A commodity recommendation method and system based on graphical decoupling representation learning

Abstract

The application discloses a commodity recommendation method and system based on graphic decoupling characteristic learning, and belongs to the technical field of electronic commerce and artificial intelligence recommendation. In view of the problem that the current electronic commerce recommendation system cannot effectively capture the complex purchase decision and behavior change of consumers and is difficult to meet the personalized demand, firstly, the popularity factor and the interest factor influencing the user decision are decoupled, and a preliminary user-item interest representation and a popularity representation are constructed; then, a contrast learning mechanism is introduced to optimize the decoupled representation space, so as to reduce the coupling degree between the interest embedding and the popularity embedding; the real intention of the user is mined, and thus accurate commodity recommendation is realized. The application effectively improves the ability of the recommendation system in the aspect of personalized modeling, and has good practicability and scalability.

Description

Technical Field

[0001] This invention belongs to the field of e-commerce and artificial intelligence recommendation technology, specifically relating to a product recommendation method and system based on graph decoupling representation learning, which is particularly suitable for online personalized product recommendations. Background Technology

[0002] As e-commerce enters a new phase where experience and efficiency are the core competitive advantages, AI-driven personalized recommendation systems have evolved from a value-added function into a core infrastructure and strategic engine supporting industry growth. Currently, this technology is undergoing a profound transformation from static collaborative filtering to deep dynamic intelligence. Static collaborative filtering algorithms primarily rely on users' historical consumption data, likes, comments, and favorites for recommendations. However, in practical applications, popular products often receive greater exposure and are selected and consumed by a large number of users. This creates a dominant interaction record in the recommendation system, easily overlooking users' potential interests in other products, exacerbating the popularity bias problem, and affecting recommendation effectiveness. Therefore, understanding users' true needs is crucial for personalized product recommendations.

[0003] In personalized product recommendations, traditional graph neural network models commonly suffer from popularity bias. As a potential bias factor in implicit feedback data, product popularity leads to homogenized recommendation results, exacerbating the long-tail effect, affecting the fairness of product exposure, and ultimately reducing the overall diversity of the system, thus harming the user experience. In contrast, graph decoupling representation learning can effectively separate popularity factors from users' true needs, thereby revealing the user's true intent. Summary of the Invention

[0004] To address the problem that current e-commerce recommendation systems cannot effectively capture consumers' complex purchasing decisions and behavioral changes, and thus fail to meet personalized needs, this invention provides a product recommendation method and system based on graph decoupling representation learning.

[0005] To achieve the above objectives, the present invention employs the following technical solutions:

[0006] A product recommendation method based on graph decoupling representation learning, the method comprising the following steps:

[0007] Step 1: Constructing the user-product interaction graph. Obtain data from the website and construct the user-product interaction graph based on various relationships.

[0008] Step 1 specifically involves:

[0009] Step 1.1: Based on the needs of the product recommendation task, obtain and organize user interaction data with products from the website;

[0010] Step 1.2: Construct a user-product interaction network using interactive data. , where the node set , Indicates inclusion A set of user nodes express A set of product nodes, i.e. , edge set To represent the observed user and product interaction records in the edge set Each pair Each edge represents a set of user and product interactions, and the weight of all edges is fixed at 1.

[0011] Step 2, Decoupled Representation of User-Product Interest and Popularity: This is used to analyze the true reasons for user interaction with products and to distinguish the impact of popularity and interest on users. By dividing the user-product interaction graph in Step 1 into four interaction sub-graphs, the representations of users and products are decoupled as follows: user's personal interest representation and user popularity representation, product content feature representation and product popularity representation;

[0012] Step 2 specifically involves:

[0013] Step 2.1: Determine the threshold using the Otsu's method. and ; Used to categorize goods into popular and non-popular goods. It is used to divide users into those who are highly dependent on trendy products (i.e., those who are more likely to follow the crowd) and those who are less dependent on trendy products (i.e., those who are less likely to follow the crowd).

[0014] Calculate each product Number of interactions / Total number of product interactions and each user Average popularity of interactive products and initialize both;

[0015] The user-product interaction graph obtained in step 1 is decomposed into four interaction subgraphs, each corresponding to a set of users with high conformity. Low herd mentality user group Popular Products Collection Long-tail product collection Four subsets, the first interactive subgraph on the product side. Depend on The second subgraph consists of the products and the users who have interacted with them. Depend on The first user-side interaction subgraph consists of the products and the users who have interacted with them. Depend on The second subgraph consists of users and the products they have interacted with. Depend on It consists of users and the products they interact with.

[0016] The user's interest representation, user's popularity representation, product's interest representation, and product's popularity representation are learned from the four interaction subgraphs respectively; the propagation process formulas for interest representation and popularity representation are as follows:

[0017] (1);

[0018] (2);

[0019] in Indicates the first The interest representation matrix propagated from the layer, Indicates the first The popularity representation matrix propagated from the layers, This represents the dimension of the embedding vector. Representing the The adjacency matrix of each subgraph Representing the Degree matrix of user-product interaction subgraph; Used to identify the interest representation matrix, Used to identify the popularity representation matrix;

[0020] After After propagation through the layers, the outputs of each propagation layer are aggregated to obtain... The final interest representation matrix of each user and the final popularity representation matrix The formula is as follows:

[0021] (3);

[0022] (4);

[0023] Encoders using shared parameters on the product side, after passing through After propagation through the layers, the outputs of each propagation layer are aggregated to obtain... Final interest representation matrix of each product and the final popularity representation matrix The formula is as follows:

[0024] (5);

[0025] (6);

[0026] Step 3, Contrastive Learning to Optimize Representations: Starting from both user and product perspectives, contrastive learning is used to design two metrics, separation degree and prediction consistency, as feedback signals to enhance the decoupled interest representation and popularity representation, enabling these representations to better capture the intrinsic characteristics of users and products;

[0027] Step 3 specifically involves:

[0028] Step 3.1: Design two metrics, separation degree and prediction consistency, as feedback signals to evaluate the initial results. and Iterative updates are performed, and the separation metric is used to measure the spatial discriminability of interest and popularity representations, which is achieved by minimizing mutual information. The core objective of minimizing mutual information is to reduce redundant information between interest and popularity representations, making them more semantically independent. The loss function for this process is as follows:

[0029] (7);

[0030] in express One of the lines represents the user. interest expression, yes One line in the middle represents the user Popularity is indicated by It is the true joint distribution of positive sample pairs. This represents a variational approximation distribution used to estimate conditional probabilities. Indicates the distribution of edge multiplication integrals The expected log probability of negative sample pairs. It is a pair of negative samples;

[0031] Step 3.2, the second type of metric is the predictive consistency metric, used to determine whether interest metrics are better at predicting interactions with long-tail products, while popularity metrics are better at predicting interactions with popular products; based on the popular product set segmented in Step 2. and long-tail product collection Extract the interaction subsets of popular products from them. long-tail product interaction subset Then for any sample Calculate interest prediction score Popularity prediction score :

[0032] (8);

[0033] (9);

[0034] in, express One of the rows represents the products. interest expression, yes One of the rows represents the product. Popularity is indicated by This means using the Sigmoid function to map interest scores to [0,1]. This means that the popularity score is mapped to [0,1] using the Sigmoid function;

[0035] Then calculate the consistency loss. :

[0036] (10);

[0037] in, The binary label is either 0 or 1. Represents the binary cross-entropy loss function; Represents a subset of long-tail product interactions; This represents a subset of interactions with popular products. This indicates the number of samples in the long-tail product interaction subset; This indicates the number of samples in the popular product interaction subset;

[0038] Combining the loss functions of the two types of feedback signals above, we obtain the overall loss function for the contrastive learning task:

[0039] (11);

[0040] Dynamic optimization through gradient descent and By gradually improving the feature decoupling effect, we obtain enhanced interest representation and popularity representation.

[0041] Step 4, Adaptive Fusion: Using the obtained user and product interest representations and popularity representations, adaptive fusion is performed to generate new user and product representations;

[0042] Step 4 specifically involves:

[0043] Step 4.1: Use the interest representation as the query and the popularity representation as the key for attention weighting. Since the query and key are vectors of different lengths, additive attention is used as the scoring function, given the user's interest representation. and user popularity The scoring function for additive attention is:

[0044] (12);

[0045] (13);

[0046] (14);

[0047] in, , The weight matrix is ​​a learnable matrix. Used to map queries to a hidden space. The formula for updating keys to the hidden space is shown above. The dimension of the hidden layer is a hyperparameter. and All are embedding dimensions of interest representation; Indicates the learning rate. Map the output of the hidden layer to a scalar. As an attention score; the query and key are concatenated and fed into a multilayer perceptron (MLP) using tanh as the activation function; This represents the total loss function; Indicates transpose;

[0048] Step 4.2, Adaptive Weighted Connection: Using an additive attention mechanism, calculate the attention weights for the user interest representation and popularity representation. The formula is as follows:

[0049] (15);

[0050] Calculate attention weights for product interest representation and popularity representation. The formula is as follows:

[0051] (16);

[0052] The comprehensive representation vector is then expressed as:

[0053] (17);

[0054] (18);

[0055] in, The final representation on behalf of the user The final representation of the product.

[0056] Step 5, Model Building and Optimization: Combine the comparison task from Step 3 with the recommendation task to build a product recommendation model based on graph decoupling representation learning. Design a joint optimization objective function and train the model using the objective function.

[0057] Step 5 specifically involves:

[0058] Step 5.1: Combining the comparison task and the recommendation task, construct a product recommendation model based on graph decoupling representation learning; for the product recommendation model, design a joint optimization objective function and train the model using the objective function; the joint optimization objective loss function is shown in the following equation:

[0059] (19);

[0060] in, This represents the Bayesian personalized ranking loss, used to optimize recommendation tasks. These are adjustable hyperparameters used to regulate the weight distribution among different tasks; This represents the overall loss function for the contrastive learning task;

[0061] Step 5.2: Based on the joint optimization objective function, backpropagation of error is performed using the gradient descent method to determine the undetermined parameters in the product recommendation model based on decoupled representation learning. Update the model and complete the training.

[0062] Step 6, Recommendation Result Output: Using the trained product recommendation model based on graph decoupling representation learning, product recommendations are made and the results are provided to users and merchants for use.

[0063] Step 6 specifically involves:

[0064] Step 6.1: When a user generates a shopping request, the product recommendation model based on graph decoupling representation learning generates a corresponding recommendation sequence based on the user's personal preferences and historical consumption behavior.

[0065] T (20);

[0066] (twenty one);

[0067] in, Used to predict users For goods The preference values ​​are sorted from highest to lowest based on a specific user's preference values ​​for different products, and the top-ranked products are selected. For each product, generate a recommendation sequence. , The final representation on behalf of the user The final representation of the product is indicated by T, which stands for transpose.

[0068] Step 6.2 provides personalized product recommendations to users; helps users quickly find products that match their personal preferences; and helps store operators develop more targeted marketing strategies to attract customers and drive business growth.

[0069] A product recommendation system based on graph decoupling representation learning is provided. The system is used to implement the product recommendation method based on graph decoupling representation learning, and includes a computer processor and memory, a user-product interest and popularity decoupling unit, a contrastive learning optimization unit, an adaptive fusion unit, a product recommendation training unit based on graph decoupling representation learning, and a product recommendation result output unit.

[0070] The user-product interest and popularity decoupling unit executes steps 1 and 2, obtains data from the website, constructs a user-product interaction graph, and quantifies user and product popularity to obtain preliminary decoupled user-product interest and popularity representations.

[0071] In the contrastive learning optimization unit, step 3 is executed. From the perspectives of both users and products, contrastive learning is used to design feedback signals to enhance the decoupled interest representation and popularity representation, so that these representations can better capture the intrinsic characteristics of users and products.

[0072] The adaptive fusion unit executes step 4. On the original representation of the user-item interaction graph learned in steps 2 and 3, it automatically calculates the attention weights of interest representation and popularity representation through a learnable attention network and dynamically adjusts the importance of the two types of features based on the intrinsic correlation between features. Finally, it obtains the predicted score through weighted fusion, and obtains the final user-item representation.

[0073] The product recommendation training unit based on graph decoupling representation learning executes step 5, constructing a product recommendation based on graph decoupling representation learning according to each step from step 1 to step 4, and training the model based on the defined objective function.

[0074] The product recommendation result output unit executes step 6, which categorizes products with high user interest on the retail platform into recommendation columns and provides the recommendation results to users and merchants on the retail platform.

[0075] Compared with the prior art, the present invention has the following advantages:

[0076] The product recommendation method of this invention can capture the user's true shopping intention and more accurate product information by decoupling the popularity and interest representation of users and products. It can also correctly utilize the effective information brought by popularity while eliminating the bias caused by popularity information, making the recommendations more diverse and personalized.

[0077] This invention utilizes a graph-based decoupling representation learning method to establish a product recommendation model, capturing the true intent of user and product interactions, enabling the model to have a certain generative capability, thereby giving the product recommendation process a stronger generalization ability. Attached Figure Description

[0079] Figure 1 This is a structural diagram of a product recommendation model based on graph decoupling representation learning according to the present invention;

[0080] Figure 2 This is a flowchart of a product recommendation method based on graph decoupling representation learning according to the present invention;

[0081] Figure 3 This is a structural diagram of a product recommendation system based on graph decoupling representation learning according to the present invention. Detailed Implementation

[0083] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0084] The product recommendation method and system based on graph decoupling representation learning described in this invention are implemented through a computer program, and will be described below according to... Figure 2 The illustrated process details the specific implementation of the technical solution proposed in this invention. The technical solution of this invention is used for recommendation on the Clothing dataset, which comes from the Amazon platform and includes user ratings, reviews, product information (such as name, category, etc.), and information about relationships between users. We divide the dataset into three parts: a training set, a validation set, and a test set, in a ratio of 7:1:2.

[0085] The implementation method mainly includes the following key aspects:

[0086] A product recommendation method based on graph decoupling representation learning, the method comprising the following steps:

[0087] Step 1: Constructing the user-product interaction graph. Obtain data from the website and construct the user-product interaction graph based on various relationships.

[0088] Step 1 specifically involves:

[0089] Step 1.1: Based on the needs of the product recommendation task, obtain and organize user interaction data with products from the website;

[0090] Step 1.2: Construct a user-product interaction network using interactive data. , where the node set , Indicates inclusion A set of user nodes express A set of product nodes, i.e. , edge set To represent the observed user and product interaction records in the edge set Each pair Each edge represents a set of user and product interactions, and the weight of all edges is fixed at 1.

[0091] Step 2, Decoupled Representation of User-Product Interest and Popularity: This is used to analyze the true reasons for user interaction with products and to distinguish the impact of popularity and interest on users. By dividing the user-product interaction graph in Step 1 into four interaction sub-graphs, the representations of users and products are decoupled as follows: user's personal interest representation and user popularity representation, product content feature representation and product popularity representation;

[0092] Step 2 specifically involves:

[0093] Step 2.1: Determine the threshold using the Otsu's method. and ; Used to categorize goods into popular and non-popular goods. It is used to divide users into those who are highly dependent on trendy products (i.e., those who are more likely to follow the crowd) and those who are less dependent on trendy products (i.e., those who are less likely to follow the crowd).

[0094] Calculate each product Number of interactions / Total number of product interactions and each user Average popularity of interactive products and initialize both;

[0095] The user-product interaction graph obtained in step 1 is decomposed into four interaction sub-graphs, each corresponding to a highly conformist user set. Low herd mentality user group Popular Products Collection Long-tail product collection Four subsets, the first interactive subgraph on the product side. Depend on The second subgraph consists of the products and the users who have interacted with them. Depend on The first user-side interaction subgraph consists of the products and the users who have interacted with them. Depend on The second subgraph consists of users and the products they have interacted with. Depend on It consists of users and the products they interact with.

[0096] The user's interest representation, user's popularity representation, product's interest representation, and product's popularity representation are learned from the four interaction subgraphs respectively; the propagation process formulas for interest representation and popularity representation are as follows:

[0097] (1);

[0098] (2);

[0099] in Indicates the first The interest representation matrix propagated from the layer, Indicates the first The popularity representation matrix propagated from the layers, This represents the dimension of the embedding vector. Representing the The adjacency matrix of each subgraph Representing the Degree matrix of user-product interaction subgraph; Used to identify the interest representation matrix, Used to identify the popularity representation matrix;

[0100] After After propagation through the layers, the outputs of each propagation layer are aggregated to obtain... The final interest representation matrix of each user and the final popularity representation matrix The formula is as follows:

[0101] (3);

[0102] (4);

[0103] Encoders using shared parameters on the product side, after passing through After propagation through the layers, the outputs of each propagation layer are aggregated to obtain... Final interest representation matrix of each product and the final popularity representation matrix The formula is as follows:

[0104] (5);

[0105] (6);

[0106] Step 3, Contrastive Learning to Optimize Representations: Starting from both user and product perspectives, contrastive learning is used to design two metrics, separation degree and prediction consistency, as feedback signals to enhance the decoupled interest representation and popularity representation, enabling these representations to better capture the intrinsic characteristics of users and products;

[0107] Step 3 specifically involves:

[0108] Step 3.1: Design two metrics, separation degree and prediction consistency, as feedback signals to evaluate the initial results. and Iterative updates are performed; the separation metric is used to measure the spatial distinguishability of interest and popularity representations, which is achieved by minimizing mutual information; the core objective of minimizing mutual information is to reduce redundant information between interest and popularity representations, making them more semantically independent; the optimization objectives are as follows:

[0109] (7);

[0110] in express One of the lines represents the user. interest expression, yes One line in the middle represents the user Popularity is indicated by It is the true joint distribution of positive sample pairs. This represents a variational approximation distribution used to estimate conditional probabilities. Indicates the distribution of edge multiplication integrals The expected log probability of negative sample pairs. It is a pair of negative samples;

[0111] Step 3.2, the second type of feedback metric is the predictive consistency metric, used to determine whether interest metrics are better at predicting interactions with long-tail products, while popularity metrics are better at predicting interactions with popular products; based on the popular product set segmented in Step 2. and long-tail product collection Extract the interaction subsets of popular products from them. long-tail product interaction subset Then for any sample Calculate interest prediction score Popularity prediction score :

[0112] (8);

[0113] (9);

[0114] in, express One of the rows represents the products. interest expression, yes One of the rows represents the product. Popularity is indicated by This means using the Sigmoid function to map interest scores to [0,1]. This means that the popularity score is mapped to [0,1] using the Sigmoid function;

[0115] Then calculate the consistency loss. :

[0116] (10);

[0117] in, The binary label is either 0 or 1. Represents the binary cross-entropy loss function; Represents a subset of long-tail product interactions; This represents a subset of interactions with popular products. This indicates the number of samples in the long-tail product interaction subset; This indicates the number of samples in the popular product interaction subset;

[0118] The total feedback signal is obtained by combining the two types of feedback signals above:

[0119] (11);

[0120] Dynamic optimization through gradient descent and By gradually improving the feature decoupling effect, we obtain enhanced interest representation and popularity representation.

[0121] Step 4, Adaptive Fusion: Using the obtained user and product interest representations and popularity representations, adaptive fusion is performed to generate new user and product representations;

[0122] Step 4 specifically involves:

[0123] Step 4.1: Use the interest representation as the query and the popularity representation as the key for attention weighting. Since the query and key are vectors of different lengths, additive attention is used as the scoring function, given the user's interest representation. and user popularity The scoring function for additive attention is:

[0124] (12);

[0125] (13);

[0126] (14);

[0127] in, , The weight matrix is ​​a learnable matrix. Used to map queries to a hidden space. The formula for updating keys to the hidden space is shown above. The dimension of the hidden layer is a hyperparameter. and All are embedding dimensions of interest representation; Indicates the learning rate. Map the output of the hidden layer to a scalar. As an attention score; the query and key are concatenated and fed into a multilayer perceptron (MLP) using tanh as the activation function; This represents the total loss function; Indicates transpose;

[0128] Step 4.2, Adaptive Weighted Connection: Using an additive attention mechanism, calculate the attention weights for the user interest representation and popularity representation. The formula is as follows:

[0129] (15);

[0130] Calculate attention weights for product interest representation and popularity representation. The formula is as follows:

[0131] (16);

[0132] The comprehensive representation vector is then expressed as:

[0133] (17);

[0134] (18);

[0135] in, The final representation on behalf of the user The final representation of the product.

[0136] Step 5, Model Building and Optimization: Combine the comparison task from Step 3 with the recommendation task to build a product recommendation model based on graph decoupling representation learning. Design a joint optimization objective function and train the model using the objective function.

[0137] Step 5 specifically involves:

[0138] Step 5.1: Combining the comparison task and the recommendation task, construct a product recommendation model based on graph decoupling representation learning; for the product recommendation model, design a joint optimization objective function and train the model using the objective function; the joint optimization objective loss function is shown in the following equation:

[0139] (19);

[0140] in, This represents the Bayesian personalized ranking loss, used to optimize recommendation tasks. These are adjustable hyperparameters used to regulate the weight distribution among different tasks; This represents the overall loss function for the contrastive learning task;

[0141] Step 5.2: Based on the joint optimization objective function, backpropagation of error is performed using the gradient descent method to determine the undetermined parameters in the product recommendation model based on decoupled representation learning. Update the model and complete the training.

[0142] Step 6, Recommendation Result Output: Using the trained product recommendation model based on graph decoupling representation learning, product recommendations are made and the results are provided to users and merchants for use.

[0143] Step 6 specifically involves:

[0144] Step 6.1: When a user generates a shopping request, the product recommendation model based on graph decoupling representation learning generates a corresponding recommendation sequence based on the user's personal preferences and historical consumption behavior.

[0145] T (20);

[0146] (twenty one);

[0147] in, Used to predict users For goods The preference values ​​are sorted from highest to lowest based on a specific user's preference values ​​for different products, and the top-ranked products are selected. For each product, generate a recommendation sequence. , The final representation on behalf of the user The final representation of the product is indicated by T, which stands for transpose.

[0148] Step 6.2 provides personalized product recommendations to users; helps users quickly find products that match their personal preferences; and helps store operators develop more targeted marketing strategies to attract customers and drive business growth.

[0149] A product recommendation system based on graph decoupling representation learning is provided. The system is used to implement the product recommendation method based on graph decoupling representation learning, and includes a computer processor and memory, a user-product interest and popularity decoupling unit, a contrastive learning optimization unit, an adaptive fusion unit, a product recommendation training unit based on graph decoupling representation learning, and a product recommendation result output unit.

[0150] The user-product interest and popularity decoupling unit executes steps 1 and 2, obtains data from the website, constructs a user-product interaction graph, and quantifies user and product popularity to obtain preliminary decoupled user-product interest and popularity representations.

[0151] In the contrastive learning optimization unit, step 3 is executed. From the perspectives of both users and products, contrastive learning is used to design feedback signals to enhance the decoupled interest representation and popularity representation, so that these representations can better capture the intrinsic characteristics of users and products.

[0152] The adaptive fusion unit executes step 4. On the original representation of the user-item interaction graph learned in steps 2 and 3, it automatically calculates the attention weights of interest representation and popularity representation through a learnable attention network and dynamically adjusts the importance of the two types of features based on the intrinsic correlation between features. Finally, it obtains the predicted score through weighted fusion, and obtains the final user-item representation.

[0153] The product recommendation training unit based on graph decoupling representation learning executes step 5, constructing a product recommendation model according to each step from step 1 to step 4, and training the model based on the defined objective function.

[0154] The product recommendation results output unit executes step 6, categorizing products with high user interest on the retail platform into recommended categories and providing the product recommendation results to users and merchants on the retail platform. This helps users quickly find products of interest, improves the shopping experience, and increases the convenience of purchasing. At the same time, for merchants, accurate recommendations can effectively increase product exposure, help merchants better understand consumer preferences and needs, thereby optimizing inventory and marketing strategies and achieving a win-win situation.

[0155] To verify the effectiveness and advancement of the proposed technical solution, the recommendation performance of this invention is compared with baseline methods LightGCN, CasuE, DECL, and LGMRec. Recall@20 and the standardized discount cumulative gain NDCG@20 are used as evaluation metrics for recommendation performance. The product recommendation results of the above methods are evaluated using the Clothing dataset in the examples, and the results are shown in Table 1.

[0156] Table 1 Comparison and Analysis of Results

[0157] Method indicators LightGCN CasuE DECL LGMRec This invention Recall@20 0.0526 0.0542 0.0604 0.0828 0.1023 NDCG@20 0.0236 0.0255 0.0273 0.0371 0.0452

[0158] The results in the table show that the technical solution of the present invention can obtain relatively accurate recommendation results when making product recommendations.

[0159] like Figure 3 As shown, a product recommendation system based on graph decoupling representation learning includes a computer processor and memory, a user-product interest and popularity decoupling unit, a contrastive learning optimization unit, an adaptive fusion unit, a product recommendation training unit based on graph decoupling representation learning, and a product recommendation result output unit. The user-product interest and popularity decoupling unit executes steps 1 and 2, acquiring data from the website, including user ratings, reviews, and product names and categories, constructing a user-product interaction graph, and quantifying user and product popularity to obtain preliminary decoupled user-product interest and popularity representations. The contrastive learning optimization unit executes step 3, approaching the issue from both user and product perspectives, using contrastive learning to design feedback signals to enhance the decoupled interest and popularity representations, enabling these representations to better capture the intrinsic characteristics of users and products. The adaptive fusion unit executes step 4, automatically calculating the attention weights of the interest and popularity representations based on the original representations of the user-product interaction graph learned in steps 2 and 3 using a learnable attention network, and then applying these weights based on the intrinsic characteristics between features. The importance of the two types of features is dynamically adjusted, and a weighted fusion is used to obtain the predicted score, resulting in the final user and product representations. The product recommendation training unit, based on graph decoupling representation learning, executes step 5, constructing a product recommendation model according to each step from 1 to 4, and training the model based on the defined objective function. The product recommendation result output unit executes step 6, categorizing products with high user interest in the platform into recommendation columns and providing the product recommendation results to users and merchants on the retail platform. This helps users quickly find products of interest, improves the shopping experience, and increases the convenience of purchasing. Simultaneously, for merchants, accurate recommendations effectively increase product exposure, helping them better understand consumer preferences and needs, thereby optimizing inventory and marketing strategies and achieving a win-win situation.

[0160] The foregoing has shown and described the main features and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0161] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0162] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A product recommendation method based on graph decoupling representation learning, characterized in that: The method includes the following steps: Step 1: Constructing the user-product interaction graph. Obtain data from the website and construct the user-product interaction graph based on various relationships. Step 2, Decoupled representation of user-product interests and popularity: By dividing the user-product interaction graph in Step 1 into four interaction subgraphs, the representations of users and products are decoupled into: user personal interest representation and user popularity representation, product content feature representation and product popularity representation; Step 3, Contrastive Learning to Optimize Representations: Starting from both user and product perspectives, contrastive learning is used to design two metrics, separation degree and prediction consistency, as feedback signals to enhance the decoupled interest representation and popularity representation; Step 4, Adaptive Fusion: Using the obtained user and product interest representations and popularity representations, adaptive fusion is performed to generate new user and product representations; Step 5, Model Building and Optimization: Combine the comparison task from Step 3 with the recommendation task to build a product recommendation model based on graph decoupling representation learning. Design a joint optimization objective function and train the model using the objective function. Step 6, Recommendation Result Output: Using the trained product recommendation model based on graph decoupling representation learning, product recommendations are made and the results are provided to users and merchants for use.

2. The product recommendation method based on graph decoupling representation learning according to claim 1, characterized in that: Step 1 specifically involves: Step 1.1: Based on the needs of the product recommendation task, obtain and organize user interaction data with products from the website; Step 1.2: Construct a user-product interaction network using interactive data. , where the node set , Indicates inclusion A set of user nodes express A set of product nodes, i.e. , Edge set This represents the observed user and product interaction records in the edge set. Each pair Each edge represents a set of user and product interactions, and the weight of all edges is fixed at 1.

3. The product recommendation method based on graph decoupling representation learning according to claim 2, characterized in that: Step 2 specifically involves: Step 2.1: Determine the threshold using the Otsu's method. and ; Used to categorize goods into popular and non-popular goods. It is used to divide users into those who are highly dependent on trendy products and those who are less dependent on trendy products; Calculate each product Number of interactions / Total number of product interactions and each user Average popularity of interactive products and initialize both; The user-product interaction graph obtained in step 1 is decomposed into four interaction subgraphs, each corresponding to a set of users with high conformity. Low herd mentality user group Popular Products Collection Long-tail product collection Four subsets, the first interactive subgraph on the product side. Depend on The second subgraph consists of the products and the users who have interacted with them. Depend on The product and the users who have interacted with it constitute the user's first interaction subgraph. Depend on The second subgraph consists of users and the products they have interacted with. Depend on The user base and the products it interacts with constitute the user base; The user's interest representation, user's popularity representation, product's interest representation, and product's popularity representation are learned from the four interaction subgraphs respectively; the propagation process formulas for interest representation and popularity representation are as follows: (1)； (2)； in, Indicates the first The interest representation matrix propagated from the layer, Indicates the first The popularity representation matrix propagated from the layers, This represents the dimension of the embedding vector. Representing the The adjacency matrix of each subgraph Representing the Degree matrix of user-product interaction subgraph; Used to identify the interest representation matrix, Used to identify the popularity representation matrix; After After propagation through the layers, the outputs of each propagation layer are aggregated to obtain... The final interest representation matrix of each user and the final popularity representation matrix The formula is as follows: (3)； (4)； Encoders using shared parameters on the product side, after passing through After propagation through the layers, the outputs of each propagation layer are aggregated to obtain... Final interest representation matrix of each product and the final popularity representation matrix The formula is as follows: (5)； (6)。 4. The product recommendation method based on graph decoupling representation learning according to claim 3, characterized in that: Step 3 specifically involves: Step 3.1: Design two metrics, separation degree and prediction consistency, as feedback signals to evaluate the initial results. and We will conduct iterative optimization and updates, with the following optimization objectives: (7)； in, for One of the lines represents the user. interest expression, for One line in the middle represents the user Popularity is indicated by It is the true joint distribution of positive sample pairs. This represents a variational approximation distribution used to estimate conditional probabilities. Represents the distribution of edge multiplication integrals The expected log probability of negative sample pairs. It is a pair of negative samples; Step 3.2, the second type of feedback metric is the prediction consistency metric, used to determine whether interest representation is better at predicting interactions with long-tail products, based on the popular product set segmented in Step 2. and long-tail product collection Extract the interaction subsets of popular products from them. long-tail product interaction subset Then for any sample Calculate interest prediction score Popularity prediction score : (8)； (9)； in, for One of the rows represents the products. interest expression, for One of the rows represents the product. Popularity is indicated by This means using the Sigmoid function to map interest scores to [0,1]. This means that the popularity score is mapped to [0,1] using the Sigmoid function; Then calculate the consistency loss. : (10)； in, The binary label is either 0 or 1. Represents the binary cross-entropy loss function; Represents a subset of long-tail product interactions; This represents a subset of interactions with popular products. This indicates the number of samples in the long-tail product interaction subset; This indicates the number of samples in the popular product interaction subset; The total feedback signal is obtained by combining the two types of feedback signals above: (11)； Dynamic optimization through gradient descent and By gradually improving the feature decoupling effect, we obtain enhanced interest representation and popularity representation.

5. The product recommendation method based on graph decoupling representation learning according to claim 4, characterized in that: Step 4 specifically involves: Step 4.1: Use the interest representation as the query and the popularity representation as the key for attention weighting, and use additive attention as the scoring function, given the user's interest representation. and user popularity The scoring function for additive attention is: (12)； (13)； (14)； in, , The weight matrix is ​​a learnable matrix. Used to map queries to a hidden space. The formula for updating keys to the hidden space is shown above. The dimension of the hidden layer is a hyperparameter. and All are embedding dimensions of interest representation; Indicates the learning rate. Map the output of the hidden layer to a scalar. As an attention score; the query and key are concatenated and input into a multilayer perceptron, with tanh as the activation function; This represents the total loss function; Indicates transpose; Step 4.2, Adaptive Weighted Connection: Using an additive attention mechanism, calculate the attention weights for the user interest representation and popularity representation. The formula is as follows: (15)； Calculate attention weights for product interest representation and popularity representation. The formula is as follows: (16)； The comprehensive representation vector is then expressed as: (17)； (18)； in, The final representation on behalf of the user The final representation of the product.

6. The product recommendation method based on graph decoupling representation learning according to claim 5, characterized in that: Step 5 specifically involves: Step 5.1: Combining the comparison task and the recommendation task, construct a product recommendation model based on graph decoupling representation learning; for the product recommendation model, design a joint optimization objective function and train the model using the objective function; the joint optimization objective loss function is shown in the following equation: (19)； in, This represents the Bayesian personalized ranking loss, used to optimize recommendation tasks. These are adjustable hyperparameters used to regulate the weight distribution among different tasks; This represents the overall loss function for the contrastive learning task; Step 5.2: Based on the joint optimization objective function, backpropagation of error is performed using the gradient descent method to determine the undetermined parameters in the product recommendation model based on decoupled representation learning. Update the model and complete the training.

7. The product recommendation method based on graph decoupling representation learning according to claim 6, characterized in that: Step 6 specifically involves: Step 6.1: When a user generates a shopping request, the product recommendation model based on graph decoupling representation learning generates a corresponding recommendation sequence based on the user's personal preferences and historical consumption behavior. T (20)； (21)； in, Used to predict users For goods The preference values ​​are sorted from highest to lowest based on a specific user's preference values ​​for different products, and the top-ranked products are selected. For each product, generate a recommendation sequence. , The final representation on behalf of the user The final representation of the product, T indicates transpose; Step 6.2: Provide personalized product recommendation results to the user.

8. A product recommendation system based on graph decoupling representation learning, characterized in that, The system is used to implement a product recommendation method based on graph decoupling representation learning as described in any one of claims 1-7, including a computer processor and memory, a user-product interest and popularity decoupling unit, a contrastive learning optimization unit, an adaptive fusion unit, a product recommendation training unit based on graph decoupling representation learning, and a product recommendation result output unit; The user-product interest and popularity decoupling unit executes steps 1 and 2, obtains data from the website, constructs a user-product interaction graph, and quantifies user and product popularity to obtain preliminary decoupled user-product interest and popularity representations. In the contrastive learning optimization unit, step 3 is executed, starting from both the user and product perspectives, using contrastive learning to design feedback signals to enhance the decoupled interest representation and popularity representation; The adaptive fusion unit executes step 4. On the original representation of the user-item interaction graph learned in steps 2 and 3, it automatically calculates the attention weights of interest representation and popularity representation through a learnable attention network and dynamically adjusts the importance of the two types of features based on the intrinsic correlation between features. Finally, it obtains the predicted score through weighted fusion, and obtains the final user-item representation. The product recommendation training unit based on graph decoupling representation learning executes step 5, constructing a product recommendation based on graph decoupling representation learning according to each step from step 1 to step 4, and training the model based on the defined objective function. The product recommendation result output unit executes step 6, which categorizes products with high user interest on the retail platform into recommended categories and provides the product recommendation results to users and merchants on the retail platform.

Images