Two-channel graph convolution contrast learning recommendation method and system fusing multi-dimensional scoring

Abstract

The invention discloses a two-channel graph convolution contrast learning recommendation method and system fused with multi-dimensional scoring, and belongs to the technical field of deep learning. The multi-dimensional scoring system is a new scoring system which is developed on each large online platform in recent years, comprises total scores of the user on the items and scores of the user on the items on different attributes, and can comprehensively and meticulously reflect user preferences. However, most of the existing personalized recommendation algorithms are oriented to a single-dimensional scoring system only containing a total score, and ignore fine-grained preferences of a user in different dimensions and a complex relationship between the fine-grained preferences and the total score. On the basis, a two-channel graph convolution contrast learning recommendation method fusing multi-dimensional scores is provided, a graph convolution neural network and an attention mechanism are comprehensively utilized to learn coarse and fine granularity characterization of users and items, and a high-order complex relation between each attribute and a total score is modeled through two-channel contrast learning. While the user attribute preference and the representation vector are optimized, the introduction of external noise is effectively avoided, and a high-quality personalized recommendation result is provided for the user.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a dual-channel graph convolution comparative learning recommendation method and system integrating multi-dimensional scoring. Background Art

[0002] With the continuous development of the Internet industry, users have an increasing demand for personalized services, and recommendation systems have become a research direction of common concern in the business and academic fields. Most commonly used personalized recommendation algorithms learn user preferences based on historical user-item interactions or total scores. In fact, there are a large number of multidimensional rating systems in real life, which include both the user's total score for the project and the user's score on different attributes of the project, which can more comprehensively reflect the user's subjective evaluation, thereby providing more detailed feedback information and helping to accurately model user preferences.

[0003] At present, there are relatively few personalized recommendation algorithms for multidimensional rating systems. Existing studies can be roughly divided into the following two categories: one is to use matrix decomposition methods to obtain the representation vectors of users and items, and determine the attribute weights by calculating the Euclidean distance between the total score and the scores of different attributes; the other is to input the total score and the scores of each attribute into the neural network to learn the relationship between the attributes, and then obtain the representation vectors of users and items. However, most existing studies only mine user preferences from the perspective of user-item interaction, ignoring the association between users and failing to fully model user representation. At the same time, the attribute weight calculation method used is too simple to capture the complex relationship between the attribute score and the total score, and it is impossible to obtain the user's fine-grained attribute preferences, which limits the performance of the recommendation system. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems and propose a dual-channel graph convolutional contrastive learning recommendation method and system that integrates multi-dimensional ratings. The user preferences are jointly modeled from multiple perspectives in the two channels of total rating and multi-attribute rating, and the coarse-grained representation of the user under the total rating and the fine-grained representation under each attribute are obtained, and then the user's attribute preferences are calculated using the cosine similarity measure. On this basis, dual-channel contrastive learning is used to optimize the user's attribute preferences and their representation vectors, and the user's comprehensive representation is obtained through weighted fusion and splicing operations. Finally, the predicted score is obtained by performing an inner product operation on the user's comprehensive representation and the item representation, thereby generating a personalized recommendation list.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A dual-channel graph convolutional contrastive learning recommendation method integrating multi-dimensional ratings includes the following steps:

[0007] Step 1: By calculating the cosine similarity between different user representations, we get the user-user association matrix. Based on the user-item total score matrix and the user-user association matrix, we construct the user-item score graph and the user-user association graph. Then, we use the graph convolutional neural network and attention mechanism to jointly model the coarse-grained representation of users and items under the total score channel.

[0008] Step 2: Based on the attribute rating matrix, model the user preferences and item features from the user-item rating perspective and the user-user association perspective, obtain the fine-grained representation of users and items under the multi-attribute channel, and obtain the multi-attribute fusion representation of the item through average pooling;

[0009] Step 3: By calculating the cosine similarity between the user's total score and the representation of each attribute, the attribute weight is preliminarily determined, and then dual-channel contrastive learning is used to capture the complex relationship between the user's coarse-grained and fine-grained representations. While optimizing the user's attribute preferences, the user's coarse-grained representation under the total score and the multi-attribute fusion representation are enhanced;

[0010] Step 4: respectively concatenate the coarse-grained representation and multi-attribute fusion representation of the user and item under the total score to obtain the final representation of the user and item, input the two into the multi-layer perceptron for linear transformation, and obtain the predicted score of the item through inner product operation;

[0011] Step 5: By maximizing the consistency of the representation of the same user under the two channels and the difference of the representation of different users under the two channels, a dual-channel contrast loss function is designed, and the total loss function is formed by combining the mean square error loss function, and then the method is optimized by multi-task joint learning.

[0012] Furthermore, the implementation method of step 1 is as follows:

[0013] We learn the coarse-grained representation of users in the total rating channel from the perspective of user-item interaction and user-user association respectively;

[0014] From the user-item interaction perspective, is the user's initial embedding, is the initial embedding of the project. First, based on the total score matrix R n×m Construct a user-item rating graph, and use a lightweight graph convolutional neural network to aggregate node information and update the representation of each node to obtain the user representation of the lth layer. and project characterization The specific formula is as follows:

[0015]

[0016] Among them, N irepresents the set of users who interact with the project, N u Represents a collection of items that the user has interacted with.

[0017] On this basis, the node representations of layer L are averaged and pooled to obtain the user representation e from the perspective of user-item interaction. u and project characterization i The specific formula is as follows:

[0018]

[0019] From the user-user association perspective, let user u h The total score vector for all items is User v The total score vector for all items is By calculating user u h and user u v The cosine similarity sim(u h ,u v ), we can get the user-user association matrix S under the total score n×n The specific formula is as follows:

[0020]

[0021] By setting the similarity threshold γ, similar users of the target user are found, thereby constructing a user-user association graph. Considering that the scale of the user-user association graph is much smaller than the user-item rating graph, a graph convolutional neural network is used to aggregate the neighbor node information in the user-user association graph layer by layer to obtain the user representation of the lth layer. The specific formula is as follows:

[0022]

[0023] in, is the adjacency matrix of the user-user association graph, represents the degree matrix, W represents the learnable weight matrix, and σ represents the nonlinear activation function.

[0024] Average pooling is performed on the user representations of the L layer to obtain a coarse-grained representation of the user from the perspective of user-user association.

[0025]

[0026] After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the coarse-grained representation of the user under the total score channel. The specific formula is as follows:

[0027]

[0028] Among them, ω u and is a trainable parameter matrix.

[0029] Furthermore, the implementation method of step 2 is as follows:

[0030] Learn fine-grained representations of users under multi-attribute channels from the perspective of user-item interaction and user-user association respectively;

[0031] From the perspective of user-item interaction, the rating matrix based on each attribute Construct the corresponding user-item rating graph, and use the lightweight graph neural network to aggregate node information, update the representation of each node, and obtain the user representation of the lth layer and project characterization Then, the user representation under each attribute is obtained through average pooling and project characterization The specific formula is as follows:

[0032]

[0033] In getting the attribute a k Fine-grained representation of projects under Finally, the average pooling strategy is used to fuse the fine-grained representations of the project under multiple attributes to obtain the project representation after multi-attribute fusion.

[0034]

[0035] Where P represents the number of attributes.

[0036] From the user-user association perspective, let user u h For all items in attribute a k The rating vector under User v For all items in attribute a k The rating vector under Computed property a k Next user u h and user u v Cosine similarity between Get attribute a k The user-user association matrix under The specific formula is as follows:

[0037]

[0038] Based on the attribute a k The user-user association matrix under Build property a k The user-user association graph under , and the graph convolutional neural network is used to aggregate the neighbor node information layer by layer, so as to obtain the user representation of the lth layer under the attribute The specific formula is as follows:

[0039]

[0040] in, For attribute a k The adjacency matrix of the user-user association graph below.

[0041] Perform average pooling on the user representations of the L layer to obtain a fine-grained representation of the user from the perspective of user-user association

[0042]

[0043] After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the fine-grained representation of the user under the multi-attribute channel. The specific formula is as follows:

[0044]

[0045] in, and is a trainable parameter matrix.

[0046] Furthermore, the implementation method of step 3 is as follows:

[0047] In order to capture the relationship between each attribute and the total score and realize the effective fusion of the user's fine-grained representation, it is necessary to obtain the user's attribute preferences. First, the cosine similarity between the user's representation under each attribute and its representation under the total score is calculated. Determine the initial weights of the attributes Then, the user representation after multi-attribute fusion is obtained by weighted fusion. The specific formula is as follows:

[0048]

[0049] In order to further optimize the user's attribute preferences and representation vectors, a dual-channel contrast loss function Loss is designed. CL , maximize the consistency of the representation of the same user under two channels and the difference of the representation of different users under two channels. The specific formula is as follows:

[0050]

[0051] Where τ is the temperature coefficient.

[0052] Furthermore, the implementation method of generating the prediction score in step 4 is as follows:

[0053] The representations of the user and item under the total score are concatenated with the representations after multi-attribute fusion to obtain the final representation of the user. and the final representation of the project The specific formula is as follows:

[0054]

[0055] Use multi-layer perceptron (MLP) to finally characterize the user and final characterization of the project Perform linear transformation and obtain the final prediction score through inner product operation The specific formula is as follows:

[0056]

[0057] The specific optimization method in step 5 is as follows:

[0058] After getting the prediction score, use the mean square error loss function Loss MSE And contrast loss function Loss CL The mixed total loss function Loss optimizes the method;

[0059] Among them, the mean square error loss is used to measure the average difference between the predicted value and the true value; the contrast loss is used to enhance the user's attribute preference and representation vector. The method is jointly optimized through multi-task learning to obtain more personalized recommendation results. The specific formula is as follows:

[0060]

[0061] Loss=Loss MSE +λLoss CL +||Θ|| 2 (30)

[0062] Among them, I represents the project set, y ui represents the actual rating of item i by user u in the training set, represents the predicted score of user u for item i, and Θ represents the set of hyperparameters in the algorithm.

[0063] A dual-channel graph convolutional contrastive learning recommendation system integrating multi-dimensional ratings, including the following modules:

[0064] Total score channel representation learning module: learn the coarse-grained representation of users and items under the total score. First, calculate the cosine similarity sim between different users to obtain the user-user association matrix S n×n, and then based on the user-item total score matrix R n×m and the user-user association matrix S n×n , construct the user-item rating graph and the user-user association graph, and then use the graph convolutional neural network and attention mechanism to jointly model the coarse-grained representation of users under this channel and the coarse-grained representation of the project i ;

[0065] Multi-attribute channel representation learning module: Learn the fine-grained representation of users and projects under each attribute. Similar to the total score channel, based on each attribute score matrix, model user preferences and project features from the user-project rating perspective and the user-user association perspective, and obtain the fine-grained representation of users and projects under each attribute. and The average pooling strategy is used to fuse the fine-grained representations of the project under multiple attributes to obtain the multi-attribute fusion representation of the project.

[0066] Attribute preference optimization module: Enhance the user's representation by updating the attribute preference. Use the cosine similarity between the user's total score and the representation under each attribute. Preliminary determination of attribute weights Then, the complex relationship between the coarse-grained and fine-grained representations of users is captured through dual-channel contrastive learning. While optimizing user attribute preferences, the coarse-grained representation and multi-attribute fusion representation of users under the total score are enhanced.

[0067] Prediction score module: concatenates the coarse-grained representation and multi-attribute fusion representation of users and items under the total score to obtain the final representation of users and items and The two are input into the multi-layer perceptron for linear transformation, and the prediction score is obtained through inner product operation;

[0068] Method optimization module: Use multi-task joint learning method to optimize the total loss function obtained by mixing the mean square error loss function and the contrastive learning loss function;

[0069] Recommendation list generation module: The system generates a corresponding item recommendation list for the user to be recommended based on the final representation of the user to be recommended.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] 1. The present invention can more comprehensively and accurately model the individual preferences of users from multiple perspectives, provide users with personalized recommendation lists that better meet their needs, assist users in making quick choices and efficient decisions, and thus improve user satisfaction and experience;

[0072] 2. The present invention can be used in various network applications such as e-commerce, social media, online education, smart medical care, etc. It can provide users with decision-making references by integrating various scoring information, reduce user selection costs, and help improve the platform's product conversion rate and customer loyalty, and give full play to the economic and market value of the recommendation system in the field of multi-dimensional scoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a schematic diagram of the process of the method recommended by the present invention;

[0074] Figure 2 It is a schematic diagram of the framework of the method recommended by the present invention;

[0075] Figure 3 Schematic diagram of the structure of the system recommended by the present invention. DETAILED DESCRIPTION

[0076] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0077] like Figures 1 to 3 As shown, the dual-channel graph convolution contrastive learning recommendation method integrating multi-dimensional scoring described in this embodiment includes the following steps:

[0078] Step 1: By calculating the cosine similarity between different user representations, we get the user-user association matrix. Based on the user-item total score matrix and the user-user association matrix, we construct the user-item score graph and the user-user association graph. Then, we use the graph convolutional neural network and the attention mechanism to jointly model the coarse-grained representation of users and items under the total score channel. The implementation method is as follows:

[0079] We learn the coarse-grained representation of users in the total rating channel from the perspective of user-item interaction and user-user association respectively;

[0080] From the user-item interaction perspective, is the user's initial embedding, is the initial embedding of the project. First, based on the total score matrix R n×m Construct a user-item rating graph, and use a lightweight graph convolutional neural network to aggregate node information and update the representation of each node to obtain the user representation of the lth layer. and project characterization The specific formula is as follows:

[0081]

[0082] Among them, N i represents the set of users who interact with the project, N u Represents a collection of items that the user has interacted with.

[0083] On this basis, the node representations of layer L are averaged and pooled to obtain the user representation e from the perspective of user-item interaction. u and project characterization i The specific formula is as follows:

[0084]

[0085] From the user-user association perspective, let user u h The total score vector for all items is User v The total score vector for all items is By calculating user u h and user u v The cosine similarity sim(u h ,u v ), we can get the user-user association matrix S under the total score n×n The specific formula is as follows:

[0086]

[0087] By setting the similarity threshold γ, similar users of the target user are found, thereby constructing a user-user association graph. Considering that the scale of the user-user association graph is much smaller than the user-item rating graph, a graph convolutional neural network is used to aggregate the neighbor node information in the user-user association graph layer by layer to obtain the user representation of the lth layer. The specific formula is as follows:

[0088]

[0089] in, is the adjacency matrix of the user-user association graph, represents the degree matrix, W represents the learnable weight matrix, and σ represents the nonlinear activation function.

[0090] Average pooling is performed on the user representations of the L layer to obtain a coarse-grained representation of the user from the perspective of user-user association.

[0091]

[0092] After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the coarse-grained representation of the user under the total score channel. The specific formula is as follows:

[0093]

[0094] Among them, ω u and is a trainable parameter matrix.

[0095] Step 2: Based on the attribute rating matrix, model the user preferences and item features from the user-item rating perspective and the user-user association perspective, obtain the fine-grained representation of users and items under the multi-attribute channel, and obtain the multi-attribute fusion representation of the item through average pooling. The implementation method is as follows:

[0096] Learn fine-grained representations of users under multi-attribute channels from the perspective of user-item interaction and user-user association respectively;

[0097] From the perspective of user-item interaction, the rating matrix based on each attribute Construct the corresponding user-item rating graph, and use the lightweight graph neural network to aggregate node information, update the representation of each node, and obtain the user representation of the lth layer and project characterization Then, the user representation under each attribute is obtained through average pooling and project characterization The specific formula is as follows:

[0098]

[0099] In getting the attribute a k Fine-grained representation of projects under Finally, the average pooling strategy is used to fuse the fine-grained representations of the project under multiple attributes to obtain the project representation after multi-attribute fusion.

[0100]

[0101] Where P represents the number of attributes.

[0102] From the user-user association perspective, let user u h For all items in attribute a k The rating vector under User v For all items in attribute a k The rating vector under Computed property a k Next user u h and user u v Cosine similarity between Get attribute a k The user-user association matrix under The specific formula is as follows:

[0103]

[0104] Based on the attribute a kThe user-user association matrix under Build property a k The user-user association graph under , and the graph convolutional neural network is used to aggregate the neighbor node information layer by layer, so as to obtain the user representation of the lth layer under the attribute The specific formula is as follows:

[0105]

[0106] in, For attribute a k The adjacency matrix of the user-user association graph below.

[0107] Perform average pooling on the user representations of the L layer to obtain a fine-grained representation of the user from the perspective of user-user association

[0108]

[0109] After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the fine-grained representation of the user under the multi-attribute channel. The specific formula is as follows:

[0110]

[0111] in, and is a trainable parameter matrix.

[0112] Step 3: By calculating the cosine similarity between the user's total score and the representation of each attribute, the attribute weight is preliminarily determined, and then dual-channel contrastive learning is used to capture the complex relationship between the user's coarse-grained and fine-grained representations. While optimizing the user's attribute preferences, the user's coarse-grained representation under the total score and the multi-attribute fusion representation are enhanced. The implementation method is as follows:

[0113] In order to capture the relationship between each attribute and the total score and realize the effective fusion of the user's fine-grained representation, it is necessary to obtain the user's attribute preferences. First, the cosine similarity between the user's representation under each attribute and its representation under the total score is calculated. Determine the initial weights of the attributes Then, the user representation after multi-attribute fusion is obtained by weighted fusion. The specific formula is as follows:

[0114]

[0115] In order to further optimize the user's attribute preferences and representation vectors, a dual-channel contrast loss function Loss is designed. CL, maximize the consistency of the representation of the same user under two channels and the difference of the representation of different users under two channels. The specific formula is as follows:

[0116]

[0117] Where τ is the temperature coefficient.

[0118] Step 4: Concatenate the coarse-grained representation and multi-attribute fusion representation of the user and item under the total score to obtain the final representation of the user and item. Input the two into the multi-layer perceptron for linear transformation, and obtain the predicted score of the item through inner product operation. The implementation method is as follows:

[0119] The representations of the user and item under the total score are concatenated with the representations after multi-attribute fusion to obtain the final representation of the user. and the final representation of the project The specific formula is as follows:

[0120]

[0121] Use multi-layer perceptron (MLP) to finally characterize the user and final characterization of the project Perform linear transformation and obtain the final prediction score through inner product operation The specific formula is as follows:

[0122]

[0123] Step 5: By maximizing the consistency of the representation of the same user under the two channels and the difference of the representation of different users under the two channels, a dual-channel contrast loss function is designed, and the total loss function is formed by combining the mean square error loss function. Then, the method is optimized by multi-task joint learning. The implementation method is as follows:

[0124] After getting the prediction score, use the mean square error loss function Loss MSE And contrast loss function Loss CL The mixed total loss function Loss optimizes the method;

[0125] Among them, the mean square error loss is used to measure the average difference between the predicted value and the true value; the contrast loss is used to enhance the user's attribute preference and representation vector. The method is jointly optimized through multi-task learning to obtain more personalized recommendation results. The specific formula is as follows:

[0126]

[0127] Loss=Loss MSE +λLoss CL +||Θ||2 (30)

[0128] Among them, I represents the project set, y ui represents the actual rating of item i by user u in the training set, represents the predicted score of user u for item i, and Θ represents the set of hyperparameters in the algorithm.

[0129] A dual-channel graph convolutional contrastive learning recommendation system integrating multi-dimensional ratings, including the following modules:

[0130] Total score channel representation learning module: is the user's initial embedding, is the initial embedding of the project from the perspective of user-item interaction, based on the total rating matrix R n×m Construct a user-item rating graph, use a lightweight graph convolutional neural network to aggregate node information, update the representation of each node, and obtain the user representation of the lth layer and project characterization Then, the node representations of layer L are averaged and pooled to obtain the user representation e under the total score u and project characterization i From the user-user association perspective, let user u h The total score vector for all items is User v The total score vector for all items is By calculating user u h and user u v The cosine similarity sim(u h ,u v ), we can get the user-user association matrix S under the total score n×n By setting the similarity threshold γ, similar users of the target user are found, thereby constructing a user-user association graph. Considering that the scale of the user-user association graph is much smaller than the user-item rating graph, a graph convolutional neural network is used to aggregate the neighbor node information in the user-user association graph layer by layer to obtain the user representation of the lth layer. Then, the user representations of the L layer are averaged and pooled to obtain the coarse-grained representation of the user from the perspective of user-user association in the total score channel. After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the coarse-grained representation of the user under the total score channel.

[0131] Multi-attribute channel representation learning module: Based on the rating matrix of each attribute from the perspective of user-item interaction Construct the corresponding user-item rating graph, and use the lightweight graph neural network to aggregate node information, update the representation of each node, and obtain the user representation of the lth layer and project characterization Then, the user representation under each attribute is obtained through average pooling and project characterization In getting the attribute a k Fine-grained representation of projects under Finally, the average pooling strategy is used to fuse the fine-grained representations of the project under multiple attributes to obtain the project representation after multi-attribute fusion. From the user-user association perspective, let user u h For all items in attribute a k The rating vector under User v For all items in attribute a k The rating vector under Computed property a k Next user u h and user u v Cosine similarity between Get attribute a k The user-user association matrix under Similar to the method of learning node representation from the perspective of user-user association in the total score channel, based on attribute a k The user-user association matrix under Build property a k The user-user association graph under , and the graph convolutional neural network is used to aggregate the neighbor node information layer by layer, so as to obtain the user representation of the lth layer under different attributes After that, the user representation of layer L is averaged and pooled to obtain the fine-grained representation of the user from the perspective of multi-attribute channel user-user association. After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the fine-grained representation of the user under the multi-attribute channel.

[0132] Attribute preference optimization module: by calculating the cosine similarity between the user's representation under each attribute and its representation under the total score To determine the initial weight of the attribute Then, dual-channel contrastive learning is designed to optimize the user representation vector and attribute weights, and the user representation after multi-attribute fusion is obtained by weighted fusion.

[0133] Prediction score module: The representation of the user and item under the total score is spliced ​​with the representation after multi-attribute fusion to obtain the final representation of the user and the final representation of the project Use multi-layer perceptron (MLP) to finally characterize the user and final characterization of the project Perform linear transformation and then obtain the final prediction score through inner product operation

[0134] Method optimization module: After obtaining the prediction score, the multi-task joint learning method is used to optimize the mean square error loss function Loss MSE And contrastive learning loss function Loss CL The total loss function Loss obtained by mixing;

[0135] Recommendation list generation module: The system generates a corresponding item recommendation list for the user to be recommended based on the final representation of the user to be recommended.

Claims

1. A dual-channel graph convolutional contrastive learning recommendation method integrating multi-dimensional ratings, characterized in that: The following steps are involved: Step 1: By calculating the cosine similarity between different user representations, we get the user-user association matrix. Based on the user-item total score matrix and the user-user association matrix, we construct the user-item score graph and the user-user association graph. Then, we use the graph convolutional neural network and attention mechanism to jointly model the coarse-grained representation of users and items under the total score channel. Step 2: Based on the attribute rating matrix, model the user preferences and item features from the user-item rating perspective and the user-user association perspective, obtain the fine-grained representation of users and items under the multi-attribute channel, and obtain the multi-attribute fusion representation of the item through average pooling; Step 3: By calculating the cosine similarity between the user's total score and the representation of each attribute, the attribute weight is preliminarily determined, and then the relationship between the user's coarse-grained and fine-grained representations is captured using dual-channel contrastive learning. While optimizing the user's attribute preferences, the user's coarse-grained representation under the total score and the multi-attribute fusion representation are enhanced; Step 4: respectively concatenate the coarse-grained representation and multi-attribute fusion representation of the user and item under the total score to obtain the final representation of the user and item, input the two into the multi-layer perceptron for linear transformation, and obtain the predicted score of the item through inner product operation; Step 5: By maximizing the consistency of the representation of the same user under the two channels and the difference of the representation of different users under the two channels, a dual-channel contrast loss function is designed, and the total loss function is formed by combining the mean square error loss function, and then the method is optimized by multi-task joint learning.

2. According to the dual-channel graph convolution contrastive learning recommendation method integrating multi-dimensional scoring according to claim 1, it is characterized in that: The implementation method of step 1 is as follows: We learn the coarse-grained representation of users in the total rating channel from the perspective of user-item interaction and user-user association respectively; From the user-item interaction perspective, is the user's initial embedding, is the initial embedding of the project; first, based on the total score matrix R n×m Construct a user-item rating graph, and use a lightweight graph convolutional neural network to aggregate node information and update the representation of each node to obtain the user representation of the lth layer. and project characterization The specific formula is as follows: Among them, N i represents the set of users who interact with the project, N u Represents a collection of items with which the user interacts; On this basis, the node representations of layer L are averaged and pooled to obtain the user representation e from the perspective of user-item interaction. u and project characterization i ; The specific formula is as follows: From the user-user association perspective, let user u h The total score vector for all items is e uh =[r h1 ,r h2 …r hm ], user u v The total score vector for all items is e uv =[r v1 ,r v2 …r vm ], by calculating user u h and user u v The cosine similarity sim(u h ,u v ), we can get the user-user association matrix S under the total score n×n The specific formula is as follows: By setting the similarity threshold γ, similar users of the target user are found to construct a user-user association graph. Considering that the scale of the user-user association graph is much smaller than the user-item rating graph, a graph convolutional neural network is used to aggregate the neighbor node information in the user-user association graph layer by layer to obtain the user representation of the lth layer. The specific formula is as follows: in, is the adjacency matrix of the user-user association graph, represents the degree matrix, W represents the learnable weight matrix, and σ represents the nonlinear activation function; Average pooling is performed on the user representations of the L layer to obtain a coarse-grained representation of the user from the perspective of user-user association. After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the coarse-grained representation of the user under the total score channel. The specific formula is as follows: Among them, ω u and is a trainable parameter matrix.

3. According to the dual-channel graph convolution contrastive learning recommendation method integrating multi-dimensional scoring according to claim 1, it is characterized in that: The implementation method of step 2 is as follows: Learn fine-grained representations of users under multi-attribute channels from the perspective of user-item interaction and user-user association respectively; From the perspective of user-item interaction, the rating matrix based on each attribute Construct the corresponding user-item rating graph, and use the lightweight graph neural network to aggregate node information, update the representation of each node, and obtain the user representation of the lth layer and project characterization Then, the user representation under each attribute is obtained through average pooling and project characterization The specific formula is as follows: In getting the attribute a k Fine-grained representation of projects under Finally, the average pooling strategy is used to fuse the fine-grained representations of the project under multiple attributes to obtain the project representation after multi-attribute fusion. Where P represents the number of attributes. From the user-user association perspective, let user u h For all items in attribute a k The rating vector under User v For all items in attribute a k The rating vector under Computed property a k Next user u h and user u v Cosine similarity between Get attribute a k The user-user association matrix under The specific formula is as follows: Based on the attribute a k The user-user association matrix under Build property a k The user-user association graph under , and the graph convolutional neural network is used to aggregate the neighbor node information layer by layer, so as to obtain the user representation of the lth layer under the attribute The specific formula is as follows: in, For attribute a k The adjacency matrix of the user-user association graph is as follows; Perform average pooling on the user representations of the L layer to obtain a fine-grained representation of the user from the perspective of user-user association After obtaining the user representations from the above two perspectives, the two are weightedly fused through the attention mechanism to obtain the fine-grained representation of the user under the multi-attribute channel. The specific formula is as follows: in, and is a trainable parameter matrix.

4. According to claim 1, a dual-channel graph convolution contrastive learning recommendation method integrating multi-dimensional scoring is characterized in that: The implementation method of step 3 is as follows: In order to capture the relationship between each attribute and the total score and realize the effective fusion of the user's fine-grained representation, it is necessary to obtain the user's attribute preference. First, the cosine similarity between the user's representation under each attribute and its representation under the total score is calculated. Determine the initial weights of the attributes Then, the user representation after multi-attribute fusion is obtained by weighted fusion. The specific formula is as follows: In order to further optimize the user's attribute preferences and representation vectors, a dual-channel contrast loss function Loss is designed. CL , maximize the consistency of the representation of the same user under two channels and the difference of the representation of different users under two channels; the specific formula is as follows: Where τ is the temperature coefficient.

5. The dual-channel graph convolution contrastive learning recommendation method integrating multi-dimensional scoring according to claim 1, characterized in that: The implementation method of generating the prediction score in step 4 is as follows: The representations of the user and item under the total score are concatenated with the representations after multi-attribute fusion to obtain the final representation of the user. and the final representation of the project The specific formula is as follows: Use multi-layer perceptron (MLP) to finally characterize the user and final characterization of the project Perform linear transformation and obtain the final prediction score through inner product operation The specific formula is as follows:

6. The dual-channel graph convolution contrastive learning recommendation method integrating multi-dimensional scoring according to claim 1, characterized in that: The specific optimization method in step 5 is as follows: After getting the prediction score, use the mean square error loss function Loss MSE And contrast loss function Loss CL The mixed total loss function Loss optimizes the method; Among them, the mean square error loss is used to measure the average difference between the predicted value and the true value; the contrast loss is used to enhance the user's attribute preference and representation vector; the method is jointly optimized through multi-task learning to obtain more personalized recommendation results; the specific formula is as follows: Among them, I represents the project set, y ui represents the actual rating of item i by user u in the training set, represents the predicted score of user u for item i, and Θ represents the set of hyperparameters in the algorithm.

7. A dual-channel graph convolutional contrastive learning recommendation system integrating multi-dimensional ratings, characterized in that: Includes the following modules: Total score channel representation learning module: learn the coarse-grained representation of users and items under the total score; first calculate the cosine similarity sim between different users to obtain the user-user association matrix S n×n , and then based on the user-item total score matrix R n×m and the user-user association matrix S n×n , construct the user-item rating graph and the user-user association graph, and then use the graph convolutional neural network and attention mechanism to jointly model the coarse-grained representation of users under this channel and the coarse-grained representation of the project i ; Multi-attribute channel representation learning module: learns the fine-grained representation of users and projects under each attribute; similar to the total score channel, based on each attribute score matrix, model user preferences and project features from the user-project rating perspective and the user-user association perspective, and obtain the fine-grained representation of users and projects under each attribute and The average pooling strategy is used to fuse the fine-grained representations of the project under multiple attributes to obtain the multi-attribute fusion representation of the project. Attribute preference optimization module: Enhance the user's representation by updating the attribute preference; use the cosine similarity between the user's total score and the representation under each attribute Preliminary determination of attribute weights Then, the complex relationship between the coarse-grained and fine-grained representations of users is captured through dual-channel contrastive learning. While optimizing user attribute preferences, the coarse-grained representation and multi-attribute fusion representation of users under the total score are enhanced. Prediction score module: concatenates the coarse-grained representation and multi-attribute fusion representation of users and items under the total score to obtain the final representation of users and items and The two are input into the multi-layer perceptron for linear transformation, and the prediction score is obtained through inner product operation; Method optimization module: Use multi-task joint learning method to optimize the total loss function obtained by mixing the mean square error loss function and the contrastive learning loss function; Recommendation list generation module: The system generates a corresponding item recommendation list for the user to be recommended based on the final representation of the user to be recommended.

Images