Method, device and medium for training and applying a recommendation model based on multi-platform behavior

Abstract

The application discloses a recommendation model training and application method and device based on multi-platform behaviors, and a medium, relates to the technical field of user recommendation, and the method comprises the steps of constructing a mixed neural network model, the mixed neural network model comprising a feature embedding module, a multiple-hole convolution module, a multi-level attention mechanism module and a prediction module; the feature embedding module processes a multi-platform behavior sequence to obtain a feature embedding vector; the multiple-hole convolution module processes the feature embedding vector to obtain a multi-scale feature vector; the multi-level attention mechanism module processes the multi-scale feature vector and a commodity feature vector to obtain a user interest vector; the prediction module predicts a predicted click rate of a user on a candidate commodity based on the user interest vector and the commodity feature vector; the mixed neural network model is trained to obtain a recommendation model; and the recommendation model is used for recommendation. The application can train a recommendation model with better performance and improve recommendation performance.

Description

Technical Field

[0001] This application relates to the field of user recommendation technology, and in particular to a method, device and medium for training and applying a recommendation model based on multi-platform behavior. Background Technology

[0002] With the rapid development of internet technology, user behavior data across various platforms, such as e-commerce, content, and social media, is becoming increasingly abundant. Effectively utilizing this cross-platform behavioral data to build comprehensive user interest profiles and provide recommendations has become crucial for improving recommendation performance.

[0003] Current recommendation technologies mainly suffer from the following problems.

[0004] (1) Limitations of single-platform data: Most traditional recommendation technologies are based on user behavior data on a single platform, which cannot capture the complete decision-making process of users on multiple platforms. For example, a user may see product introductions on a content platform, get reviews on a social platform, and finally complete the purchase on an e-commerce platform. Behavioral data from a single platform cannot reflect this complete decision-making process.

[0005] (2) Difficulty in cross-platform data fusion: Traditional recommendation technology usually adopts simple feature splicing or early fusion strategy for processing multi-platform data, ignoring the heterogeneity of data from different platforms. Data from different platforms have different semantic meanings and confidence levels. Simple merging will lead to information loss and noise introduction.

[0006] (3) Insufficient sequence modeling capability: When traditional recommendation technology models user behavior sequences, although RNN (Recurrent Neural Network) and GRU (Gated Recurrent Unit) can process temporal information, they have the gradient vanishing problem when modeling long sequences. Although Transformer can handle long sequence dependencies, it has high computational complexity and is not sensitive enough to capturing local patterns.

[0007] (4) Insufficient feature crossover: Traditional recommendation technology models such as matrix factorization and factorization machine mainly rely on manual feature engineering or shallow feature crossover, which makes it difficult to discover deep and complex patterns in user behavior sequences. Summary of the Invention

[0008] The purpose of this application is to provide a method, device, and medium for training and applying a recommendation model based on multi-platform behavior. This method can train a recommendation model with better performance based on multi-platform behavior sequences and hybrid neural network models, thereby improving recommendation performance.

[0009] To achieve the above objectives, this application provides the following solution.

[0010] Firstly, this application provides a method for training a recommendation model based on multi-platform behavior, the method comprising:

[0011] Obtain the dataset; the dataset includes multiple sample data, the sample data includes user multi-platform behavior sequences, product feature vectors of candidate products, and user actual click-through rates on candidate products, the multi-platform behavior sequences are sequence data obtained by arranging user behavior data on multiple platforms in chronological order;

[0012] A hybrid neural network model is constructed. The hybrid neural network model includes a feature embedding module, a multi-dilated convolution module, a multi-level attention mechanism module, and a prediction module connected in sequence. The feature embedding module is used to process the multi-platform behavior sequence to obtain a feature embedding vector. The multi-dilated convolution module is used to process the feature embedding vector to obtain a multi-scale feature vector. The multi-level attention mechanism module is used to process the multi-scale feature vector and the product feature vector to obtain a user interest vector. The prediction module is used to predict the user's click-through rate for candidate products based on the user interest vector and the product feature vector.

[0013] The hybrid neural network model is trained using the dataset to obtain a recommendation model.

[0014] Secondly, this application provides a method for applying a recommendation model based on multi-platform behavior, the method comprising:

[0015] Obtain the multi-platform behavior sequence of the user to be recommended and the product feature vector of each product to be recommended;

[0016] For each product to be recommended, the multi-platform behavior sequence of the user to be recommended and the product feature vector of the product to be recommended are used as inputs. The recommendation model is used to determine the predicted click-through rate of the user to be recommended for the product to be recommended. The recommendation model is a model trained using the multi-platform behavior-based recommendation model training method described above.

[0017] Each product to be recommended is sorted in descending order of predicted click-through rate to obtain the sorting results, and recommendations are made to the users to be recommended based on the sorting results.

[0018] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the above-described method for training a recommendation model based on multi-platform behavior or the above-described method for applying a recommendation model based on multi-platform behavior.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for training a recommendation model based on multi-platform behavior or the above-described method for applying a recommendation model based on multi-platform behavior.

[0020] According to the specific embodiments provided in this application, this application has the following technical effects.

[0021] This application provides a method, device, and medium for training and applying a recommendation model based on multi-platform behavior. The method involves arranging user behavior data generated on multiple platforms in chronological order to obtain a multi-platform behavior sequence. A hybrid neural network model is then constructed, comprising a feature embedding module, a multi-dilated convolution module, a multi-level attention mechanism module, and a prediction module connected sequentially. The feature embedding module processes the multi-platform behavior sequence to obtain a feature embedding vector. The multi-dilated convolution module processes the feature embedding vector to obtain a multi-scale feature vector. The multi-level attention mechanism module processes the multi-scale feature vector and the product feature vector to obtain a user interest vector. The prediction module predicts the user's click-through rate for candidate products based on the user interest vector and the product feature vector. The hybrid neural network model is then trained to obtain a recommendation model, which is subsequently used for recommendations. This application addresses the limitations of single-platform data by incorporating user behavior sequences across multiple platforms. Furthermore, the design of a hybrid neural network model, including multiple dilated convolutional modules and multi-level attention mechanisms, overcomes the difficulties in cross-platform data fusion and insufficient sequence modeling capabilities. Finally, the processing of user behavior sequences using a hybrid neural network model resolves the issue of insufficient feature cross-fertilization. Through innovative network architecture design, intelligent fusion of cross-platform data and collaborative modeling of multi-scale sequence patterns are achieved. Consequently, a superior recommendation model can be trained based on multi-platform behavior sequences and the hybrid neural network model, thereby improving recommendation performance. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This application provides an application environment diagram for a recommendation model training and application method based on multi-platform behavior sequences.

[0024] Figure 2 This is a flowchart illustrating a recommendation model training method based on multi-platform behavior sequences provided in Embodiment 1 of this application.

[0025] Figure 3 This is a schematic diagram illustrating the overall implementation process of a recommendation model training method based on multi-platform behavior sequences provided in Embodiment 1 of this application.

[0026] Figure 4 This is a schematic diagram of the network structure of the feature embedding module provided in Embodiment 1 of this application.

[0027] Figure 5 This is a schematic diagram of the network structure of the multi-dilated convolutional module provided in Embodiment 1 of this application.

[0028] Figure 6 This is a schematic diagram of the network structure of the inter-platform importance attention layer provided in Embodiment 1 of this application.

[0029] Figure 7 This is a flowchart illustrating a recommendation model application method based on multi-platform behavior sequences provided in Embodiment 2 of this application.

[0030] Figure 8 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] Example 1.

[0033] The recommendation model training method based on multi-platform behavior provided in this application can be applied to, for example... Figure 1The application environment shown illustrates this. The terminal communicates with the server via a network. A data storage system stores the data the server needs to process. This data storage system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send a training request to the server. Upon receiving the request, the server acquires a dataset, which includes multiple sample data points. These sample data include user behavior sequences across multiple platforms, product feature vectors of candidate products, and the actual click-through rate of users on candidate products. A hybrid neural network model is constructed, consisting of sequentially connected feature embedding modules, multiple dilated convolutional modules, multi-level attention mechanisms, and a prediction module. The hybrid neural network model is trained using the dataset to obtain a recommendation model. The server can then feed back the training result—the recommendation model tailored to the training request—to the terminal.

[0034] Furthermore, in some embodiments, the recommendation model training method based on multi-platform behavior can also be implemented by the server or the terminal alone. For example, the terminal can directly process the training request to be processed, or the server can obtain the training request to be processed from the data storage system and process it.

[0035] In one exemplary embodiment, such as Figure 2 As shown, a recommendation model training method based on multi-platform behavior is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps are used as an example of a server in the example.

[0036] Step S1, obtain the dataset; the dataset includes multiple sample data, the sample data includes the user's multi-platform behavior sequence, the product feature vector of the candidate product, and the user's actual click rate on the candidate product. The multi-platform behavior sequence is the sequence data obtained by arranging the user's behavior data on multiple platforms in chronological order.

[0037] Step S2: Construct a hybrid neural network model. The hybrid neural network model includes a feature embedding module, a multi-dilated convolution module, a multi-level attention mechanism module, and a prediction module connected in sequence. The feature embedding module processes the multi-platform behavior sequence to obtain a feature embedding vector. The multi-dilated convolution module processes the feature embedding vector to obtain a multi-scale feature vector. The multi-level attention mechanism module processes the multi-scale feature vector and the product feature vector to obtain a user interest vector. The prediction module predicts the user's click-through rate for candidate products based on the user interest vector and the product feature vector.

[0038] Step S3: Train the hybrid neural network model using the dataset to obtain the recommendation model.

[0039] By implementing steps S1 to S3 above, this embodiment uses a multi-platform behavior sequence and a hybrid neural network model to model user interests and obtain a user interest vector, which represents the user's interest in candidate products. It further predicts the user's predicted click-through rate for candidate products and trains the hybrid neural network model in conjunction with the user's actual click-through rate for candidate products to obtain a recommendation model. The recommendation model is then used for targeted recommendations.

[0040] The purpose of this embodiment is to overcome the shortcomings of traditional technologies and provide a recommendation method that can effectively integrate cross-platform user behavior data and achieve refined user interest modeling through an innovative hybrid neural network model architecture, thereby improving the accuracy of personalized recommendations and user experience. The following, in conjunction with... Figure 3 The following is a detailed description of the recommendation model training method based on multi-platform behavior used in this embodiment, which includes the following steps.

[0041] (a) Dataset construction.

[0042] This embodiment acquires a dataset, which includes multiple sample data. The sample data includes user multi-platform behavior sequences, product feature vectors of candidate products, and the actual click-through rate of users on candidate products. The multi-platform behavior sequences are sequence data obtained by arranging user behavior data generated on multiple platforms in chronological order. The behavior data includes user ID (Identifier), behavior type, behavior object ID, timestamp, and platform ID. The user ID is the ID obtained after privacy protection processing of the real user ID. Behavior types include, but are not limited to, clicking, browsing, favoriting, and purchasing. The behavior object ID is the ID of the object operated by the user through the behavior. The timestamp is the time when the user behavior was generated. The platform ID is the platform on which the user behavior was generated.

[0043] like Figure 3 As shown, when obtaining a user's multi-platform behavior sequence, the process first acquires user behavior data across multiple platforms (at least two different platforms), including but not limited to clicks, browsing, favorites, and purchases. This data is then preprocessed to obtain the user's behavior sequence on each platform, such as platform A, platform B, and platform C. Finally, data alignment and preprocessing are performed, specifically user identity privacy alignment, cross-platform sequence concatenation, and time series standardization, resulting in the user's multi-platform behavior sequence. During user identity privacy alignment, privacy-preserving processing (i.e., privacy-preserving user identifier mapping) technology can be used to align cross-platform user identities. Specifically, a cryptographic hash function can be used for user identifier mapping. The cryptographic hash function processes the real user IDs from different platforms, mapping them to the same user ID. This involves a one-way hash calculation, converting the data into an irreversible, fixed-length string as the user's anonymous identifier (i.e., user ID), ensuring user privacy and security while achieving cross-platform behavior association. Subsequently, the behavior data of users with the same user ID on different platforms are arranged chronologically (i.e., cross-platform sequence concatenation) to generate a unified cross-platform behavior sequence, further yielding the user's multi-platform behavior sequence.

[0044] The specific steps for obtaining a user's multi-platform behavior sequence are as follows.

[0045] First, user behavior logs (i.e., behavior data) are collected from multiple platforms. Each user behavior log includes information such as user ID (processed for privacy protection), behavior type, behavior object ID, timestamp, and platform ID.

[0046] Then, the behavioral data for each platform is preprocessed, including sorting, categorizing, and arranging in ascending order by timestamp, to construct the user's behavioral sequence for each platform. The sorting and categorization process may include cleaning and denoising, data standardization, etc., which are existing mature technologies and will not be elaborated upon here.

[0047] Furthermore, by using privacy-preserving user identifier mapping technology (such as cryptographic hash functions), the behavioral data of the same user on different platforms can be linked together, and all the behavioral data of the user on multiple platforms can be arranged in ascending order by timestamp to form a unified cross-platform behavioral sequence.

[0048] Finally, the compiled cross-platform behavior sequences are normalized (i.e., time series standardization, which can use any normalization method, such as min-max normalization), sequence filling, and data augmentation preprocessing to eliminate noise information in the data and obtain the user's multi-platform behavior sequences.

[0049] For each user, construct a multi-platform behavior sequence in the following form:

[0050] ;

[0051] in, For multi-platform behavior sequences; For the first Individual behavioral data, , The sequence length; For the first Each action includes user ID, action type, and action object ID; For the first The timestamp of each action; For the first The platform ID of the platform where the action occurred.

[0052] For each candidate product, the characteristics of the candidate product are manually determined. For example, the category, packaging shape, color, size and other features of the candidate product are used as the characteristics of the candidate product. Then the characteristics of the candidate product are encoded. Specifically, the features of the candidate product are processed by feature embedding through an independent embedding layer to obtain the vector representation of the candidate product and obtain the product feature vector of the candidate product.

[0053] After obtaining multiple sample data (user-candidate product samples), a dataset can be constructed. ,in, For the first Multi-platform behavior sequences of users in a sample dataset For the first The product feature vector of candidate products in each sample data. For the first The actual click-through rate (ATR) of user clicks on candidate products in each sample dataset is the actual click percentage, which is equal to the number of times a user clicked on a product divided by the number of times the product appeared. , This represents the number of sample data in the dataset.

[0054] The aforementioned dataset is used to train the subsequently established hybrid neural network model, resulting in a recommendation model. Based on this model, the user interest vector and click-through rate for a particular product can be predicted.

[0055] (ii) Model building.

[0056] like Figure 3As shown, this embodiment constructs a hybrid neural network model, which includes a feature embedding module, a multi-dilated convolution module, a multi-level attention mechanism module, and a prediction module connected in sequence. The feature embedding module is used to process multi-platform behavior sequences to obtain feature embedding vectors. The multi-dilated convolution module is used to process the feature embedding vectors to obtain multi-scale feature vectors, which are intermediate representations of user interests obtained through feature extraction. The multi-level attention mechanism module is used to process the multi-scale feature vectors and product feature vectors to obtain user interest vectors. The prediction module is used to predict the user's predicted click-through rate for candidate products based on the user interest vectors and product feature vectors.

[0057] In this embodiment, the breadth, depth, and training method of the hybrid neural network model can be determined first. The breadth is the number of neurons / channels in each layer, the depth is the number of trainable hidden layers, and the training method includes the selection of loss function and optimizer.

[0058] (1) Feature embedding module.

[0059] The multi-platform behavior sequence is input into the feature embedding module, which extracts and embeds features into the multi-platform behavior sequence. Specifically, discrete behavior features are mapped into continuous vectors to obtain feature embedding vectors.

[0060] Specifically, such as Figure 4 As shown, the feature embedding module performs feature embedding processing on each element in the multi-platform behavior sequence, including behavior type embedding, behavior object ID embedding, time interval embedding, and platform ID embedding, to obtain the embedding vector of each element. After concatenating the embedding vectors of each element, the concatenated vectors are then fused into a unified vector representation through a fully connected layer or a multilayer perceptron to obtain the feature embedding vector.

[0061] In this embodiment, the feature embedding module includes a first feature embedding layer, a second feature embedding layer, a third feature embedding layer, a fourth feature embedding layer, a first splicing layer, and a first fusion layer.

[0062] The first feature embedding layer is used to perform feature embedding processing on the behavior types in the multi-platform behavior sequence to obtain the first embedding vector.

[0063] Behavior type embedding:

[0064] ;

[0065] in, This is the first embedding vector; For feature embedding layer; For behavior type; As the first dimension, It can be / 4, For custom dimensions. By embedding behavior types, behavior types are mapped to... / 4-dimensional first embedding vector.

[0066] The second feature embedding layer is used to perform feature embedding processing on the behavior object ID in the multi-platform behavior sequence to obtain the second embedding vector.

[0067] Behavior object ID embedding:

[0068] ;

[0069] in, This is the second embedding vector; For the behavior object ID; As the second dimension, It can be By embedding the behavior object ID, the behavior object ID is mapped to... The second embedding vector of dimension.

[0070] The third feature embedding layer is used to perform feature embedding processing on the time interval between two adjacent timestamps in the multi-platform behavior sequence to obtain the third embedding vector.

[0071] Time interval embedding:

[0072] ;

[0073] in, This is the third embedding vector; The time interval between two adjacent timestamps; As the third dimension, It can be / 4. By embedding time intervals, the time intervals of adjacent actions are mapped to... / 4-dimensional third embedding vector.

[0074] The fourth feature embedding layer is used to perform feature embedding processing on the platform ID in the multi-platform behavior sequence to obtain the fourth embedding vector.

[0075] Platform ID Embedded:

[0076] ;

[0077] in, This is the fourth embedding vector; Platform ID; As the fourth dimension, It can be / 2. By embedding the platform ID, the platform ID is mapped to... The fourth embedding vector in 2D.

[0078] The input of the first concatenation layer is connected to the outputs of the first feature embedding layer, the second feature embedding layer, the third feature embedding layer, and the fourth feature embedding layer. The first concatenation layer is used to concatenate the first embedding vector, the second embedding vector, the third embedding vector, and the fourth embedding vector to obtain the first concatenated vector.

[0079] The input of the first fusion layer is connected to the output of the first concatenation layer. The first fusion layer is used to process the first concatenated vector to obtain the feature embedding vector. The first fusion layer is either a fully connected layer or a first multilayer perceptron.

[0080] After concatenating all the embedded vectors, they are then fused using either a fully connected layer or a first-level multilayer perceptron.

[0081] ;

[0082] in, For feature embedding vectors; It is a fully connected layer; This is a splicing layer.

[0083] (2) Multiple dilated convolution module.

[0084] The feature embedding vector is input into the multi-dilated convolution module, which extracts multi-scale local patterns from the sequence to obtain multi-scale feature vectors.

[0085] like Figure 5 As shown, this multi-dilated convolution module includes multiple parallel dilated convolution branches with different dilation rates, used to extract multi-scale local patterns in the sequence, that is, to extract multi-scale local interest patterns from the behavioral sequence. For example, it can include three parallel dilated convolution branches with dilation rates of 1, 2, and 4, or four parallel dilated convolution branches with dilation rates of 1, 2, 4, and 8. Each dilated convolution branch includes a dilated convolution layer, a batch normalization layer, a GELU (Gaussian Error Linear Units) activation function layer, and a Dropout layer connected in sequence. The dilation rate refers to the dilation rate of the dilated convolution layer. The feature vectors output by each dilated convolution branch are concatenated along the feature dimension, and then fused through a 1×1 convolution layer and a normalization layer. Finally, the fused output vector is residually concatenated with the input feature embedding vector to obtain a multi-scale feature vector.

[0086] Specifically, the multi-dilated convolution module includes three parallel dilated convolution branches with dilation rates of 1, 2, and 4, which are used to capture behavioral dependencies at different intervals.

[0087] The specific structure of each dilated convolution branch is as follows.

[0088] 1) Dilated convolutional layer: One-dimensional dilated convolution with a kernel size of 3.

[0089] ;

[0090] in, This is the output of the dilated convolutional layer; It is a one-dimensional convolutional layer; This is the input to the dilated convolutional layer; The kernel size; Void ratio; The void ratio can take values ​​of 1, 2, and 4.

[0091] 2) Batch Normalization Layer: Stabilizing the Training Process

[0092] ;

[0093] in, This is the output of the batch normalization layer; It is a one-dimensional batch normalization layer.

[0094] 3) GELU activation function layer: using the GELU activation function

[0095] ;

[0096] in, The output of the GELU activation function layer; This is the GELU activation function layer.

[0097] 4) Dropout layer: Prevents overfitting

[0098] ;

[0099] in, This is the output of the Dropout layer, i.e., the feature vector output by the dilated convolution branch; For Dropout layer; This represents the discard rate.

[0100] The feature vectors output from the three dilated convolution branches are concatenated along the feature dimension, and then fused through a 1×1 convolutional layer and a normalization layer.

[0101] ;

[0102] in, This is the fused vector; For standardization layer; This is the feature vector output by the first dilated convolution branch; This is the feature vector output by the second dilated convolution branch; This is the feature vector output by the third dilated convolution branch.

[0103] The fused output vector is residually concatenated with the input feature embedding vector:

[0104] ;

[0105] in, These are multi-scale feature vectors; For feature embedding vectors; This is a learnable scaling parameter, with an initial value of 0.1.

[0106] In this embodiment, the multi-dilated convolution module includes multiple dilated convolution branches, a second splicing layer, a second fusion layer, and an output layer.

[0107] Each dilated convolution branch consists of a dilated convolutional layer, a batch normalization layer, a GELU activation function layer, and a Dropout layer connected in sequence. The dilation rate of the dilated convolutional layers in different dilated convolutional branches is different.

[0108] The input of the second concatenation layer is connected to the output of each dilated convolution branch. The second concatenation layer is used to concatenate the feature vectors output by each dilated convolution branch to obtain the second concatenated vector.

[0109] The input of the second fusion layer is connected to the output of the second splicing layer. The second fusion layer is used to process the second spliced ​​vector to obtain the fused vector. The second fusion layer includes a convolutional layer and a normalization layer connected in sequence.

[0110] The first input of the output layer is connected to the input of the multi-dilated convolution module, and the second input of the output layer is connected to the output of the second fusion layer. The output layer is used to perform a weighted summation of the feature embedding vector and the fused vector to obtain a multi-scale feature vector.

[0111] (3) Multi-level attention mechanism module.

[0112] The multi-scale feature vectors are input into the multi-level attention mechanism module, which integrates cross-platform information and models global dependencies to obtain the user interest vector.

[0113] A three-layer attention network is constructed in the multi-layer attention mechanism module. This three-layer attention network includes three levels of attention mechanisms: project-level attention, intra-platform self-attention, and inter-platform importance attention.

[0114] 1) Project-level attention:

[0115] Use the product feature vector of the candidate product as the query vector Using the multi-scale feature vectors of multi-platform behavioral sequences as key vectors Sum value vector It utilizes a multi-head attention mechanism to calculate attention, capture user interests related to the current candidate product, calculate attention weights, and further calculate item-level vectors, which are weighted representations of historical behaviors related to the candidate product.

[0116] ;

[0117] ;

[0118] in, Attention weights; For the Softmax function; Key vector The dimension; This is a project-level vector.

[0119] 2) Self-attention within the platform:

[0120] Feature extraction is performed on the behavior sequences within each platform to obtain platform feature vectors. Self-attention mechanism and attention pooling are applied respectively to model the temporal dependencies within the platform, and the attention calculation results of each platform are obtained. Then, they are concatenated to obtain the platform-specific vectors.

[0121] Specifically, for each platform, when applying self-attention mechanism and attention pooling, the user's behavioral data (belonging to a multi-platform behavioral sequence) is arranged in chronological order to obtain the platform's behavioral sequence. This behavioral sequence is then used as input to the feature embedding module. The feature embedding module, multiple dilated convolutional modules, and attention pooling layers are sequentially connected to extract features from the behavioral sequence, resulting in the platform's feature vector. This platform's feature vector is then used as the query vector. Key vector Sum value vector The calculation is performed using two formulas in project-level attention, and the result is... This is the initial attention calculation result of the platform. Then, the attention pooling layer is used to process the initial attention calculation result of the platform to obtain the final attention calculation result of the platform.

[0122] Attention pooling layers are commonly used network layers and can employ general query vector attention pooling.

[0123] 3) Inter-platform importance attention:

[0124] The inter-platform importance attention is generated through a learnable weighted network to dynamically evaluate the importance of behavioral data from different platforms. Based on the importance, the platform feature vectors of each platform are weighted and summed to achieve intelligent fusion of cross-platform information and obtain inter-platform vectors, which are cross-platform fusion representations.

[0125] like Figure 6 As shown, the platform feature vector and product feature vector of each platform are concatenated to obtain the joint vector of each platform. The joint vector of each platform is processed by a multilayer perceptron to obtain the initial weight of each platform. The initial weight of each platform is normalized by the Softmax function to obtain the platform importance weight of each platform. Based on the platform importance weight of each platform, the platform feature vector of each platform is weighted and summed to obtain the inter-platform vector.

[0126] The importance weight of a platform can also be set manually. For example, the importance weight of a platform can be determined manually by the user based on the statistical information of the behavioral data of each platform and the context information of the current recommendation task.

[0127] The output features of the three levels of attention mechanisms are weighted and fused. Specifically, the output features of the three levels of attention mechanisms are first concatenated, and then processed by a multilayer perceptron.

[0128] ;

[0129] in, User interest vector; It is a multilayer perceptron; This is a project-level vector; This is an in-platform vector; This is an inter-platform vector.

[0130] In this embodiment, the multi-layered attention mechanism module includes a project-level attention layer, an intra-platform self-attention layer, an inter-platform importance attention layer, and a third fusion layer.

[0131] The project-level attention layer is used to perform attention calculations using the product feature vector as the query vector and the multi-scale feature vector as the key vector and value vector, to obtain the project-level vector.

[0132] The self-attention layer within the platform is used to perform self-attention calculation and attention pooling on the platform feature vector of each platform to obtain the attention calculation result of each platform. The attention calculation results of each platform are then concatenated to obtain the platform vector. The platform feature vector is the feature vector obtained by extracting features from the platform's behavior sequence. The behavior sequence is the sequence data obtained by arranging the user's behavior data on the platform in chronological order from the behavior sequences of multiple platforms.

[0133] The inter-platform importance attention layer is used to concatenate the platform feature vector and the product feature vector of each platform to obtain the joint vector of each platform. The joint vector of each platform is then processed to obtain the platform importance weight of each platform. Based on the platform importance weight of each platform, the platform feature vector of each platform is weighted and summed to obtain the inter-platform vector.

[0134] The joint vector of each platform is processed separately to obtain the platform importance weight of each platform. Specifically, the joint vector of each platform is processed separately using the second multilayer perceptron to obtain the initial weight of each platform, and the initial weight of each platform is normalized using the Softmax function to obtain the platform importance weight of each platform.

[0135] The input of the third fusion layer is connected to the output of the project-level attention layer, the output of the platform-intra-platform self-attention layer, and the output of the inter-platform importance attention layer. The third fusion layer is used to fuse the project-level vector, the intra-platform vector, and the inter-platform vector to obtain the user interest vector.

[0136] The user interest vector is obtained by fusing the project-level vector, the intra-platform vector, and the inter-platform vector. Specifically, the project-level vector, the intra-platform vector, and the inter-platform vector are concatenated using a third concatenation layer to obtain a third concatenated vector. The third multilayer perceptron is then used to process the third concatenated vector to obtain the user interest vector.

[0137] (4) Prediction module.

[0138] After designing the aforementioned three modules, and obtaining the user's multi-platform behavior sequence and the product feature vectors of candidate products, the feature embedding module converts the multi-platform behavior sequence into a vector representation to obtain the feature embedding vector. Based on the feature embedding vector, the multi-scale local patterns in the sequence are extracted through the multi-dilated convolution module to obtain the multi-scale feature vector. Based on the multi-scale feature vector and the product feature vector, the multi-level attention mechanism module generates the user interest vector, which is then used for downstream recommendation tasks. At this point, the user interest vector, product feature vector, and context feature vector interact. Specifically, the prediction module is designed to perform feature fusion and click-through rate prediction. The user interest vector, product feature vector, and context feature vector are concatenated, and the probability of the user clicking on the candidate product is predicted using a multilayer perceptron and a sigmoid activation function to obtain the predicted click-through rate of the user on the candidate product.

[0139] The formula for predicting click-through rate is:

[0140] ;

[0141] in, To predict click-through rate; Use the Sigmoid activation function; For product feature vectors; This is the context feature vector.

[0142] The context feature vector is determined by encoding context features such as time (e.g., timestamp), location (e.g., delivery address), and device (e.g., the user's device) in the multi-platform behavior sequence into vectors using a multilayer perceptron.

[0143] In this embodiment, the prediction module includes a fourth splicing layer, a fourth multilayer perceptron, and a Sigmoid activation function layer connected in sequence.

[0144] The fourth concatenation layer is used to concatenate the user interest vector, product feature vector, and context feature vector to obtain the fourth concatenated vector. The context feature vector is the feature vector obtained by extracting features from the timestamp, location, and device in the multi-platform behavior sequence.

[0145] The fourth multilayer perceptron is used to process the fourth concatenated vector to obtain the prediction vector.

[0146] The Sigmoid activation function layer is used to process the prediction vector to obtain the predicted click-through rate of the user for the candidate products.

[0147] (III) Model training.

[0148] This embodiment uses a dataset to train a hybrid neural network model to obtain a recommendation model.

[0149] For each sample in the dataset, the multi-platform behavior sequence and product feature vector from the sample data are input into the hybrid neural network model to obtain the predicted click-through rate (CTR) of users for candidate products. The predicted CTR of users for candidate products and the actual CTR of users for candidate products in the sample data are used to calculate the loss, resulting in the total loss. The hybrid neural network model is then iteratively updated using the total loss to obtain the recommendation model. After obtaining the recommendation model, the multi-platform behavior sequence of the user to be recommended and the product feature vector of the product to be recommended are input into the recommendation model to obtain the predicted CTR of the user to be recommended for the product to be recommended. A personalized recommendation list can then be generated to complete the recommendation process.

[0150] The hybrid neural network model is trained using a dataset to obtain a recommendation model, which includes the following steps.

[0151] (1) Randomly select multiple sample data from the dataset as training sample data.

[0152] (2) For each training sample data, the multi-platform behavior sequence and product feature vector in the training sample data are used as inputs, and the hybrid neural network model is used to determine the user's predicted click rate for the candidate product.

[0153] (3) The total loss is calculated based on the actual and predicted click-through rates of users for candidate products.

[0154] The total loss is the click-through rate prediction loss.

[0155] (4) Update the hybrid neural network model using the total loss to obtain the updated model.

[0156] (5) Determine whether the iteration termination condition has been met; if yes, use the updated model as the recommendation model; if no, use the updated model as the hybrid neural network model for the next iteration and return to the step of "randomly select multiple sample data in the dataset as training sample data".

[0157] The iteration termination condition can be reaching the maximum number of iterations or the total loss being less than a preset value.

[0158] This embodiment can adopt a multi-task learning framework for joint training of multiple tasks, and simultaneously optimize the main task (click rate prediction task) and multiple auxiliary tasks (next platform prediction, next behavior prediction and contrastive learning tasks, etc.) to improve the generalization ability of the model.

[0159] At this point, the total loss is the weighted sum of the click-through rate prediction loss, the next platform prediction loss, the next action prediction loss, and the contrastive learning loss. The click-through rate prediction loss is calculated based on the user's actual click-through rate and predicted click-through rate for candidate products. The next platform prediction loss is calculated based on the actual value and predicted value of the platform ID of the next platform used by the user. The next action prediction loss is calculated based on the actual value and predicted value of the action type of the user's next action.

[0160] When predicting the next platform, a first mean pooling layer and a next platform prediction head are added after the feature embedding module and the multiple dilated convolution module. The first mean pooling layer is used to perform mean pooling on the multi-scale feature vector output by the multiple dilated convolution module to obtain the first global sequence representation vector. The next platform prediction head can use multiple multilayer perceptrons connected in sequence and a Softmax function to output a first probability vector. The first probability vector includes the probability of each platform being the next platform. The platform ID of the platform with the highest probability is selected as the predicted value of the platform ID of the next platform used by the user.

[0161] Similarly, when predicting the next action, a second average pooling layer and a next action prediction head are added after the feature embedding module and the multi-dilated convolution module. The second average pooling layer is used to perform average pooling on the multi-scale feature vector output by the multi-dilated convolution module to obtain the second global sequence representation vector. The next action prediction head can use multiple multilayer perceptrons connected in sequence and a Softmax function to output a second probability vector. The second probability vector includes the probability of each action as the next action. The action type with the highest probability is selected as the predicted value of the action type of the user's next action.

[0162] During training, after calculating the total loss, backpropagation based on the total loss will simultaneously update the network parameters of the hybrid neural network model, the first mean pooling layer, the next platform prediction head, the second mean pooling layer, and the next row prediction head. After training is completed, the trained hybrid neural network model is selected as the recommendation model.

[0163] The formula for calculating the total loss is:

[0164] ;

[0165] in, Total loss; Weighting parameters for click-through rate prediction loss; Predicting losses based on click-through rate; Weighting parameters for predicting loss on the next platform; Predict losses for the next platform; The weighting parameters for predicting the loss are used for the next row; Predict the loss for the next step; The weight parameters are used to compare the learning loss. To compare learning loss. They can be set to 1.0, 0.3, 0.2, and 0.1 respectively.

[0166] Click-through rate prediction loss can be achieved using binary classification cross-entropy:

[0167] ;

[0168] in, This refers to the batch size, which is the number of training sample data inputs in one training session. For the first The actual click-through rate of candidate products by users corresponding to the training sample data; For the first The predicted click-through rate of candidate products by users corresponding to each training sample data.

[0169] The next platform predicts loss using multi-class cross-entropy:

[0170] ;

[0171] in, For the number of platforms; For the first The actual value of the platform ID of the next platform used by the user corresponding to each training sample data is The probability, if the true value is If the probability is 1, then the probability is 1; otherwise, the probability is 0. For the first The predicted value of the platform ID of the next platform used by the user corresponding to each training sample data is The probability of.

[0172] The next line uses multi-class cross-entropy to predict the loss:

[0173] ;

[0174] in, Number of behavior types; For the first The true value of the next action type of the user corresponding to the training sample data is... The probability, if the true value is If the probability is 1, then the probability is 1; otherwise, the probability is 0. For the first The predicted behavior type of the user's next action corresponding to each training sample data is: The probability of.

[0175] The contrastive learning loss uses InfoNCE loss (Information Noise-Contrastive Estimation):

[0176] ;

[0177] in, The cosine similarity function; For the first The query vector corresponding to each training sample data; For the first A query vector for positive sample data corresponding to each training sample data; For temperature parameters; For the first A query vector for positive sample data corresponding to each training sample data; For indicator functions, when satisfying ,but If it is 1, otherwise, =0; For the first A query vector for the negative sample data corresponding to the training sample data.

[0178] The query vector can be a user interest vector, or it can be a vector obtained by processing the user interest vector. For example, a multilayer perceptron can be used to perform a nonlinear projection transformation on the user interest vector to obtain a transformed vector, and then the transformed vector can be normalized to obtain the query vector.

[0179] When positive sample data is obtained, data augmentation is performed on the multi-behavior user sequences in the sample data. Data augmentation methods can include random masking (randomly discarding a certain proportion (e.g., 20%) of the behavior data in the multi-behavior user sequence), random rearrangement (shuffling the order of behavior data within a local window (e.g., 3 behavior data) in the multi-behavior user sequence), and feature discarding (randomly discarding certain fields (e.g., platform ID) in certain behavior data in the multi-behavior user sequence).

[0180] When negative sample data is obtained, other sample data (different from the sample data) within the same batch are selected as negative sample data.

[0181] When updating the hybrid neural network model using the total loss, use Optimizer, learning rate set to The cosine annealing learning rate scheduling strategy is adopted, which dynamically adjusts the learning rate of each parameter by using the first moment estimation and second moment estimation of the gradient. After bias correction, the learning rate of each iteration has a certain range, which makes the parameter update stable.

[0182] After the iteration termination condition is met, the training ends and the optimal recommendation model is obtained.

[0183] This embodiment can also be retrained based on user feedback to update the network parameters of the recommendation model.

[0184] Example 2.

[0185] The recommendation model application method based on multi-platform behavior provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown depicts a scenario where the terminal communicates with the server via a network. A data storage system stores the data the server needs to process. This system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send application requests to the server. Upon receiving the request, the server obtains the multi-platform behavior sequence of the user to be recommended and the product feature vector of each product to be recommended. For each product, using the multi-platform behavior sequence of the user to be recommended and the product feature vector of the product as input, the server uses a recommendation model to determine the predicted click-through rate (CTR) of the product to be recommended. The server then sorts each product according to its predicted CTR from highest to lowest, obtaining the sorting result, and recommends it to the user based on this result. The server can then feed back the sorting result (the application result for the application request) to the terminal.

[0186] Furthermore, in some embodiments, the recommendation model application method based on multi-platform behavior can also be implemented by the server or the terminal alone. For example, the terminal can directly process the application request to be processed, or the server can obtain the application request to be processed from the data storage system and process it.

[0187] In one exemplary embodiment, such as Figure 7 As shown, a recommendation model application method based on multi-platform behavior is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps are used as an example of a server in the example.

[0188] Step T1: Obtain the multi-platform behavior sequence of the user to be recommended and the product feature vector of each product to be recommended.

[0189] Step T2: For each product to be recommended, the multi-platform behavior sequence of the user to be recommended and the product feature vector of the product to be recommended are used as inputs to determine the predicted click-through rate of the user to be recommended for the product to be recommended using the recommendation model; the recommendation model is a model trained using the above-mentioned multi-platform behavior-based recommendation model training method.

[0190] Step T3: Sort each product to be recommended in descending order of predicted click-through rate to obtain the sorting result, and recommend it to the user to be recommended based on the sorting result.

[0191] This embodiment deploys a recommendation model for online personalized recommendations, processing user requests in real time and generating personalized recommendation results. Specifically, when a user visits, the multi-platform behavior sequence of the user is acquired in real time. The multi-platform behavior sequence of the user to be recommended and the product feature vector of the product to be recommended are input into the recommendation model to obtain the predicted click-through rate of the user for the product to be recommended. Each product to be recommended is sorted in descending order of predicted click-through rate to generate a personalized recommendation list (i.e., the sorting result), which includes the predicted click-through rate of the user for each recommended product.

[0192] Of course, after sorting, this embodiment can also be filtered in conjunction with business rules to generate a personalized recommendation list, and finally recommend the top-ranked products to the users to be recommended.

[0193] This embodiment supports both real-time inference and batch inference modes to meet the needs of different business scenarios.

[0194] This embodiment supports A / B testing and hot model updates to ensure continuous optimization of recommendation performance.

[0195] This embodiment relates to artificial intelligence, deep learning, and recommendation technologies. Specifically, it describes a method for achieving accurate personalized recommendations by utilizing cross-platform user behavior data and employing a hybrid neural network model that integrates dilated convolution and attention mechanisms. This method is a user interest modeling and recommendation approach based on multi-platform behavior sequences and a hybrid neural network model. First, it acquires user behavior data from multiple platforms, processes it for privacy protection, and aligns it chronologically to generate multi-platform behavior sequences. Then, it uses a feature embedding module to embed features into these sequences, converting them into vector representations to obtain feature embedding vectors. Next, it uses a multi-dilated convolution module to extract multi-scale local patterns from the sequences, obtaining multi-scale feature vectors. Then, it uses a multi-level attention mechanism module for sequence modeling, including item-level attention, intra-platform self-attention, and inter-platform importance attention, to obtain user interest vectors. Finally, it fuses user interest vectors, product feature vectors, and contextual feature vectors to predict click-through rates (CTRs) for candidate products. This method effectively integrates cross-platform information, captures multi-scale patterns of user interests, and significantly improves recommendation accuracy and user experience.

[0196] This embodiment has the following advantages compared to traditional technologies.

[0197] (1) Panoramic User Interest Modeling: By integrating behavioral data from multiple platforms, a more comprehensive and accurate user interest profile was constructed, breaking through the data limitations of a single platform.

[0198] (2) Multi-scale pattern capture: The multi-dilated convolution module can capture local neighbor patterns and long-distance dependencies in the behavior sequence at the same time, which improves the ability of sequence modeling.

[0199] (3) Intelligent cross-platform integration: The inter-platform importance attention mechanism can dynamically assess the importance of data from different platforms, realize the intelligent integration of cross-platform information, and avoid the noise problem caused by simple splicing.

[0200] (4) End-to-end joint optimization: The multi-task learning framework enables the model to learn useful representations from auxiliary tasks, improving the performance of the main task and the generalization ability of the model.

[0201] (5) Improved actual business performance: A / B testing in actual e-commerce recommendation scenarios shows that compared with the baseline model, the click-through rate is increased by 15%-25%, the conversion rate is increased by 10%-20%, and the user experience is significantly improved.

[0202] (6) Good scalability: The model architecture is flexible and can be easily extended to more platforms and more complex behavior types to meet the needs of different business scenarios.

[0203] This application also provides an application scenario in which the above-described recommendation model application method based on multi-platform behavior is applied. Specifically, the recommendation model application method based on multi-platform behavior provided in this embodiment can be applied to a product recommendation scenario. The product recommendation scenario includes a prediction stage and a recommendation stage. The prediction stage is used to determine the predicted click-through rate for each product and further generate ranking results. The recommendation stage is used to recommend products to the user based on the ranking results. The recommendation model application method based on multi-platform behavior provided in this embodiment belongs to the prediction stage.

[0204] Example 3.

[0205] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When executed by the processor, the computer program implements either a recommendation model training method based on multi-platform behavior or a recommendation model application method based on multi-platform behavior.

[0206] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0207] In one exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the recommendation model training method based on multi-platform behavior in Embodiment 1 or the recommendation model application method based on multi-platform behavior in Embodiment 2.

[0208] Example 4.

[0209] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the recommendation model training method based on multi-platform behavior in Embodiment 1 or the recommendation model application method based on multi-platform behavior in Embodiment 2.

[0210] 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, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0211] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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 specification.

[0212] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for training a recommendation model based on multi-platform behavior, characterized in that, The method for training the recommendation model based on multi-platform behavior includes: Obtain the dataset; the dataset includes multiple sample data, the sample data includes user multi-platform behavior sequences, product feature vectors of candidate products, and user actual click-through rates on candidate products, the multi-platform behavior sequences are sequence data obtained by arranging user behavior data on multiple platforms in chronological order; A hybrid neural network model is constructed. The hybrid neural network model includes a feature embedding module, a multi-dilated convolution module, a multi-level attention mechanism module, and a prediction module connected in sequence. The feature embedding module is used to process the multi-platform behavior sequence to obtain a feature embedding vector. The multi-dilated convolution module is used to process the feature embedding vector to obtain a multi-scale feature vector. The multi-level attention mechanism module is used to process the multi-scale feature vector and the product feature vector to obtain a user interest vector. The prediction module is used to predict the user's click-through rate for candidate products based on the user interest vector and the product feature vector. The hybrid neural network model is trained using the dataset to obtain a recommendation model.

2. The method for training a recommendation model based on multi-platform behavior according to claim 1, characterized in that, The behavioral data includes user ID, behavior type, behavior object ID, timestamp, and platform ID. The user ID is the ID obtained after privacy protection processing of the real user ID. At this time, the feature embedding module includes a first feature embedding layer, a second feature embedding layer, a third feature embedding layer, a fourth feature embedding layer, a first splicing layer, and a first fusion layer. The first feature embedding layer is used to perform feature embedding processing on the behavior types in the multi-platform behavior sequence to obtain a first embedding vector; The second feature embedding layer is used to perform feature embedding processing on the behavior object ID in the multi-platform behavior sequence to obtain a second embedding vector; The third feature embedding layer is used to perform feature embedding processing on the time interval between two adjacent timestamps in the multi-platform behavior sequence to obtain a third embedding vector. The fourth feature embedding layer is used to perform feature embedding processing on the platform ID in the multi-platform behavior sequence to obtain the fourth embedding vector; The input of the first splicing layer is connected to the output of the first feature embedding layer, the output of the second feature embedding layer, the output of the third feature embedding layer, and the output of the fourth feature embedding layer; the first splicing layer is used to splice the first embedding vector, the second embedding vector, the third embedding vector, and the fourth embedding vector to obtain the first spliced ​​vector; The input of the first fusion layer is connected to the output of the first splicing layer; the first fusion layer is used to process the first spliced ​​vector to obtain a feature embedding vector; wherein, the first fusion layer is a fully connected layer or a first multilayer perceptron.

3. The method for training a recommendation model based on multi-platform behavior according to claim 1, characterized in that, The multi-dilated convolution module includes multiple dilated convolution branches, a second splicing layer, a second fusion layer, and an output layer; Each of the dilated convolution branches includes a dilated convolution layer, a batch normalization layer, a GELU activation function layer, and a Dropout layer connected in sequence; the dilation rate of the dilated convolution layers in different dilated convolution branches is different; The input of the second concatenation layer is connected to the output of each of the dilated convolution branches; the second concatenation layer is used to concatenate the feature vectors output by each of the dilated convolution branches to obtain a second concatenated vector. The input of the second fusion layer is connected to the output of the second splicing layer; the second fusion layer is used to process the second spliced ​​vector to obtain the fused vector; wherein, the second fusion layer includes a convolutional layer and a normalization layer connected in sequence; The first input of the output layer is connected to the input of the multi-dilated convolution module, and the second input of the output layer is connected to the output of the second fusion layer; the output layer is used to perform a weighted summation of the feature embedding vector and the fused vector to obtain a multi-scale feature vector.

4. The method for training a recommendation model based on multi-platform behavior according to claim 1, characterized in that, The multi-level attention mechanism module includes a project-level attention layer, an intra-platform self-attention layer, an inter-platform importance attention layer, and a third fusion layer; The project-level attention layer is used to perform attention calculations using the product feature vector as the query vector and the multi-scale feature vector as the key vector and value vector to obtain the project-level vector. The self-attention layer within the platform is used to perform self-attention calculation and attention pooling on the platform feature vector of each platform to obtain the attention calculation result of each platform, and then concatenates the attention calculation results of each platform to obtain the platform vector; the platform feature vector is the feature vector obtained by feature extraction of the platform behavior sequence, and the behavior sequence is the sequence data obtained by arranging the user's behavior data on the platform in chronological order in the multi-platform behavior sequence; The inter-platform importance attention layer is used to concatenate the platform feature vector of each platform with the product feature vector to obtain the joint vector of each platform. The joint vector of each platform is processed to obtain the platform importance weight of each platform. Based on the platform importance weight of each platform, the platform feature vector of each platform is weighted and summed to obtain the inter-platform vector. The input of the third fusion layer is connected to the output of the project-level attention layer, the output of the platform-intra-platform self-attention layer, and the output of the inter-platform importance attention layer, respectively. The third fusion layer is used to fuse the project-level vector, the platform-intra-platform vector, and the inter-platform vector to obtain the user interest vector.

5. The method for training a recommendation model based on multi-platform behavior according to claim 4, characterized in that, The joint vector of each platform is processed separately to obtain the platform importance weight of each platform. Specifically, this includes: processing the joint vector of each platform using a second multilayer perceptron to obtain the initial weight of each platform; and normalizing the initial weight of each platform using the Softmax function to obtain the platform importance weight of each platform. The project-level vector, the intra-platform vector, and the inter-platform vector are fused to obtain the user interest vector. Specifically, this includes: using a third concatenation layer to concatenate the project-level vector, the intra-platform vector, and the inter-platform vector to obtain a third concatenated vector; and using a third multilayer perceptron to process the third concatenated vector to obtain the user interest vector.

6. The method for training a recommendation model based on multi-platform behavior according to claim 1, characterized in that, The prediction module includes a fourth splicing layer, a fourth multilayer perceptron, and a Sigmoid activation function layer connected in sequence. The fourth concatenation layer is used to concatenate the user interest vector, the product feature vector, and the context feature vector to obtain the fourth concatenated vector; the context feature vector is a feature vector obtained by extracting features from the timestamp, location, and device in the multi-platform behavior sequence; The fourth multilayer perceptron is used to process the fourth concatenated vector to obtain a prediction vector; The Sigmoid activation function layer is used to process the prediction vector to obtain the predicted click-through rate of the user for the candidate product.

7. The method for training a recommendation model based on multi-platform behavior according to claim 1, characterized in that, The hybrid neural network model is trained using the dataset to obtain a recommendation model, specifically including: Multiple sample data are randomly selected from the dataset as training sample data; For each training sample data, the multi-platform behavior sequence and product feature vector in the training sample data are used as inputs, and the hybrid neural network model is used to determine the user's predicted click-through rate for candidate products; The total loss is calculated based on the actual and predicted click-through rates of users for candidate products; The hybrid neural network model is updated using the total loss to obtain the updated model; Determine whether the iteration termination condition has been met; if yes, use the updated model as the recommendation model; if no, use the updated model as the hybrid neural network model for the next iteration, and return to the step of "randomly selecting multiple sample data from the dataset as training sample data".

8. A method for applying a recommendation model based on multi-platform behavior, characterized in that, The recommendation model application method based on multi-platform behavior includes: Obtain the multi-platform behavior sequence of the user to be recommended and the product feature vector of each product to be recommended; For each product to be recommended, the multi-platform behavior sequence of the user to be recommended and the product feature vector of the product to be recommended are used as inputs, and the predicted click-through rate of the user to be recommended for the product to be recommended is determined by the recommendation model; the recommendation model is a model trained using the recommendation model training method based on multi-platform behavior as described in any one of claims 1-7. Each product to be recommended is sorted in descending order of predicted click-through rate to obtain the sorting results, and recommendations are made to the users to be recommended based on the sorting results.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the recommendation model training method based on multi-platform behavior as described in any one of claims 1-7 or the recommendation model application method based on multi-platform behavior as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the recommendation model training method based on multi-platform behavior as described in any one of claims 1-7 or the recommendation model application method based on multi-platform behavior as described in claim 8.

Images