Recommendation information click probability prediction method and device and computer equipment

Abstract

The invention relates to a recommendation information click probability prediction method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: converting data in a recommendation information click data table into feature elements to obtain an embedded matrix; performing different pooling processing on the embedded matrix, and performing cross fusion on feature vectors obtained after different pooling processing to obtain pooling feature vectors; converting the pooling feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector and the embedded matrix to obtain a first feature representation; converting the embedded matrix into a second feature representation through a preset projection weight and a second activation function; and fusing the first feature representation and the second feature representation to obtain a combined feature representation, and determining the probability that the target object clicks the recommendation information according to the combined feature representation. By adopting the method, the prediction accuracy of the recommendation information click rate can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, computer device, computer-readable storage medium, and computer program product for predicting click probability of recommended information. Background Art

[0002] The prediction of click-through rate of recommended information plays a vital role in the fields of online recommended information display and recommendation system. The importance of predicting the click-through rate of recommended information lies in that it can accurately rank a small number of items from a large number of candidate items, which can help service providers attract high-quality users. Accurate prediction of click-through rate of recommended information can improve user positioning strategies and increase revenue.

[0003] The challenge of improving prediction accuracy is how to continuously explore the implicit relationships between features of different categories, thereby improving the model's fitting ability. However, traditional methods for estimating click-through rates of recommended information usually rely on manual feature selection, which is time-consuming and labor-intensive, and may miss some potentially important features. They are often based on simple statistical methods or linear models, which are difficult to capture the complex nonlinear relationship between user information and recommended information, resulting in limited prediction accuracy. Summary of the invention

[0004] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium and computer program product for predicting the click probability of recommended information, which can improve the prediction accuracy of the click rate of recommended information, in order to address the above technical problems.

[0005] In a first aspect, the present application provides a method for predicting click probability of recommended information, comprising:

[0006] Obtaining a recommendation information click data table for the target object and the recommendation information, and converting the data in the recommendation information click data table into feature elements to obtain an embedding matrix;

[0007] Performing different pooling processes on the embedding matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain a pooled feature vector;

[0008] Converting the pooled feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector with the embedding matrix to obtain a first feature representation;

[0009] By presetting projection weights and a second activation function, the embedding matrix is ​​converted into a second feature representation;

[0010] The first feature representation and the second feature representation are merged to obtain a combined feature representation, and the probability of the target object clicking on the recommended information is determined according to the combined feature representation.

[0011] In one embodiment, the data in the recommended information click data table includes a plurality of columns of data; converting the data in the recommended information click data table into feature elements to obtain an embedding matrix includes:

[0012] For each column data in the recommended information click data table, based on the type of the column data, feature encoding is performed on the column data to obtain a plurality of initial feature elements; a preset embedding dimension is obtained, and according to the preset embedding dimension, the plurality of initial feature elements are mapped into a plurality of embedding vectors; the plurality of embedding vectors are combined to obtain an embedding matrix.

[0013] In one embodiment, the number of columns of the embedding matrix is ​​the number of embedding dimensions; and performing different pooling processes on the embedding matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain the pooled feature vector, comprises:

[0014] For each embedding dimension of the embedding matrix, extract the maximum value of the embedding matrix on the embedding dimension to obtain a maximum eigenvector; for each embedding dimension of the embedding matrix, extract the average value of the embedding matrix on the embedding dimension to obtain an average eigenvector; multiply the maximum eigenvector and the average eigenvector element by element to obtain a cross eigenvector; linearly combine the maximum eigenvector, the average eigenvector and the cross eigenvector to obtain a pooled eigenvector.

[0015] In one embodiment, the number of columns of the pooled feature vector is the number of feature dimensions; the first activation function is used to convert the pooled feature vector into an attention weight feature vector, and the attention weight feature vector is fused with the embedding matrix to obtain a first feature representation, including:

[0016] For each feature dimension of the linear feature vector, a first activation function is used to convert the feature elements of the linear feature vector on the feature dimension into attention scores to obtain an attention weight feature vector; the attention weight feature vector and the embedding matrix are multiplied to obtain a first feature representation.

[0017] In one embodiment, the preset projection weights include a first projection weight, a second projection weight, and a third projection weight; and converting the embedding matrix into a second feature representation by using the preset projection weights and a second activation function includes:

[0018] The embedding matrix is ​​transformed into an initial projection matrix by respectively using the first projection weight and the second projection weight; for each feature dimension of the initial projection matrix, the feature elements of the initial projection matrix on the feature dimension are converted into attention scores by using a second activation function to obtain an attention weight matrix; and the attention weight matrix is ​​transformed into a second feature representation by using the third projection weight.

[0019] In one embodiment, the fusing the first feature representation and the second feature representation to obtain a combined feature representation, and determining the probability of the target object clicking on the recommended information according to the combined feature representation, includes:

[0020] The first feature representation and the embedding matrix are multiplied to obtain a first matrix to be projected, and the first matrix to be projected is transformed into a first network output matrix through a fourth projection weight; the first feature representation and the second feature representation are concatenated to obtain a second matrix to be projected, and the second matrix to be projected is transformed into a second network output matrix through a fifth projection weight; the first network output matrix and the second network output matrix are added to obtain a merged feature representation; the merged feature representation is converted into an attention weight feature representation through a third activation function, and based on the attention score in the attention weight feature representation, the probability of the target object clicking on the recommended information is determined.

[0021] In a second aspect, the present application also provides a device for predicting click probability of recommended information, comprising:

[0022] An acquisition module is used to acquire a recommendation information click data table for the target object and the recommendation information, and convert the data in the recommendation information click data table into feature elements to obtain an embedding matrix;

[0023] A cross module, used to perform different pooling processes on the embedding matrix respectively, and cross-fuse the feature vectors obtained after the different pooling processes to obtain a pooled feature vector;

[0024] A fusion module, configured to convert the pooled feature vector into an attention weight feature vector through a first activation function, and fuse the attention weight feature vector with the embedding matrix to obtain a first feature representation;

[0025] A conversion module, used for converting the embedding matrix into a second feature representation by presetting a projection weight and a second activation function;

[0026] A determination module is used to merge the first feature representation and the second feature representation to obtain a combined feature representation, and determine the probability of the target object clicking on the recommended information according to the combined feature representation.

[0027] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0028] Obtaining a recommendation information click data table for the target object and the recommendation information, and converting the data in the recommendation information click data table into feature elements to obtain an embedding matrix;

[0029] Performing different pooling processes on the embedding matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain a pooled feature vector;

[0030] Converting the pooled feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector with the embedding matrix to obtain a first feature representation;

[0031] By presetting projection weights and a second activation function, the embedding matrix is ​​converted into a second feature representation;

[0032] The first feature representation and the second feature representation are merged to obtain a combined feature representation, and the probability of the target object clicking on the recommended information is determined according to the combined feature representation.

[0033] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0034] Obtaining a recommendation information click data table for the target object and the recommendation information, and converting the data in the recommendation information click data table into feature elements to obtain an embedding matrix;

[0035] Performing different pooling processes on the embedding matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain a pooled feature vector;

[0036] Converting the pooled feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector with the embedding matrix to obtain a first feature representation;

[0037] By presetting projection weights and a second activation function, the embedding matrix is ​​converted into a second feature representation;

[0038] The first feature representation and the second feature representation are merged to obtain a combined feature representation, and the probability of the target object clicking on the recommended information is determined according to the combined feature representation.

[0039] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0040] Obtaining a recommendation information click data table for the target object and the recommendation information, and converting the data in the recommendation information click data table into feature elements to obtain an embedding matrix;

[0041] Performing different pooling processes on the embedding matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain a pooled feature vector;

[0042] Converting the pooled feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector with the embedding matrix to obtain a first feature representation;

[0043] By presetting projection weights and a second activation function, the embedding matrix is ​​converted into a second feature representation;

[0044] The first feature representation and the second feature representation are merged to obtain a combined feature representation, and the probability of the target object clicking on the recommended information is determined according to the combined feature representation.

[0045] The above-mentioned recommendation information click probability prediction method, device, computer equipment, computer-readable storage medium and computer program product obtain the recommendation information click data table for the target object and the recommendation information, and convert the data in the recommendation information click data table into feature elements to obtain an embedding matrix; convert the original data into feature elements to facilitate model processing. In addition, through the form of an embedding matrix, high-dimensional sparse features can be effectively converted into low-dimensional dense vectors, reducing the computational complexity while retaining the relationship between features.

[0046] Different pooling processes are performed on the embedding matrix respectively, and the feature vectors obtained after different pooling processes are cross-fused to obtain pooled feature vectors; features can be extracted from different angles to capture different types of feature information. The feature vectors obtained by different pooling operations are cross-fused to combine the advantages of the two pooling operations, obtain more comprehensive feature information, and improve the robustness and prediction ability of the model.

[0047] Through the first activation function, the pooled feature vector is converted into an attention weight feature vector, and the attention weight feature vector and the embedding matrix are fused to obtain the first feature representation; through the attention mechanism, important features can be automatically learned and emphasized, and irrelevant features can be ignored. This helps to improve the model's sensitivity to key features and improve the accuracy of prediction. Fusion of the attention weight feature vector and the embedding matrix can further enhance the influence of important features and improve the expressiveness of the model.

[0048] By presetting the projection weights and the second activation function, the embedding matrix is ​​converted into a second feature representation; the presetting projection weights can convert the embedding matrix into a new feature representation and further extract high-order feature information. This helps to capture more complex feature relationships and improve the model's fitting ability. The use of activation functions can introduce nonlinear transformations, enabling the model to better capture nonlinear relationships between features and improve prediction accuracy.

[0049] The first feature representation and the second feature representation are fused to obtain a merged feature representation, and the probability of the target object clicking on the recommended information is determined based on the merged feature representation. The fusion of the two feature representations can combine the advantages of the two feature representations and improve the overall performance of the model. The first feature representation enhances the influence of important features through the attention mechanism, while the second feature representation extracts high-order feature information through projection transformation. The combination of the two can more comprehensively capture the complex relationship between features. The final merged feature representation is used to predict the probability of the target object clicking on the recommended information, which can more accurately predict user behavior.

[0050] Through the transformation of the embedding matrix in the above steps, the cross-fusion of different pooling processes, the introduction of the attention mechanism and the comprehensive representation of features, the complex relationship between user information and recommendation information can be captured more accurately, and the fitting ability and prediction accuracy of the model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0052] FIG1 is a diagram showing an application environment of a method for predicting click probability of recommended information in one embodiment;

[0053] FIG2 is a schematic diagram of a flow chart of a method for predicting click probability of recommended information in one embodiment;

[0054] FIG3 is a schematic diagram of an overall network of a method for predicting click probability of recommended information in one embodiment;

[0055] FIG4 is a schematic diagram of a maximum average cross-attention network of a method for predicting click probability of recommended information in one embodiment;

[0056] FIG5 is a schematic diagram of an element-level cross-attention network of a method for predicting click probability of recommended information in one embodiment;

[0057] FIG6 is a schematic diagram of a flow chart of a method for predicting click probability of recommended information in another embodiment;

[0058] FIG7 is a structural block diagram of a device for predicting click probability of recommended information in one embodiment;

[0059] FIG. 8 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0061] The method for predicting the click probability of recommended information provided in the embodiment of the present application can be applied in the application environment shown in FIG. 1. The terminal 102 communicates with the server 104 through a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on a cloud or other network server. The terminal 102 is used to generate a click probability prediction request for recommended information, and send the click probability prediction request for recommended information to the server 104, so that the server 104 merges the first feature representation and the second feature representation to obtain a merged feature representation, and determines the probability of the target object clicking the recommended information according to the merged feature representation. The terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices, and the Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The server 104 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0062] In an exemplary embodiment, as shown in FIG2 , a method for predicting click probability of recommended information is provided, and the method is applied to the server 104 in FIG1 as an example for description, including the following steps 202 to 210. Among them:

[0063] Step 202: Obtain a recommendation information click data table for the target object and the recommendation information, and convert the data in the recommendation information click data table into feature elements to obtain an embedding matrix.

[0064] The target object refers to the user in the advertising system or recommendation system. These users are the target audience of the advertisement or recommendation information, and it is necessary to predict whether they will click on a certain recommendation information. It can be understood that the target object can be one user or multiple users. This application does not make specific restrictions and can be set according to actual needs.

[0065] Recommended information refers to advertisements or recommended content displayed to users in advertising systems or recommendation systems. These contents can be in various forms such as products, articles, videos, etc., with the purpose of attracting users' attention and prompting them to click.

[0066] The recommended information click data table refers to an Excel worksheet that records the user's interaction with the recommended information. The Excel worksheet includes multiple columns of data, and different columns of data are different feature data, including user feature data and recommended information feature data. For example, user feature data can be user age, user gender, and user interests, etc., and recommended information feature data can be recommended information category, recommended information release time, etc.

[0067] Feature elements refer to specific feature values ​​extracted from the recommended information click data table. These feature values ​​can be user attributes, recommended information attributes, or other related variables. Feature elements are the basis of model input and are used to describe various attributes of users and recommended information.

[0068] The embedding matrix converts discrete features such as user feature data and recommendation information feature data into continuous vector representations, so that the model can more efficiently process and learn the relationship between these features. It can be understood that each row of the embedding matrix corresponds to the vector representation of a feature element.

[0069] Specifically, firstly, a recommendation information click data table for the target object and the recommendation information is obtained, including obtaining user feature data and recommendation information feature data from the data storage system, and arranging the user feature data and recommendation information feature data into a structured table form to ensure the consistency and integrity of the data.

[0070] Then convert the data in the recommended information click data table into feature elements, including extracting the attribute values ​​of each user and recommended information from the data table to form feature data. Select feature data that has an impact on the prediction of the recommended information click rate, such as user ID, recommended information ID, user age, user gender, recommended information category, etc. Preprocess the feature data and finally obtain feature elements. For example, characterize the classification features and process missing values ​​(replace with fixed values).

[0071] Finally, an embedding layer is used to convert discrete features (such as user ID, recommendation information ID) into continuous vector representations. A fixed-length vector representation is generated for each feature element. All feature vectors are combined into a matrix, with each row representing the vector representation of a feature element.

[0072] Step 204 , performing different pooling processes on the embedded matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain a pooled feature vector.

[0073] Among them, the pooled feature vector refers to the feature vector obtained by performing different pooling operations on the embedded matrix, and then the comprehensive feature vector is obtained by cross-fusion of these feature vectors.

[0074] Specifically, different types of feature information are first extracted from the embedding matrix through different pooling operations, including extracting the maximum value from each eigenvector of the embedding matrix to generate a new eigenvector. Maximum pooling can capture the most significant features, and calculating the average value from each eigenvector of the embedding matrix to generate a new eigenvector. Average pooling can capture the overall feature distribution.

[0075] Then, the feature vectors obtained by different pooling operations are combined to generate a comprehensive pooling feature vector, including concatenating the maximum pooling feature vector and the average pooling feature vector to form a longer feature vector. The maximum pooling feature vector and the average pooling feature vector are fused in a weighted manner to generate a comprehensive pooling feature vector.

[0076] Step 206: Convert the pooled feature vector into an attention weight feature vector through a first activation function, and fuse the attention weight feature vector and the embedding matrix to obtain a first feature representation.

[0077] Among them, the first activation function refers to a function used to introduce nonlinear transformation in a neural network, so that the model can better capture the complex relationship between features. In this application, the first activation function is used to convert the pooled feature vector into a nonlinear attention weight feature vector, so that the model can capture more complex feature relationships.

[0078] The attention weight feature vector is a set of weight values ​​used to emphasize the importance of certain features in the embedding matrix. In this application, the attention weight feature vector can be used to assign a weight to each feature, highlighting important features and ignoring unimportant features. The attention weight feature vector is dot-multiplied with the embedding matrix to generate a weighted feature representation.

[0079] The first feature representation refers to the feature representation obtained by fusing the attention weight feature vector with the embedding matrix. Through this fusion, a feature representation that pays more attention to important features can be generated.

[0080] Specifically, the pooled feature vector is first converted into an attention weight feature vector through a first activation function, including selecting a suitable activation function such as a Sigmoid activation function (commonly used as an activation function of the output layer for binary classification problems), and inputting the pooled feature vector into the selected activation function to obtain an attention weight feature vector.

[0081] Then, the attention weight feature vector and the embedding matrix are fused to obtain a first feature representation, including element-by-element multiplication of the attention weight feature vector and the embedding matrix to generate the first feature representation. It can be understood that the first feature representation, as a weighted feature matrix, can highlight important features.

[0082] Step 208: Convert the embedding matrix into a second feature representation by presetting the projection weight and the second activation function.

[0083] The preset projection weights refer to a set of fixed weight values ​​used to convert the embedding matrix into a new feature representation. These weights are usually obtained through pre-training or setting to capture high-order feature information in the embedding matrix. The preset projection weights can change the dimension of the feature representation to make it more suitable for subsequent processing and model training. By presetting the projection weights, the embedding matrix is ​​converted into a new feature representation to capture higher-order feature relationships.

[0084] The second activation function refers to a function used to introduce nonlinear transformations in a neural network, so that the model can better capture the complex relationships between features. In this application, the second activation function can be used to convert new feature representations into nonlinear forms, so that the model can capture more complex feature relationships.

[0085] The second feature representation refers to the new feature representation obtained by transforming the embedding matrix through the preset projection weights and the second activation function. Through this transformation, a more advanced and complex feature representation can be generated to capture the high-order feature relationships in the embedding matrix.

[0086] Specifically, the embedding matrix is ​​first used to perform matrix multiplication with the preset projection weight matrix to generate a new feature representation, and then the new feature representation is input into the second activation function to generate a nonlinear second feature representation.

[0087] Step 210: The first feature representation and the second feature representation are merged to obtain a combined feature representation, and the probability of the target object clicking on the recommended information is determined based on the combined feature representation.

[0088] The combined feature representation refers to fusing the first feature representation and the second feature representation to generate a comprehensive feature representation. It can be understood that this comprehensive feature representation can more comprehensively capture the complex relationship between users and recommendation information and improve the prediction ability of the model.

[0089] Specifically, the first feature representation and the second feature representation are first fused to generate a merged feature representation. Optionally, the first feature representation and the second feature representation can be spliced ​​along the feature dimension. The first feature representation and the second feature representation can also be fused by weighted fusion or other methods.

[0090] The generated merged feature representation is then input into the fully connected layer to generate an intermediate representation. An activation function (such as the Sigmoid activation function) is used in the output layer to convert the intermediate representation into a click probability, that is, the probability of the target object clicking on the recommended information.

[0091] In one of the embodiments, for each column data in the recommended information click data table, based on the type of the column data, feature encoding is performed on the column data to obtain multiple initial feature elements; a preset embedding dimension is obtained, and according to the preset embedding dimension, the multiple initial feature elements are mapped into multiple embedding vectors; and the multiple embedding vectors are combined to obtain an embedding matrix.

[0092] The initial feature element refers to the specific value of each feature extracted from the recommended information click data table. These feature values ​​can be user attributes, recommended information attributes or other related variables.

[0093] The preset embedding dimension refers to the dimension of the vector set when mapping the initial feature elements to the embedding vector. The embedding dimension determines the length of the embedding vector and is usually a small positive integer.

[0094] An embedding vector is a low-dimensional dense vector that maps the initial feature element. Each embedding vector represents the position of a feature element in the embedding space.

[0095] Specifically, referring to Figure 3, first extract the specific value of each column data from the recommended information click data table. For example, the recommended information click data table contains column data for user gender, then the specific value of the user gender of the column data extracted from the recommended information click data table is male, female, female. Then, according to the type of column data, feature encoding is performed on the extracted specific values. For example, according to the type of user gender, feature encoding is performed on the extracted specific values ​​male, female, female (male=1, female=0), and the initial feature elements 1, 0, 0 are obtained. Next, according to the preset embedding dimension, the embedding layer is used to map the initial feature elements into embedding vectors. Finally, the embedding vectors of each feature are arranged in order to form an embedding matrix. The number of rows of the embedding matrix is ​​the number of samples, and the number of columns is the embedding dimension.

[0096] Through feature encoding, different feature types can be unified into numerical representations, making feature processing and analysis more intuitive and explainable. And by mapping the initial feature elements into embedded vectors through the embedding layer, high-dimensional sparse features can be converted into low-dimensional dense vectors, enriching the representation of features and enabling the model to capture more feature information.

[0097] In one of the embodiments, for each embedding dimension of the embedding matrix, the maximum value of the embedding matrix on the embedding dimension is extracted to obtain the maximum eigenvector; for each embedding dimension of the embedding matrix, the average value of the embedding matrix on the embedding dimension is extracted to obtain the average eigenvector; the maximum eigenvector and the average eigenvector are multiplied element by element to obtain a cross eigenvector; the maximum eigenvector, the average eigenvector and the cross eigenvector are linearly combined to obtain a pooled eigenvector.

[0098] The embedding dimension refers to the fact that when processing text or category data, an embedding layer is usually used to convert the original high-dimensional sparse features (such as words or category IDs) into a low-dimensional dense vector, which is called an embedding vector. The dimension of the embedding vector is the embedding dimension, which determines the length of the embedding vector, that is, the size of the space represented by each feature after conversion. For example, in the given example, the embedding dimension is set to 10.

[0099] The maximum eigenvector means that for each embedding dimension, the maximum value of that dimension is extracted from the embedding matrix to form a new vector, namely the maximum eigenvector. This operation can be achieved through max pooling, the purpose of which is to capture the most significant eigenvalues ​​in each dimension, which helps the model focus on those features that are critical to the task.

[0100] The average feature vector means that for each embedding dimension, the average value on that dimension is extracted from the embedding matrix to form the average feature vector. This is usually done through average pooling, which aims to reflect the overall level of all feature values ​​rather than just focusing on the largest value.

[0101] The cross eigenvector is the result of element-wise multiplication of the maximum eigenvector and the average eigenvector. This operation can be understood as an interaction between two eigenvectors, the purpose of which is to capture the relationship between the two, which may reveal new information or patterns.

[0102] Specifically, referring to Figure 4, in the maximum average cross attention network, first from the embedding matrix, for each embedding dimension, the maximum value on that dimension is extracted to form the maximum eigenvector. For example, the shape of the embedding matrix is ​​(N, D), where N is the number of samples and D is the embedding dimension. For each embedding dimension d (d = 1, 2, ..., D), the maximum value of all samples on that dimension is found to form a vector of length D, i.e., the maximum eigenvector.

[0103] Then, for each embedding dimension, the average value on the dimension is extracted from the embedding matrix to form the average feature vector. For example, the shape of the embedding matrix is ​​(N, D). For each embedding dimension d (d = 1, 2, ..., D), the average value of all samples on this dimension is calculated to form a vector of length D, which is the average feature vector.

[0104] Next, the maximum eigenvector and the average eigenvector are multiplied element by element to obtain the cross eigenvector.

[0105] Finally, the maximum eigenvector, average eigenvector and cross eigenvector are linearly combined using the following formula to generate the final pooled eigenvector:

[0106] V pool =w1V max + w2V avg +w3V cross

[0107] Among them, w1, w2, and w3 are weight coefficients, which can be fixed values ​​or obtained through model learning. max is the largest eigenvector, V avg is the average eigenvector, V cross is the cross eigenvector, V pool is the pooled feature vector.

[0108] By extracting the maximum value on each embedding dimension, the most significant feature value on each dimension can be captured, helping the model focus on features that are critical to the task. By extracting the average value on each embedding dimension, the overall level of all feature values ​​can be reflected, providing a global perspective and avoiding focusing only on extreme values. By combining the maximum value and the average value, the model can not only focus on extreme values, but also take into account the overall distribution, so that it performs more stably under different data distributions.

[0109] In one of the embodiments, for each feature dimension of the linear feature vector, a first activation function is used to convert the feature elements of the linear feature vector in the feature dimension into attention scores to obtain an attention weight feature vector; the attention weight feature vector and the embedding matrix are multiplied to obtain a first feature representation.

[0110] The feature dimension refers to the position corresponding to each element in the feature vector. It can be understood that the feature vector is usually a multi-dimensional vector, and each dimension represents a specific feature or attribute.

[0111] Attention score is a value used to measure the importance of a feature or feature dimension in the current task. Through the attention mechanism, different weights can be dynamically assigned to different feature dimensions, so that the model can pay more attention to important features. In this application, the attention score is obtained by transforming the feature elements of the linear feature vector in each feature dimension through a first activation function.

[0112] Specifically, the attention score is first calculated, including selecting an activation function, transforming each element in the linear feature vector, and obtaining the attention score. For example, the linear feature vector is [0.1, 0.5, 0.3], the Sigmoid function is selected as the activation function, and the attention score of the linear feature vector in each feature dimension is calculated to obtain the attention weight feature vector = [0.525, 0.622, 0.575].

[0113] Then, the attention weight feature vector and the embedding matrix are multiplied to obtain the first feature representation. For example, the shape of the embedding matrix is ​​(N, D), where N is the number of samples and D is the embedding dimension. The attention weight feature vector is multiplied element by element with the embedding matrix to obtain the first feature representation.

[0114] Because of the attention mechanism, the model can dynamically assign different weights to each feature dimension. This allows the model to pay more attention to features that are more important to the current task and pay less attention to unimportant features. And the attention mechanism can capture complex relationships between features, especially those non-linear relationships. By multiplying the attention weight feature vector with the embedding matrix, a richer feature representation can be generated, which can better capture the complex patterns in the data.

[0115] In one of the embodiments, the embedding matrix is ​​transformed into an initial projected matrix by using a first projection weight and a second projection weight respectively; for each feature dimension of the initial projected matrix, the feature elements of the initial projected matrix on the targeted feature dimension are converted into attention scores by using a second activation function to obtain an attention weight matrix; and the attention weight matrix is ​​transformed into a second feature representation by using a third projection weight.

[0116] The first projection weight refers to a weight coefficient used to transform the embedding matrix into the initial projection matrix, which is learned during the model training process.

[0117] The second projection weight also refers to a weight coefficient used to transform the embedding matrix into the initial projection matrix, which is learned during the model training process.

[0118] The third projection weight refers to a weight coefficient used to transform the attention weight matrix into the second feature representation. This weight matrix is ​​also learned during the model training process.

[0119] The initial post-projection matrix refers to a matrix obtained by transforming the embedding matrix by the first projection weight or the second projection weight.

[0120] The attention weight matrix refers to the matrix obtained by converting the feature elements of the initial projection matrix in each feature dimension into attention scores through the second activation function. It is used to dynamically assign different weights to different feature dimensions so that the model can pay more attention to important features.

[0121] Specifically, referring to Figure 5, in the element-level cross attention network, the embedding matrix is ​​first transformed into the initial projection matrix through the first projection weight and the second projection weight. For example, the shape of the embedding matrix E is (N, D), where N is the number of samples and D is the embedding dimension. The first projection weight is the weight coefficient A, and the second projection weight is the weight coefficient B. Perform the projection transformation and use A×E×B as the initial projection matrix obtained after projection.

[0122] Then, for each feature dimension of the initial projection matrix, the feature elements on each feature dimension are converted into attention scores through the Sigmoid function, and then all the attention scores obtained after the conversion are combined into a matrix, namely the attention weight matrix.

[0123] Next, the attention weight matrix is ​​transformed into the second feature representation through the third projection weight. For example, the third projection weight is the weight coefficient C, and the attention weight matrix is ​​Q. C×Q is taken as the second feature representation obtained after projection.

[0124] Because of multiple projections and transformations, the model can better capture the complex relationships between features, especially those nonlinear relationships. This helps to generate richer feature representations, thereby better capturing complex patterns in the data. And by dynamically assigning weights, the attention mechanism can reduce the impact of noise features on the model, making the model more stable and robust. In addition, when processing high-dimensional sparse data, the attention mechanism can effectively focus on important features, reduce the interference of invalid features, and improve the performance of the model on sparse data.

[0125] In one of the embodiments, the first feature representation and the embedding matrix are multiplied to obtain a first matrix to be projected, and the first matrix to be projected is transformed into a first network output matrix through a fourth projection weight; the first feature representation and the second feature representation are concatenated to obtain a second matrix to be projected, and the second matrix to be projected is transformed into a second network output matrix through a fifth projection weight; the first network output matrix and the second network output matrix are added to obtain a merged feature representation; the merged feature representation is converted into an attention weight feature representation through a third activation function, and based on the attention score in the attention weight feature representation, the probability of the target object clicking on the recommended information is determined.

[0126] The first matrix to be projected refers to a matrix obtained by multiplying the first feature representation and the embedding matrix, which is used for further projection transformation to generate a first network output matrix.

[0127] The fourth projection weight refers to a weight coefficient used to transform the first matrix to be projected into the first network output matrix. This weight matrix is ​​learned during the model training process.

[0128] The first network output matrix refers to the matrix obtained by transforming the first matrix to be projected through the fourth projection weight, which contains comprehensive information of the first feature representation and the embedding matrix, and is used for subsequent feature fusion and output.

[0129] The second matrix to be projected refers to a matrix obtained by concatenating the first feature representation and the second feature representation, which is used for further projection transformation to generate a second network output matrix.

[0130] The fifth projection weight refers to a weight coefficient used to transform the second matrix to be projected into a second network output matrix. This weight matrix is ​​learned during the model training process.

[0131] The second network output matrix refers to the matrix obtained by transforming the second matrix to be projected through the fifth projection weight. This matrix contains the comprehensive information of the first feature representation and the second feature representation, which is used for subsequent feature fusion and output.

[0132] The third activation function is a nonlinear function used to convert the merged feature representation into the attention weight feature representation. Through the third activation function, nonlinear characteristics can be introduced to enhance the expressiveness of the model.

[0133] The attention weight feature representation refers to the matrix obtained by transforming the merged feature representation through the third activation function. The attention weight feature representation contains the final attention score, which is used to determine the probability of the target object clicking on the recommended information.

[0134] Specifically, first multiply the first feature representation and the embedding matrix to obtain the first matrix to be projected. For example, the shape of the first feature representation F1 is (N1, M1), where N is the number of samples and M1 is the dimension of the first feature representation. The shape of the embedding matrix E is (N, D), where D is the embedding dimension. Calculate the first matrix to be projected, M1=F1×E, where the shape of M1 is (N, D).

[0135] Then, the first matrix to be projected is transformed into the first network output matrix by the fourth projection weight. The first matrix to be projected is transformed into the first network output matrix. For example, the fourth projection weight is the weight coefficient O, and the first matrix to be projected is P. O×P is taken as the first network output matrix obtained after projection.

[0136] Next, the first feature representation and the second feature representation are concatenated to obtain the second matrix to be projected. For example, the shape of the first feature representation F1 is (N, M1), and the shape of the second feature representation F2 is (N, M2), where M2 is the dimension of the second feature representation. The concatenated feature representation is M2=[F1, F2], where the shape of M2 is (N, M1+ M2).

[0137] The second matrix to be projected is then transformed into a second network output matrix by the fifth projection weight. The first network output matrix and the second network output matrix are added to obtain a merged feature representation. For example, the shape of the first network output matrix Y1 is (N, K1), and the shape of the second network output matrix Y2 is (N, K2). Optionally, if K1 and K2 are inconsistent, their dimensions can be made the same by padding or cropping. The first network output matrix Y1 and the second network output matrix Y2 are added to obtain a merged feature representation. For example, U = Y1 + Y2, where the shape of U is (N, K), and K is the dimension of the merged feature representation.

[0138] Finally, the merged feature representation is converted into an attention weight feature representation through the third activation function. Based on the attention score in the attention weight feature representation, the probability of the target object clicking on the recommended information is determined. For example, the shape of the attention weight feature representation is (N, K), which is obtained by taking a specific dimension of the attention weight feature representation or by other aggregation methods such as averaging, and can be set according to actual needs.

[0139] Since the first feature representation is concatenated with the second feature representation, feature information from different sources or different types can be effectively integrated, so that the model can understand the data from a more comprehensive perspective and make more accurate predictions. And using the third activation function to convert the merged feature representation into an attention weight feature representation can significantly improve the model's sensitivity to personalized factors such as user interest preferences, further optimizing the recommendation effect.

[0140] In an exemplary embodiment, as shown in FIG6 , steps 602 to 610 are included. Among them:

[0141] Step 602, obtaining a recommendation information click data table for the target object and the recommendation information, and for each column data in the recommendation information click data table, based on the type of the column data, feature encoding the column data to obtain a plurality of initial feature elements; obtaining a preset embedding dimension, and mapping the plurality of initial feature elements into a plurality of embedding vectors according to the preset embedding dimension; combining the plurality of embedding vectors to obtain an embedding matrix;

[0142] Step 604: for each embedding dimension of the embedding matrix, extract the maximum value of the embedding matrix on the embedding dimension to obtain the maximum eigenvector; for each embedding dimension of the embedding matrix, extract the average value of the embedding matrix on the embedding dimension to obtain the average eigenvector; multiply the maximum eigenvector and the average eigenvector element by element to obtain a cross eigenvector; linearly combine the maximum eigenvector, the average eigenvector and the cross eigenvector to obtain a pooled eigenvector;

[0143] Step 606: for each feature dimension of the linear feature vector, convert the feature elements of the linear feature vector on the feature dimension into attention scores through the first activation function to obtain an attention weight feature vector; multiply the attention weight feature vector and the embedding matrix to obtain a first feature representation;

[0144] Step 608, transforming the embedding matrix into an initial projection matrix by using the first projection weight and the second projection weight respectively; for each feature dimension of the initial projection matrix, transforming the feature elements of the initial projection matrix on the feature dimension into attention scores by using the second activation function to obtain an attention weight matrix; transforming the attention weight matrix into a second feature representation by using the third projection weight;

[0145] Step 610: multiply the first feature representation and the embedding matrix to obtain a first matrix to be projected, and transform the first matrix to be projected into a first network output matrix through the fourth projection weight; concatenate the first feature representation and the second feature representation to obtain a second matrix to be projected, and transform the second matrix to be projected into a second network output matrix through the fifth projection weight; add the first network output matrix and the second network output matrix to obtain a merged feature representation; convert the merged feature representation into an attention weight feature representation through the third activation function, and determine the probability of the target object clicking on the recommended information based on the attention score in the attention weight feature representation.

[0146] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0147] Based on the same inventive concept, the embodiment of the present application also provides a recommended information click probability prediction device for implementing the recommended information click probability prediction method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more recommended information click probability prediction device embodiments provided below can refer to the limitations of the recommended information click probability prediction method above, and will not be repeated here.

[0148] In an exemplary embodiment, as shown in FIG. 7 , a device 700 for predicting click probability of recommended information is provided, comprising: an acquisition module 702 , a cross module 704 , a fusion module 706 , a conversion module 708 and a determination module 710 , wherein:

[0149] An acquisition module 702 is used to acquire a recommendation information click data table for the target object and the recommendation information, and convert the data in the recommendation information click data table into feature elements to obtain an embedding matrix;

[0150] A cross module 704 is used to perform different pooling processes on the embedding matrix respectively, and cross-fuse the feature vectors obtained after the different pooling processes to obtain a pooled feature vector;

[0151] A fusion module 706 is used to convert the pooled feature vector into an attention weight feature vector through a first activation function, and fuse the attention weight feature vector with the embedding matrix to obtain a first feature representation;

[0152] A conversion module 708, configured to convert the embedding matrix into a second feature representation by presetting a projection weight and a second activation function;

[0153] The determination module 710 is used to fuse the first feature representation and the second feature representation to obtain a combined feature representation, and determine the probability of the target object clicking on the recommended information according to the combined feature representation.

[0154] In one embodiment, the acquisition module 702 is used to, for each column data in the recommendation information click data table, perform feature encoding on the column data based on the type of the column data to obtain multiple initial feature elements; obtain a preset embedding dimension, and map the multiple initial feature elements into multiple embedding vectors according to the preset embedding dimension; and combine the multiple embedding vectors to obtain an embedding matrix.

[0155] In one embodiment, the cross module 704 is used to extract the maximum value of the embedding matrix on each embedding dimension of the embedding matrix to obtain the maximum eigenvector; extract the average value of the embedding matrix on each embedding dimension of the embedding matrix to obtain the average eigenvector; multiply the maximum eigenvector and the average eigenvector element by element to obtain a cross eigenvector; linearly combine the maximum eigenvector, the average eigenvector and the cross eigenvector to obtain a pooled eigenvector.

[0156] In one embodiment, the fusion module 706 is used to convert the feature elements of the linear feature vector in the feature dimension into attention scores through a first activation function for each feature dimension of the linear feature vector, so as to obtain an attention weight feature vector; and multiply the attention weight feature vector and the embedding matrix to obtain a first feature representation.

[0157] In one embodiment, the conversion module 708 is used to transform the embedding matrix into an initial projection matrix through a first projection weight and a second projection weight respectively; for each feature dimension of the initial projection matrix, the feature elements of the initial projection matrix on the targeted feature dimension are converted into attention scores through a second activation function to obtain an attention weight matrix; and the attention weight matrix is ​​transformed into a second feature representation through a third projection weight.

[0158] In one embodiment, the determination module 710 is used to multiply the first feature representation and the embedding matrix to obtain a first matrix to be projected, and transform the first matrix to be projected into a first network output matrix through a fourth projection weight; concatenate the first feature representation and the second feature representation to obtain a second matrix to be projected, and transform the second matrix to be projected into a second network output matrix through a fifth projection weight; add the first network output matrix and the second network output matrix to obtain a merged feature representation; convert the merged feature representation into an attention weight feature representation through a third activation function, and determine the probability of the target object clicking on the recommended information based on the attention score in the attention weight feature representation.

[0159] Each module in the above-mentioned recommendation information click probability prediction device can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of each of the above modules.

[0160] In an exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG8. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data related to the prediction of the click probability of recommended information. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for predicting the click probability of recommended information is implemented.

[0161] Those skilled in the art will understand that the structure shown in FIG. 8 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0162] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0163] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0164] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0166] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0167] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0168] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for predicting click probability of recommended information, characterized in that: The method comprises: Obtaining a recommendation information click data table for the target object and the recommendation information, and converting the data in the recommendation information click data table into feature elements to obtain an embedding matrix; Performing different pooling processes on the embedding matrix respectively, and cross-merging the feature vectors obtained after the different pooling processes to obtain a pooled feature vector; Converting the pooled feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector with the embedding matrix to obtain a first feature representation; By presetting projection weights and a second activation function, the embedding matrix is ​​converted into a second feature representation; The first feature representation and the second feature representation are merged to obtain a combined feature representation, and the probability of the target object clicking on the recommended information is determined according to the combined feature representation.

2. The method according to claim 1, characterized in that The data in the recommended information click data table includes multiple columns of data; The step of converting the data in the recommended information click data table into feature elements to obtain an embedding matrix includes: For each column data in the recommended information click data table, based on the type of the targeted column data, feature encoding is performed on the targeted column data to obtain a plurality of initial feature elements; Obtaining a preset embedding dimension, and mapping the plurality of initial feature elements into a plurality of embedding vectors according to the preset embedding dimension; The multiple embedding vectors are combined to obtain an embedding matrix.

3. The method according to claim 1, characterized in that The number of columns of the embedding matrix is ​​the number of embedding dimensions; The respectively performing different pooling processes on the embedding matrix and cross-merging the feature vectors obtained after the different pooling processes to obtain the pooled feature vectors includes: For each embedding dimension of the embedding matrix, extract the maximum value of the embedding matrix on the embedding dimension to obtain a maximum eigenvector; For each embedding dimension of the embedding matrix, extracting the average value of the embedding matrix on the embedding dimension to obtain an average eigenvector; Multiplying the maximum eigenvector and the average eigenvector element by element to obtain a cross eigenvector; The maximum eigenvector, the average eigenvector and the cross eigenvector are linearly combined to obtain a pooled eigenvector.

4. The method according to claim 3, characterized in that The number of columns of the pooled feature vector is the number of feature dimensions; The method of converting the pooled feature vector into an attention weight feature vector through a first activation function, and fusing the attention weight feature vector with the embedding matrix to obtain a first feature representation includes: For each feature dimension of the linear feature vector, convert the feature elements of the linear feature vector on the feature dimension into attention scores through a first activation function to obtain an attention weight feature vector; The attention weight feature vector and the embedding matrix are multiplied to obtain a first feature representation.

5. The method according to claim 1, characterized in that The preset projection weights include a first projection weight, a second projection weight, and a third projection weight; The step of converting the embedding matrix into a second feature representation by presetting a projection weight and a second activation function includes: transforming the embedding matrix into an initial post-projection matrix by respectively using the first projection weight and the second projection weight; For each feature dimension of the initial projected matrix, convert the feature elements of the initial projected matrix on the feature dimension into attention scores through a second activation function to obtain an attention weight matrix; The attention weight matrix is ​​transformed into a second feature representation through the third projection weight.

6. The method according to claim 1, characterized in that The step of fusing the first feature representation and the second feature representation to obtain a combined feature representation, and determining the probability of the target object clicking on the recommended information according to the combined feature representation, includes: Multiplying the first feature representation and the embedding matrix to obtain a first matrix to be projected, and transforming the first matrix to be projected into a first network output matrix through a fourth projection weight; The first feature representation and the second feature representation are concatenated to obtain a second matrix to be projected, and the second matrix to be projected is transformed into a second network output matrix through a fifth projection weight; Adding the first network output matrix and the second network output matrix to obtain a combined feature representation; The merged feature representation is converted into an attention weight feature representation through a third activation function, and based on the attention score in the attention weight feature representation, the probability of the target object clicking on the recommended information is determined.

7. A device for predicting click probability of recommended information, characterized in that: The device comprises: An acquisition module is used to acquire a recommendation information click data table for the target object and the recommendation information, and convert the data in the recommendation information click data table into feature elements to obtain an embedding matrix; A cross module is used to perform different pooling processes on the embedding matrix respectively, and cross-fuse the feature vectors obtained after the different pooling processes to obtain a pooled feature vector; A fusion module, configured to convert the pooled feature vector into an attention weight feature vector through a first activation function, and fuse the attention weight feature vector with the embedding matrix to obtain a first feature representation; A conversion module, used for converting the embedding matrix into a second feature representation by presetting a projection weight and a second activation function; A determination module is used to fuse the first feature representation and the second feature representation to obtain a combined feature representation, and determine the probability of the target object clicking on the recommended information according to the combined feature representation.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images