.

[0109] Step 502: Input the user's feature information and the target object's feature information into the pre-trained click-through rate prediction model.

[0110] The click-through rate prediction model was trained using the training method of the click-through rate prediction model in any of the aforementioned embodiments, and will not be described in detail here.

[0111] Step 503: Obtain the click-through rate prediction value output by the click-through rate prediction model, and determine the click-through rate prediction value as the probability of user interaction with the target object.

[0112] By implementing the embodiments of this disclosure, the optimal feature interaction pattern specific to the target object can be obtained through a pre-trained click-through rate prediction model. Based on this feature interaction pattern, the click-through rate can be predicted, which can improve the accuracy of the click-through rate prediction value under small sample conditions. Based on the click-through rate prediction value, effective recommendations can be provided to users, thereby improving the user experience.

[0113] Figure 6 This is a structural block diagram of a training device for a click-through rate (CTR) prediction model provided in an embodiment of this disclosure. The CTR prediction model includes a hypernetwork module and a prediction module. Figure 6 As shown, the training device for the click-through rate prediction model may include: a first acquisition module 601, a second acquisition module 602, a click-through rate prediction module 603, and a training module 604.

[0114] The first acquisition module 601 is used to acquire sample feature information and label values; the sample feature information includes sample user feature information and target object feature information; the label values ​​are used to indicate whether the sample user has interactive behavior with the target object.

[0115] The second acquisition module 602 is used to process the feature information of the target object based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction.

[0116] Click-through rate prediction module 603 is used to obtain the predicted click-through rate value of the sample user for the target object based on the sample feature information and the plurality of adjacency matrices using the prediction module;

[0117] The training module 604 is used to train the click-through rate prediction model based on the label value and the click-through rate prediction value.

[0118] In some embodiments of this disclosure, the hypernetwork module includes multiple one-hot coding units and a graph generator corresponding to each one-hot coding unit; the second acquisition module 602 is specifically used for: inputting the target object feature information into the multiple one-hot coding units respectively to obtain the one-hot coding features output by each one-hot coding unit; inputting the one-hot coding features output by each one-hot coding unit into the corresponding graph generator to obtain the output result of each graph generator; concatenating the output results of each graph generator to obtain a first adjacency matrix for feature interaction; iterating over the first adjacency matrix, and determining the first adjacency matrix and the adjacency matrix obtained after iteration as the multiple adjacency matrices.

[0119] In some embodiments of this disclosure, the training module 604 is specifically used to: calculate a loss function based on the label value and the predicted click-through rate value; and adjust the model parameters of the hypernetwork module and the model parameters of the prediction module based on the loss function.

[0120] Optionally, in some embodiments of this disclosure, the prediction module includes multiple graph neural network units and a prediction unit, wherein the number of graph neural network units is the same as the number of adjacency matrices, and each graph neural network unit corresponds to one of the multiple adjacency matrices; for example Figure 7 As shown, the click-through rate prediction module 703 includes an acquisition unit 705 and a prediction unit 706. The acquisition unit 705 is used to acquire high-order feature information output by each graph neural network unit using the plurality of graph neural network units, based on the sample feature information and the plurality of adjacency matrices. The prediction unit 706 is used to acquire the predicted click-through rate value of the sample user for the target object using the prediction unit, based on the high-order feature information output by each graph neural network unit. Figure 7 701-704 and Figure 6 The 601-604 series have the same function and structure.

[0121] In some embodiments of this disclosure, the prediction module further includes a feature fusion unit; the prediction unit 706 is specifically used to: based on the feature fusion unit, use an attention mechanism to fuse the high-order feature information output by each graph neural network unit to obtain the feature representation of each feature in the sample feature information; input the feature representation of each feature in the sample feature information to the prediction unit to obtain the predicted value of the click rate of the sample user on the target object.

[0122] In some embodiments of this disclosure, the acquisition unit 705 is specifically used to: employ the first graph neural network unit to determine the neighbor features of the sample feature information based on the corresponding adjacency matrix, and fuse the sample feature information and its neighbor features to obtain the high-order feature information output by the first graph neural network unit; employ the l-th graph neural network unit to determine the neighbor features of the sample feature information based on the corresponding adjacency matrix, and fuse the high-order feature information output by the (l-1)-th graph neural network unit with the neighbor features to obtain the high-order feature information output by the l-th graph neural network unit; wherein, 1 <l≤N l N l The number of graph neural network units; until the high-order feature information output by all graph neural network units is obtained.

[0123] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0124] Figure 8 This is a structural block diagram of a click-through rate prediction device provided in an embodiment of this disclosure. Figure 8 As shown, the click-through rate prediction device may include: a first acquisition module 801 and a second acquisition module 802.

[0125] The first acquisition module 801 is used to acquire the user's feature information and the target object's feature information.

[0126] The second acquisition module 802 is used to input the feature information of the user and the feature information of the target object into a pre-trained click-through rate prediction model, acquire the click-through rate prediction value output by the click-through rate prediction model, and determine the click-through rate prediction value as the probability of the user interacting with the target object; wherein, the click-through rate prediction model is trained using the training method of the click-through rate prediction model of any of the foregoing embodiments, and will not be described in detail here.

[0127] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0128] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0129] like Figure 9 As shown, Figure 9This is a block diagram of an electronic device used to implement a training method or click-through rate prediction method for a click-through rate prediction model according to embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0130] like Figure 9 As shown, the electronic device includes one or more processors 901, a memory 902, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components are interconnected via different buses and can be mounted on a common motherboard or otherwise as required. The processors can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take the 901 processor as an example.

[0131] The memory 902 is the non-transitory computer-readable storage medium provided in this disclosure. The memory stores instructions executable by at least one processor to cause the at least one processor to perform the training method or click-through rate prediction method of the click-through rate prediction model provided in this disclosure. The non-transitory computer-readable storage medium of this disclosure stores computer instructions for causing a computer to perform the training method or click-through rate prediction method of the click-through rate prediction model provided in this disclosure.

[0132] The memory 902, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the training method for the click-through rate prediction model or the program instructions / modules corresponding to the click-through rate prediction method in the embodiments of this disclosure. The processor 901 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions, and modules stored in the memory 902, thereby implementing the training method for the click-through rate prediction model or the click-through rate prediction method in the above method embodiments.

[0133] The memory 902 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function. The data storage area may store data created by using the click-through rate prediction model training method or the electronic device implementing the click-through rate prediction method. Furthermore, the memory 902 may include high-speed random access memory and non-transient memory, such as at least one disk storage device, flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory 902 may optionally include memory remotely located relative to the processor 901. This remote memory can be connected via a network to the electronic device used to implement the click-through rate prediction model training method or the click-through rate prediction method. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0134] The training method for the click-through rate (CTR) prediction model or the electronic device for the CTR prediction method may further include: an input device 903 and an output device 904. The processor 901, memory 902, input device 903, and output device 904 can be connected via a bus or other means. Figure 9 Taking the example of a connection between China and Israel via a bus.

[0135] Input device 903 can receive input numerical or character information, as well as key signal inputs related to user settings and function control of the electronic device used for the click-through rate prediction model training method or click-through rate prediction method. Examples of input devices include touchscreens, keypads, mice, trackpads, touchpads, joysticks, one or more mouse buttons, trackballs, and joysticks. Output device 904 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The display device may include, but is not limited to, liquid crystal displays (LCDs), light-emitting diode (LED) displays, and plasma displays. In some embodiments, the display device may be a touchscreen.

[0136] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs, i.e., this disclosure also proposes a computer program that, when executed by a processor, implements the training method or click-through rate prediction method of the click-through rate prediction model described in the above embodiments. This one or more computer programs can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0137] These computational programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0138] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0139] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

[0140] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, a host product within the cloud computing service system, addressing the shortcomings of traditional physical hosts and VPS services ("Virtual Private Server," or simply "VPS") in terms of management difficulty and weak business scalability. The server can also be a server for a distributed system or a server integrated with blockchain. It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure is achieved, and this is not limited herein.

[0141] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for a click-through rate (CTR) prediction model, wherein, The click-through rate prediction model includes a hypernetwork module and a prediction module, and the training method includes: The sample feature information and tag values ​​are obtained; the sample feature information includes sample user feature information and target object feature information; the tag values ​​are used to indicate whether the sample user has interactive behavior with the target object, wherein the sample user feature information includes at least user ID, user age, and user interests; the target object is a product or an advertisement; the target object feature information includes at least target object ID, target object type, and target object attributes; the interactive behavior includes at least the sample user's click, purchase, share, and favorite operations on the target object; The target object feature information is processed based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction; Based on the sample feature information and the multiple adjacency matrices, the prediction module is used to obtain the predicted click-through rate of the sample users for the target object; The click-through rate prediction model is trained based on the tag value and the click-through rate prediction value. The hypernetwork module includes multiple one-hot coding units and a graph generator corresponding to each one-hot coding unit; the processing of the target object feature information based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction includes: The target object feature information is input into the plurality of one-hot coding units respectively to obtain the one-hot coded features output by each one-hot coding unit; The one-hot encoded features output by each one-hot encoded unit are input into the corresponding graph generator to obtain the output results of each graph generator; The outputs of each graph generator are concatenated to obtain a first adjacency matrix for feature interaction. The first adjacency matrix is ​​iterated, and the first adjacency matrix and the adjacency matrix obtained after iteration are determined as the plurality of adjacency matrices; The prediction module includes multiple graph neural network units and prediction units. The number of graph neural network units is the same as the number of adjacency matrices, and each graph neural network unit corresponds to one adjacency matrix. The step of obtaining the predicted click-through rate of the sample user for the target object using the prediction module based on the sample feature information and the multiple adjacency matrices includes: Based on the sample feature information and the multiple adjacency matrices, the higher-order feature information output by each of the multiple graph neural network units is obtained using the multiple graph neural network units; Based on the high-order feature information output by each graph neural network unit, the prediction unit obtains the predicted click-through rate of the sample users for the target object.

2. The method as described in claim 1, wherein, The prediction module further includes a feature fusion unit; the step of obtaining the predicted click-through rate of the sample users for the target object using the prediction unit based on the high-order feature information output by each graph neural network unit includes: Based on the feature fusion unit, an attention mechanism is used to fuse the high-order feature information output by each graph neural network unit to obtain the feature representation of each feature in the sample feature information. The feature representations of each feature in the sample feature information are input into the prediction unit to obtain the predicted click-through rate of the sample user for the target object.

3. The method as described in claim 1 or 2, wherein, The step of obtaining high-order feature information output by each graph neural network unit using the multiple graph neural network units based on the sample feature information and the multiple adjacency matrices includes: Using the first graph neural network unit, the neighbor features of the sample feature information are determined based on the corresponding adjacency matrix, and the sample feature information and its neighbor features are fused to obtain the high-order feature information output by the first graph neural network unit; Adopting the first The first graph neural network unit determines the neighbor features of the sample feature information based on the corresponding adjacency matrix, and then... The high-order feature information output by the -1th graph neural network unit is fused with the neighbor features to obtain the -1st... The high-order feature information output by each of the graph neural network units; where 1 < ≤ , The number of units in the graph neural network; This continues until the high-order feature information output by all the graph neural network units is obtained.

4. The method of claim 1, wherein, The step of training the click-through rate prediction model based on the tag value and the click-through rate prediction value includes: Calculate the loss function based on the tag value and the predicted click-through rate value; The model parameters of the hypernetwork module and the model parameters of the prediction module are adjusted according to the loss function.

5. A click-through rate prediction method, comprising: Obtain user and target object feature information; The user's feature information and the target object's feature information are input into a pre-trained click-through rate prediction model; wherein the click-through rate prediction model is trained using the method described in any one of claims 1 to 4; Obtain the click-through rate prediction value output by the click-through rate prediction model, and determine the click-through rate prediction value as the probability that the user interacts with the target object.

6. A training device for a click-through rate prediction model, wherein, The click-through rate prediction model includes a hypernetwork module and a prediction module, and the training device includes: The first acquisition module is used to acquire sample feature information and tag values; the sample feature information includes sample user feature information and target object feature information; the tag values ​​are used to indicate whether the sample user has interactive behavior with the target object, wherein the sample user feature information includes at least user ID, user age, and user interests; the target object is a product or an advertisement; the target object feature information includes at least target object ID, target object type, and target object attributes; the interactive behavior includes at least the sample user's click, purchase, share, and favorite operations on the target object; The second acquisition module is used to process the feature information of the target object based on the hypernetwork module to obtain multiple adjacency matrices for feature interaction. The click-through rate prediction module is used to obtain the predicted click-through rate value of the sample user for the target object based on the sample feature information and the multiple adjacency matrices. The training module is used to train the click-through rate prediction model based on the label value and the click-through rate prediction value. The hypernetwork module includes multiple one-hot coding units and a graph generator corresponding to each one-hot coding unit; the second acquisition module is specifically used for: The target object feature information is input into the plurality of one-hot coding units respectively to obtain the one-hot coded features output by each one-hot coding unit; The one-hot encoded features output by each one-hot encoded unit are input into the corresponding graph generator to obtain the output results of each graph generator; The outputs of each graph generator are concatenated to obtain a first adjacency matrix for feature interaction. The first adjacency matrix is ​​iterated, and the first adjacency matrix and the adjacency matrix obtained after iteration are determined as the plurality of adjacency matrices; The prediction module includes multiple graph neural network units and prediction units. The number of graph neural network units is the same as the number of adjacency matrices, and each graph neural network unit corresponds to one adjacency matrix. The click-through rate prediction module includes: The acquisition unit is used to acquire the high-order feature information output by each of the graph neural network units based on the sample feature information and the plurality of adjacency matrices. The prediction unit is used to obtain the predicted click-through rate of the sample user for the target object based on the high-order feature information output by each graph neural network unit.

7. The apparatus of claim 6, wherein, The prediction module further includes a feature fusion unit; the prediction unit is specifically used for: Based on the feature fusion unit, an attention mechanism is used to fuse the high-order feature information output by each graph neural network unit to obtain the feature representation of each feature in the sample feature information. The feature representations of each feature in the sample feature information are input into the prediction unit to obtain the predicted click-through rate of the sample user for the target object.

8. The apparatus of claim 6 or 7, wherein, The acquisition unit is specifically used for: Using the first graph neural network unit, the neighbor features of the sample feature information are determined based on the corresponding adjacency matrix, and the sample feature information and its neighbor features are fused to obtain the high-order feature information output by the first graph neural network unit; Adopting the first The first graph neural network unit determines the neighbor features of the sample feature information based on the corresponding adjacency matrix, and then... The high-order feature information output by the -1th graph neural network unit is fused with the neighbor features to obtain the -1st... The high-order feature information output by each of the graph neural network units; where 1 < ≤ , The number of units in the graph neural network; This continues until the high-order feature information output by all the graph neural network units is obtained.

9. The apparatus of claim 6, wherein, The training module is specifically used for: Calculate the loss function based on the tag value and the predicted click-through rate value; The model parameters of the hypernetwork module and the model parameters of the prediction module are adjusted according to the loss function.

10. A click-through rate prediction device, comprising: The first acquisition module is used to acquire the user's feature information and the target object's feature information; The second acquisition module is used to input the feature information of the user and the feature information of the target object into a pre-trained click-through rate prediction model, acquire the click-through rate prediction value output by the click-through rate prediction model, and determine the click-through rate prediction value as the probability of the user interacting with the target object; wherein the click-through rate prediction model is trained using the method described in any one of claims 1 to 4.

11. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 4, or to perform the method of claim 5.

12. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method of any one of claims 1 to 4, or to perform the method of claim 5.

13. A computer program product comprising a computer program, wherein, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 4, or the steps of the method according to claim 5.

Images