CN117495511A - A product recommendation system and method based on comparative learning and community perception - Google Patents

Abstract

The present invention proposes a product recommendation system and method based on contrastive learning and community perception. First, an adaptive graph enhancement strategy is designed. When performing data enhancement on the original graph, the importance of nodes and edges is considered and edges with high importance are retained. and node attributes, requiring the enhanced graph to be significantly different from the original graph; secondly, an encoder based on graph neural network and multi-layer perceptron is used to generate the representation vectors of the original graph and the enhanced graph; thirdly, the design is based on the relative distance of nodes. Comparing the selection strategy, select multiple nodes with the closest relative distance to each node as its positive samples, and use the remaining nodes as negative samples; then, use a clustering algorithm to divide the learned node representation vector into communities; finally, Based on the obtained community division results, product recommendations are made; the present invention can improve the accuracy of final personalized product recommendation while improving the cohesion of the generated community structure, thereby improving user satisfaction and product sales.

Description

A product recommendation system and method based on comparative learning and community perception

Technical field

The present invention relates to the technical field of representation learning community discovery, and in particular to a product recommendation system and method based on comparative learning and community perception.

Background technique

With the rapid development of technology and the popularity of the Internet, people's social networking and shopping behaviors have gradually shifted online. People, things, objects and their connections in the real world can be abstracted into complex networks. Community structure characteristics are an important characteristic of complex networks. Community structure refers to a group organization composed of individuals with close connections, frequent interactions, and high similarity. The connections within the community are tight and the connections between communities are sparse. Studying social structure can help us identify user groups with similar interests, occupations and other characteristics, help companies better understand user preferences, provide them with more personalized services and products, and improve user satisfaction. Current research and technology on product recommendation methods based on representation learning community discovery still have the following shortcomings: random data enhancement strategies may interfere with or even destroy the key community structure in the graph, leading to a decrease in community discovery accuracy, which in turn affects the accuracy of product recommendation. .

Contents of the invention

The present invention proposes a product recommendation system and method based on comparative learning and community perception, which can improve the accuracy of final personalized product recommendation while improving the cohesion of the generated community structure, thereby improving user satisfaction and Product sales, specific better value.

The present invention adopts the following technical solutions.

A product recommendation system based on comparative learning and community perception, which is used to recommend products based on community discovery results to help enterprises obtain user preferences, interests and demand information more effectively and accurately, and provide more personalized services and products; The system includes the following steps;

First, design an adaptive graph enhancement strategy. When performing data enhancement on the original graph, consider the importance of nodes and edges and retain the attributes of highly important edges and nodes. At the same time, it is also required that the enhanced graph has a large difference from the original graph. Differences prevent the model from falling into local optimality;

Secondly, an encoder based on graph neural network and multi-layer perceptron is used to generate the representation vectors of the original graph and the enhanced graph;

Thirdly, a comparison selection strategy based on the relative distance of nodes is designed to select multiple nodes with the closest relative distance to each node as its positive samples, and then use the remaining nodes as negative samples to ensure that the generated community structure has a high cohesion;

Next, a clustering algorithm is used to divide the learned node representation vectors into communities;

Finally, based on the obtained community division results, the user's preferences, interests and demand information are obtained, product recommendations within the community and across communities are carried out, and more personalized services are provided to users.

A product recommendation system based on contrastive learning and community perception, including an adaptive graph data enhancement module, a representation vector generation module, a comparison selection module, a loss function calculation module, a community generation module and a product recommendation module;

The adaptive graph data enhancement module is used to generate an enhanced graph G′ that is significantly different from the original graph G; it uses a data enhancement strategy that considers the importance of nodes to avoid destroying the community structure in the graph during the data enhancement process; Perform topology-level data enhancement on the original graph to remove unimportant edges in the social network, and attribute-level data enhancement to shield the attributes of unimportant nodes in the social network; at the same time, the enhanced graph is required to be larger than the original graph. differences to prevent the model from falling into local optimality;

The representation vector generation module is used to encode the original graph G and the enhanced graph G′; an encoder composed of graph convolutional neural network GCN and multi-layer perceptron MLP is used to encode the original graph and the enhanced graph to obtain node representation vectors respectively. z and z′;

The comparison selection module is used to select positive samples and negative samples that are beneficial to improving community discovery accuracy and making community boundaries clearer; multiple nodes with the closest relative distance to the target node are used as their positive samples, and the remaining nodes are used as negative samples. ; Among them, the node relative distance consists of topological distance and attribute distance;

The loss function calculation module is used to calculate the contrast loss And optimize the parameters of GCN and MLP through backpropagation; where contrast loss/> From the loss of node u in the original graph G/> and the loss of node u′ in the enhanced graph G′ composition;

The community generation module is used to generate communities; use the KMeans clustering algorithm to cluster the node representation vector z learned by the encoder to obtain community division results {C _i };

The product recommendation module is used to recommend personalized products to users based on the obtained community division results; it can recommend to users products that other users in the same community like, and can also recommend popular products in the community, and can also recommend products to users. Recommend related products in other communities with strong relevance.

A product recommendation method based on contrastive learning and community perception, using a product recommendation system based on contrastive learning and community perception, including the following steps;

Step S1: Construct _a social network G={ _V , _E , A, The edge set of e _ij = (v _i , v _j )∈E indicates that there is an edge between node v _i and node v _j ; matrix is the adjacency matrix of the network. When e _ij ∈E, A _ij =1, otherwise A _ij =0. /> is the attribute matrix of the node in the social network, m is the dimension of the node attribute, and X _ij represents the value of the j-th dimension attribute of node i;

Step S2: Perform topology-level data enhancement and attribute-level data enhancement on the input graph G respectively. At the same time, the enhanced graph is required to be significantly different from the original graph, and finally the enhanced graph G′ is obtained;

Step S3: Generate the node representation vectors z and z′ of the original graph G and the enhanced graph G′ respectively through a pair of encoders composed of graph convolutional neural GCN and multi-layer perceptron MLP;

Step S4: Calculate the relative distance between the nodes of the original graph G. Based on the calculation results, select the nodes with the closest relative distance for each node as its positive sample set, and use the remaining nodes as the negative sample set;

Step S5: Calculate the contrast loss based on the selected positive sample set and negative sample set And optimize the parameters of GCN and MLP through backpropagation;

Step S6: Use the KMeans clustering algorithm to cluster the node representation vector z learned by the encoder, and use the clusters generated by the clustering as communities to generate community division results {C _i };

Step S7: Combine the community discovery results to provide product recommendation services, including recommending products that other users are interested in in the same community and popular products in other closely related communities, so that product production companies can further understand the needs of target users and achieve more accurate results. product recommendations.

Step S2 is specifically as follows:

Step S21: Perform topology-level data enhancement on the graph G; sample the edges of nodes with higher importance from the original edge set E according to the following formula (1) to form the edge set E′ of the enhanced graph;

in, is the probability of sampling edge (u, v), which is determined by the importance of the edge and is calculated according to formula (2) and formula (3);

in, is the degree centrality of node v, and its size is equal to the degree of the node, that is, /> is the importance of edge (u, v), and its size is equal to the average importance of the two nodes it connects;/> is the maximum value of the importance of all edges in the graph, η ₁ is the coefficient used to control the probability of edge removal, and is a manually specified parameter;

Step S22: After obtaining the edge subset E′ sampled in step S21, convert it into an adjacency matrix A′ for subsequent processes;

Step S23: Perform attribute-level data enhancement on graph G. First, for each node u in the graph G, according to the following formula (4), from the Bernoulli distribution Sample a value with probability, and all sampled values form an n-dimensional vector b∈{0, 1} ⁿ , that is, /> n is the number of nodes in graph G;

is the probability of retaining the attributes of node u. Nodes with high importance have a higher probability of being retained. /> is the importance of node u,/> is the maximum value of the importance of all nodes in the graph, eta ₂ is the coefficient used to control the masking probability of node attributes, and is a manually specified parameter;

Step S24: Based on the vector b obtained in step S23 and the original feature matrix X, calculate the enhanced attribute matrix X′ according to formula (5);

X′＝diag(b)·X Formula (5)

Among them, diag(·) means expanding a vector into a diagonal matrix.

Step S25: Obtain the enhanced graph G′. The enhanced graph can be represented by the enhanced adjacency matrix A′ and attribute matrix X′, that is, G′ = (A′, X′);

Step S26: In order to prevent the model from falling into a local optimal solution because the enhanced graph is too similar to the original graph, thereby affecting the accuracy of the community discovery task, the enhanced graph G′ is required to be significantly different from the original graph G; according to the following According to formula (6), calculate the sum of the adjacency matrix and attribute matrix Euclidean distance between G′ and G;

Among them, n is the number of nodes, m is the dimension of node attributes;

Step S27: The above graph data enhancement steps will be repeated until the sum g of the Euclidean distance of the adjacency matrix and the attribute matrix is greater than the given threshold σ or the number of times of graph data enhancement reaches the maximum number of enhancements I _max .

The step S3 is specifically:

Step S31: Given a network, representation learning refers to learning a transformation function And d＜＜|V|, |V| represents the number of nodes in network G. That is, each node in the network is converted into a d-dimensional representation vector z after passing the f(v) function;

Step S32: Use the encoder f(·) composed of the graph convolutional neural network GCN and the multi-layer perceptron MLP to encode the original graph G and the enhanced graph G′, and obtain the node representation vectors z and z′ respectively for subsequent use. process.

The step S4 is specifically:

Step S41: Calculate the topological distance between target node u and other nodes Assume that the shortest path from node v ₁ to node v _n is P=(v ₁ , v ₂ ,..., v _n )∈V×V×...×V, then the distance between these two nodes/> is the shortest distance between these two nodes; where, /> is the node weight mapping function, where the unit weight f is taken: E→{1}; the shortest distance between nodes is used as the topological distance between nodes;

Step S42: Calculate the attribute distance between nodes according to formula (7)

Among them, x _u and x _v are the attribute vectors of node u and node V respectively;

Step S43: Based on the node topological distance obtained in step S41 and the node attribute distance obtained in step S42 According to formula (8), calculate the relative distance of nodes;

in, and/> Respectively represent the minimum and maximum values of the topological distance of nodes in the graph. /> and Respectively represent the minimum and maximum value of the attribute distance between nodes in the graph. λ is a parameter used to adjust the proportion of node topological distance and attribute distance;

Step S44: Based on the relative distance of the nodes obtained in step S43, sort these nodes in ascending order of relative distance to obtain the node sequence _Ru ;

Step S45: Select the first node in the sequence Nodes serve as the positive sample set P _u of the target node u, and other nodes serve as the negative sample set N _u of the node u, as shown in formula (9) and formula (10).

The step S5 is specifically:

Step S51: Based on the commonly used InfoNCE loss function in comparative learning, calculate the loss function of node u according to formula (11);

Among them, z _u represents the representation vector of node u, z _i and z′ _i are the representation vectors of node i in the original graph G and the enhanced graph G′ respectively, and sim(z _u , z _v ) represents the relationship between nodes u and v. Cosine similarity, τ is the temperature parameter used to adjust the similarity between nodes;

Step S52: According to formula (12), calculate the total loss function of the model, with a size of and Average sum across all nodes.

Step S53: Adjust the parameters of the encoder through backpropagation, and repeat the training process until the model converges.

The specific step S6 is:

Step S61: Randomly select k nodes as initial cluster centers;

Step S62: Calculate the Euclidean distance between each node and the cluster center, and assign each node to the cluster with the nearest cluster center;

Step S63: Update the cluster center of each cluster to the average value of all nodes in the cluster;

Step S64: Repeat steps S62 and S63 until the cluster center no longer changes or the maximum number of iterations is reached, and the community division result {C _i } is obtained.

The step S7 is specifically:

Step S71: According to step S6, users with similar interests and behavior patterns are divided into the same community; based on the community discovery results, recommend to the user products that are liked by other users with strong similarity and correlation in the same community to improve recommendations. accuracy and user satisfaction;

Step S72: Based on the community discovery results, identify products that are particularly popular in a certain community, and recommend popular products to other users in the community to increase user participation and help product manufacturers increase sales;

Step S73: When there are some common points of interest or correlations between different communities, use this correlation to recommend products in other communities with strong correlations to the user, and use cross-community recommendations to help users discover new interests. areas to enrich their shopping experience.

The advantage of the present invention is that the application of the present invention can avoid destroying the community structure in the graph during the data enhancement process, select appropriate positive and negative samples for comparative learning, improve the cohesion of the generated community structure, and ultimately improve product recommendation. accuracy.

Description of the drawings

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments:

Figure 1 is a schematic flow diagram of the present invention.

Detailed ways

As shown in the figure, a product recommendation system based on comparative learning and community perception is used to recommend products based on community discovery results to help enterprises obtain user preferences, interests and demand information more effectively and accurately, and provide more personalized Services and goods; the system includes the following steps;

First, design an adaptive graph enhancement strategy. When performing data enhancement on the original graph, consider the importance of nodes and edges and retain the attributes of highly important edges and nodes. At the same time, it is also required that the enhanced graph has a large difference from the original graph. Differences prevent the model from falling into local optimality;

Secondly, an encoder based on graph neural network and multi-layer perceptron is used to generate the representation vectors of the original graph and the enhanced graph;

Thirdly, a comparison selection strategy based on the relative distance of nodes is designed to select multiple nodes with the closest relative distance to each node as its positive samples, and then use the remaining nodes as negative samples to ensure that the generated community structure has a high cohesion;

Next, a clustering algorithm is used to divide the learned node representation vectors into communities;

Finally, based on the obtained community division results, the user's preferences, interests and demand information are obtained, product recommendations within the community and across communities are made, and more personalized services are provided to the user.

A product recommendation system based on contrastive learning and community perception, including an adaptive graph data enhancement module, a representation vector generation module, a comparison selection module, a loss function calculation module, a community generation module and a product recommendation module;

The adaptive graph data enhancement module is used to generate an enhanced graph G′ that is significantly different from the original graph G; it uses a data enhancement strategy that considers the importance of nodes to avoid destroying the community structure in the graph during the data enhancement process; Perform topology-level data enhancement on the original graph to remove unimportant edges in the social network, and attribute-level data enhancement to shield the attributes of unimportant nodes in the social network; at the same time, the enhanced graph is required to be larger than the original graph. differences to prevent the model from falling into local optimality;

The representation vector generation module is used to encode the original graph G and the enhanced graph G′; an encoder composed of graph convolutional neural network GCN and multi-layer perceptron MLP is used to encode the original graph and the enhanced graph to obtain node representation vectors respectively. z and z′;

The comparison selection module is used to select positive samples and negative samples that are beneficial to improving community discovery accuracy and making community boundaries clearer; multiple nodes with the closest relative distance to the target node are used as their positive samples, and the remaining nodes are used as negative samples. ; Among them, the node relative distance consists of topological distance and attribute distance;

The loss function calculation module is used to calculate the contrast loss And optimize the parameters of GCN and MLP through backpropagation; where contrast loss/> From the loss of node u in the original graph G/> and the loss of node u′ in the enhanced graph G′ composition;

The community generation module is used to generate communities; use the KMeans clustering algorithm to cluster the node representation vector z learned by the encoder to obtain community division results {C _i };

The product recommendation module is used to recommend personalized products to users based on the obtained community division results; it can recommend to users products that other users in the same community like, and can also recommend popular products in the community, and can also recommend products to users. Recommend related products in other communities with strong relevance.

A product recommendation method based on contrastive learning and community perception, using a product recommendation system based on contrastive learning and community perception, including the following steps;

Step S1: Construct _a social network G={ _V , _E , A, The edge set of e _ij = (v _i , v _j )∈E indicates that there is an edge between node v _i and node v _j ; matrix is the adjacency matrix of the network. When e _ij ∈E, A _ij =1, otherwise A _ij =0. /> is the attribute matrix of the node in the social network, m is the dimension of the node attribute, and X _ij represents the value of the j-th dimension attribute of node i;

Step S2: Perform topology-level data enhancement and attribute-level data enhancement on the input graph G respectively. At the same time, the enhanced graph is required to be significantly different from the original graph, and finally the enhanced graph G′ is obtained;

Step S3: Generate the node representation vectors z and z′ of the original graph G and the enhanced graph G′ respectively through a pair of encoders composed of graph convolutional neural GCN and multi-layer perceptron MLP;

Step S4: Calculate the relative distance between the nodes of the original graph G. Based on the calculation results, select the nodes with the closest relative distance for each node as its positive sample set, and use the remaining nodes as the negative sample set;

Step S5: Calculate the contrast loss based on the selected positive sample set and negative sample set And optimize the parameters of GCN and MLP through backpropagation;

Step S6: Use the KMeans clustering algorithm to cluster the node representation vector z learned by the encoder, and use the clusters generated by the clustering as communities to generate community division results {C _i };

Step S7: Combine the community discovery results to provide product recommendation services, including recommending products that other users are interested in in the same community and popular products in other closely related communities, so that product production companies can further understand the needs of target users and achieve more accurate results. product recommendations.

Step S2 is specifically as follows:

Step S21: Perform topology-level data enhancement on the graph G; sample the edges of nodes with higher importance from the original edge set E according to the following formula (1) to form the edge set E′ of the enhanced graph;

in, is the probability of sampling edge (u, v), which is determined by the importance of the edge and is calculated according to formula (2) and formula (3);

in, is the degree centrality of node v, and its size is equal to the degree of the node, that is, /> is the importance of edge (u, v), and its size is equal to the average importance of the two nodes it connects;/> is the maximum value of the importance of all edges in the graph, η ₁ is the coefficient used to control the probability of edge removal, and is a manually specified parameter;

Step S22: After obtaining the edge subset E′ sampled in step S21, convert it into an adjacency matrix A′ for subsequent processes;

Step S23: Perform attribute-level data enhancement on graph G. First, for each node u in the graph G, according to the following formula (4), from the Bernoulli distribution Sample a value with probability, and all sampled values form an n-dimensional vector b∈{0, 1} ⁿ , that is, /> n is the number of nodes in graph G;

is the probability of retaining the attributes of node u. Nodes with high importance have a higher probability of being retained. /> is the importance of node u,/> is the maximum value of the importance of all nodes in the graph, eta ₂ is the coefficient used to control the masking probability of node attributes, and is a manually specified parameter;

Step S24: Based on the vector b obtained in step S23 and the original feature matrix X, calculate the enhanced attribute matrix X′ according to formula (5);

X′＝diag(b)·X Formula (5)

Among them, diag(·) means expanding a vector into a diagonal matrix.

Step S25: Obtain the enhanced graph G′. The enhanced graph can be represented by the enhanced adjacency matrix A′ and attribute matrix X′, that is, G′ = (A′, X′);

Step S26: In order to prevent the model from falling into a local optimal solution because the enhanced graph is too similar to the original graph, thereby affecting the accuracy of the community discovery task, the enhanced graph G′ is required to be significantly different from the original graph G; according to the following According to formula (6), calculate the sum of the adjacency matrix and attribute matrix Euclidean distance between G′ and G;

Among them, n is the number of nodes, m is the dimension of node attributes;

Step S27: The above graph data enhancement steps will be repeated until the sum g of the Euclidean distance of the adjacency matrix and the attribute matrix is greater than the given threshold σ or the number of times of graph data enhancement reaches the maximum number of enhancements I _max .

The specific step S3 is:

Step S31: Given a network, representation learning refers to learning a transformation function And d＜＜|V|, |V| represents the number of nodes in network G. That is, each node in the network is converted into a d-dimensional representation vector z after passing the f(v) function;

Step S32: Use the encoder f(·) composed of the graph convolutional neural network GCN and the multi-layer perceptron MLP to encode the original graph G and the enhanced graph G′, and obtain the node representation vectors z and z′ respectively for subsequent use. process.

The step S4 is specifically:

Step S41: Calculate the topological distance between target node u and other nodes Assume that the shortest path from node v ₁ to node v _n is P = (v ₁ , v _{2 ,} ..., v _n )∈V×V×...×V, then the distance between these two nodes/> is the shortest distance between these two nodes; where, /> is the node weight mapping function, where the unit weight f is taken: E→{1}; the shortest distance between nodes is used as the topological distance between nodes;

Step S42: Calculate the attribute distance between nodes according to formula (7)

Among them, x _u and x _v are the attribute vectors of node u and node v respectively;

Step S43: Based on the node topological distance obtained in step S41 and the node attribute distance obtained in step S42 According to formula (8), calculate the relative distance of nodes;

in, and/> Respectively represent the minimum and maximum values of the topological distance of nodes in the graph. /> and Respectively represent the minimum and maximum value of the attribute distance between nodes in the graph. λ is a parameter used to adjust the proportion of node topological distance and attribute distance;

Step S44: Based on the relative distance of the nodes obtained in step S43, sort these nodes in ascending order of relative distance to obtain the node sequence _Ru ;

Step S45: Select the first node in the sequence Nodes serve as the positive sample set P _u of the target node u, and other nodes serve as the negative sample set N _u of the node u, as shown in formula (9) and formula (10).

The step S5 is specifically:

Step S51: Based on the commonly used InfoNCE loss function in comparative learning, calculate the loss function of node u according to formula (11);

Among them, z _u represents the representation vector of node u, z _i and z′ _i are the representation vectors of node i in the original graph G and the enhanced graph G′ respectively, and sim(z _u , z _v ) represents the relationship between nodes u and v. Cosine similarity, τ is the temperature parameter used to adjust the similarity between nodes;

Step S52: According to formula (12), calculate the total loss function of the model, with a size of and Average sum across all nodes.

Step S53: Adjust the parameters of the encoder through backpropagation, and repeat the training process until the model converges.

The specific step S6 is:

Step S61: Randomly select k nodes as initial cluster centers;

Step S62: Calculate the Euclidean distance between each node and the cluster center, and assign each node to the cluster with the nearest cluster center;

Step S63: Update the cluster center of each cluster to the average value of all nodes in the cluster;

Step S64: Repeat steps S62 and S63 until the cluster center no longer changes or the maximum number of iterations is reached, and the community division result {C _i } is obtained.

The step S7 is specifically:

Step S71: According to step S6, users with similar interests and behavior patterns are divided into the same community; based on the community discovery results, recommend to the user products that are liked by other users with strong similarity and correlation in the same community to improve recommendations. accuracy and user satisfaction;

Step S72: Based on the community discovery results, identify products that are particularly popular in a certain community, and recommend popular products to other users in the community to increase user participation and help product manufacturers increase sales;

Step S73: When there are some common points of interest or correlations between different communities, use this correlation to recommend products in other communities with strong correlations to the user, and use cross-community recommendations to help users discover new interests. areas to enrich their shopping experience.

Claims (9)

1. A product recommendation system based on comparative learning and community perception, which is used to recommend products based on community discovery results to help enterprises obtain user preferences, interests and demand information more effectively and accurately, and provide more personalized services and Commodity; characterized in that: the system includes the following steps; First, design an adaptive graph enhancement strategy. When performing data enhancement on the original graph, consider the importance of nodes and edges and retain the attributes of highly important edges and nodes. At the same time, it is also required that the enhanced graph has a large difference from the original graph. Differences prevent the model from falling into local optimality; Secondly, an encoder based on graph neural network and multi-layer perceptron is used to generate the representation vectors of the original graph and the enhanced graph; Thirdly, a comparison selection strategy based on the relative distance of nodes is designed to select multiple nodes with the closest relative distance to each node as its positive samples, and then use the remaining nodes as negative samples to ensure that the generated community structure has a high cohesion; Next, a clustering algorithm is used to divide the learned node representation vectors into communities; Finally, based on the obtained community division results, the user's preferences, interests and demand information are obtained, product recommendations within the community and across communities are carried out, and more personalized services are provided to users. 2. A product recommendation system based on contrastive learning and community perception according to claim 1, characterized by: including an adaptive graph data enhancement module, a representation vector generation module, a comparison selection module, a loss function calculation module, a community Generation module and product recommendation module; The adaptive graph data enhancement module is used to generate an enhanced graph G′ that is significantly different from the original graph G; it uses a data enhancement strategy that considers the importance of nodes to avoid destroying the community structure in the graph during the data enhancement process; Perform topology-level data enhancement on the original graph to remove unimportant edges in the social network, and attribute-level data enhancement to shield the attributes of unimportant nodes in the social network; at the same time, the enhanced graph is required to be larger than the original graph. differences to prevent the model from falling into local optimality; The representation vector generation module is used to encode the original graph G and the enhanced graph G′; an encoder composed of graph convolutional neural network GCN and multi-layer perceptron MLP is used to encode the original graph and the enhanced graph to obtain node representation vectors respectively. z and z′; The comparison selection module is used to select positive samples and negative samples that are beneficial to improving community discovery accuracy and making community boundaries clearer; multiple nodes with the closest relative distance to the target node are used as their positive samples, and the remaining nodes are used as negative samples. ; Among them, the node relative distance consists of topological distance and attribute distance; The loss function calculation module is used to calculate the contrast loss And optimize the parameters of GCN and MLP through backpropagation; where contrast loss/> From the loss of node u in the original graph G/> and the loss of node u′ in the enhanced graph G′ composition; The community generation module is used to generate communities; use the KMeans clustering algorithm to cluster the node representation vector z learned by the encoder to obtain community division results {C _i }; The product recommendation module is used to recommend personalized products to users based on the obtained community division results; it can recommend to users products that other users in the same community like, and can also recommend popular products in the community, and can also recommend products to users. Recommend related products in other communities with strong relevance. 3. A product recommendation method based on comparative learning and community perception, using a product recommendation system based on comparative learning and community perception, which is characterized by: including the following steps; Step _Sl : Construct _a social network G={ _V , E, A, The edge set of e _ij = (v _i , v _j )∈E indicates that there is an edge between node v _i and node v _j ; matrix is the adjacency matrix of the network. When e _ij ∈E, A _ij =1, otherwise A _ij =0. /> is the attribute matrix of the node in the social network, m is the dimension of the node attribute, and X _ij represents the value of the j-th dimension attribute of node i; Step S2: Perform topology-level data enhancement and attribute-level data enhancement on the input graph G respectively. At the same time, the enhanced graph is required to be significantly different from the original graph, and finally the enhanced graph G′ is obtained; Step S3: Generate the node representation vectors z and z′ of the original graph G and the enhanced graph G′ respectively through a pair of encoders composed of graph convolutional neural GCN and multi-layer perceptron MLP; Step S4: Calculate the relative distance between the nodes of the original graph G. Based on the calculation results, select the nodes with the closest relative distance for each node as its positive sample set, and use the remaining nodes as the negative sample set; Step S5: Calculate the contrast loss based on the selected positive sample set and negative sample set And optimize the parameters of GCN and MLP through backpropagation; Step S6: Use the KMeans clustering algorithm to cluster the node representation vector z learned by the encoder, and use the clusters generated by the clustering as communities to generate community division results {C _i }; Step S7: Combine the community discovery results to provide product recommendation services, including recommending products that other users are interested in in the same community and popular products in other closely related communities, so that product production companies can further understand the needs of target users and achieve more accurate results. product recommendations. 4. A product recommendation method based on comparative learning and community perception according to claim 3, characterized in that step S2 is specifically: Step S21: Perform topology-level data enhancement on the graph G; sample the edges of nodes with higher importance from the original edge set E according to the following formula (1) to form the edge set E′ of the enhanced graph; in, is the probability of sampling edge (u, v), which is determined by the importance of the edge and is calculated according to formula (2) and formula (3); in, is the degree centrality of node v, and its size is equal to the degree of the node, that is, /> is the importance of edge (u, v), and its size is equal to the average importance of the two nodes it connects;/> is the maximum value of the importance of all edges in the graph, η ₁ is the coefficient used to control the probability of edge removal, and is a manually specified parameter; Step S22: After obtaining the edge subset E′ sampled in step S21, convert it into an adjacency matrix A′ for subsequent processes; Step S23: Perform attribute-level data enhancement on graph G. First, for each node u in the graph G, according to the following formula (4), from the Bernoulli distribution Sample a value with probability, and all sampled values form an n-dimensional vector b∈{0, 1} ⁿ , that is, /> n is the number of nodes in graph G; is the probability of retaining the attributes of node u. Nodes with high importance have a higher probability of being retained. /> is the importance of node u,/> is the maximum value of the importance of all nodes in the graph, eta ₂ is the coefficient used to control the masking probability of node attributes, and is a manually specified parameter; Step S24: Based on the vector b obtained in step S23 and the original feature matrix X, calculate the enhanced attribute matrix X′ according to formula (5); X′＝diag(b)·X Formula (5) Among them, diag(·) means expanding a vector into a diagonal matrix. Step S25: Obtain the enhanced graph G′. The enhanced graph can be represented by the enhanced adjacency matrix A′ and attribute matrix X′, that is, G′ = (A′, X′); Step S26: In order to prevent the model from falling into a local optimal solution because the enhanced graph is too similar to the original graph, thereby affecting the accuracy of the community discovery task, the enhanced graph G′ is required to be significantly different from the original graph G; according to the following According to formula (6), calculate the sum of the adjacency matrix and attribute matrix Euclidean distance between G′ and G; Among them, n is the number of nodes, m is the dimension of node attributes; Step S27: The above graph data enhancement steps will be repeated until the sum g of the Euclidean distance of the adjacency matrix and the attribute matrix is greater than the given threshold σ or the number of times of graph data enhancement reaches the maximum number of enhancements I _max . 5. A product recommendation method based on comparative learning and community perception according to claim 3, characterized in that: the step S3 is specifically: Step S31: Given a network, representation learning refers to learning a transformation function f(v): v→z, v∈V, and d<<|V|, |V| represents the number of nodes of network G. That is, each node in the network is converted into a d-dimensional representation vector z after passing the f(v) function; Step S32: Use the encoder f(·) composed of the graph convolutional neural network GCN and the multi-layer perceptron MLP to encode the original graph G and the enhanced graph G′, and obtain the node representation vectors z and z′ respectively for subsequent use. process. 6. A product recommendation method based on comparative learning and community perception according to claim 3, characterized in that: the step S4 is specifically: Step S41: Calculate the topological distance between target node u and other nodes Assume that the shortest path from node v ₁ to node v _n is P=(v ₁ , v ₂ ,..., v _n )∈V×V×...×V, then the distance between these two nodes/> is the shortest distance between these two nodes; where, /> is the node weight mapping function, where the unit weight f is taken: E→{1}; the shortest distance between nodes is used as the topological distance between nodes; Step S42: Calculate the attribute distance between nodes according to formula (7) Among them, x _u and x _v are the attribute vectors of node u and node v respectively; Step S43: Based on the node topological distance obtained in step S41 and the node attribute distance obtained in step S42/> According to formula (8), calculate the relative distance of nodes; in, and/> Respectively represent the minimum and maximum values of the topological distance of nodes in the graph. /> and/> Respectively represent the minimum and maximum value of the attribute distance between nodes in the graph. λ is a parameter used to adjust the proportion of node topological distance and attribute distance; Step S44: Based on the relative distance of the nodes obtained in step S43, sort these nodes in ascending order of relative distance to obtain the node sequence _Ru ; Step S45: Select the first node in the sequence Nodes serve as the positive sample set P _u of the target node u, and other nodes serve as the negative sample set N _u of the node u, as shown in formula (9) and formula (10). 7. A product recommendation method based on comparative learning and community perception according to claim 3, characterized in that: the step S5 is specifically: Step S51: Based on the commonly used InfoNCE loss function in comparative learning, calculate the loss function of node u according to formula (11); Among them, z _u represents the representation vector of node u, z _i and z′ _i are the representation vectors of node i in the original graph G and the enhanced graph G′ respectively, and sim(z _u , z _v ) represents the relationship between nodes u and v. Cosine similarity, τ is the temperature parameter used to adjust the similarity between nodes; Step S52: According to formula (12), calculate the total loss function of the model, with a size of and/> Average sum across all nodes. Step S53: Adjust the parameters of the encoder through backpropagation, and repeat the training process until the model converges. 8. A product recommendation method based on comparative learning and community perception according to claim 3, characterized in that: the step S6 is specifically: Step S61: Randomly select k nodes as initial cluster centers; Step S62: Calculate the Euclidean distance between each node and the cluster center, and assign each node to the cluster with the nearest cluster center; Step S63: Update the cluster center of each cluster to the average value of all nodes in the cluster; Step S64: Repeat steps S62 and S63 until the cluster center no longer changes or the maximum number of iterations is reached, and the community division result {C _i } is obtained. 9. A product recommendation method based on comparative learning and community perception according to claim 3, characterized in that: the step S7 is specifically: Step S71: According to step S6, users with similar interests and behavior patterns are divided into the same community; based on the community discovery results, recommend to the user products that are liked by other users with strong similarity and correlation in the same community to improve recommendations. accuracy and user satisfaction; Step S72: Based on the community discovery results, identify products that are particularly popular in a certain community, and recommend popular products to other users in the community to increase user participation and help product manufacturers increase sales; Step S73: When there are some common points of interest or correlations between different communities, use this correlation to recommend products in other communities with strong correlations to the user, and use cross-community recommendations to help users discover new interests. areas to enrich their shopping experience.