Recommendation method based on parallel graph neural network
By using technical means such as parallel graph neural network and lightweight graph convolution module in the recommendation system, the existing recommendation methods are solved inadequate recommendation efforts for long-tail items and homogeneous recommendation results, achieving more personalized and high-quality recommendation effects.
Patent Information
- Application Number
- CN202411693861.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-03-14
AI Technical Summary
The existing recommendation methods perform well when recommending popular items, but the recommendations for long-tail items are insufficient and the unique differences between items are ignored, resulting in homogeneous recommendation results and it is difficult to meet the diverse needs of users.
Using a recommendation method based on parallel graph neural network, the vector representation of users and projects is updated through lightweight graph convolution module and knowledge graph information, combined with the attention mechanism, the user's interaction probability of the project is calculated, and personalized recommendation is provided.
On the MovieLens-20M, Last-FM and Book-Crossing datasets, the recall rate and normalized loss cumulative gain are significantly improved, the performance and user experience of the recommendation system are improved, and long-tail items can be better recommended and data sparseness and insufficient recommendation explanations can be alleviated.
Smart Images
Figure CN119622090A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent recommendation technology, and in particular to a recommendation method based on parallel graph neural network. Background Art
[0002] In the era of information explosion, the rapid growth of information has far exceeded human processing capabilities, making information overload an urgent problem to be solved. Recommendation systems have emerged as a result. They are like a key that unlocks the way for users to quickly find the content they need from massive amounts of information. They are widely used in news, movies, commodities and other fields. Given the profound impact of the big data environment, it is particularly important to build an intelligent recommendation system that can accurately and personally meet user needs.
[0003] However, existing recommendation methods have some shortcomings. They face challenges such as cold start, data sparsity, and lack of recommendation interpretability. In particular, these methods often perform well in recommending popular items, but are unable to recommend long-tail items. In addition, they recommend similar items based on users' historical behavior data, but ignore the unique differences between items, resulting in homogenized recommendation results that are difficult to meet the diverse needs of users.
[0004] Therefore, in order to overcome these limitations, we need an intelligent recommendation system that can span multiple scenarios. Such a system should be committed to providing personalized recommendation services while enhancing the interpretability of recommendation results so that users can understand the logic behind the recommendation, thereby improving the quality of recommendations and optimizing user experience. Summary of the invention
[0005] The present invention aims to provide a recommendation method based on parallel graph neural network to solve the problems existing in the above-mentioned prior art.
[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a recommendation method based on parallel graph neural network, comprising the following steps:
[0007] S1, input user information, project information, and knowledge graph information into the initialization vector layer, initialize users, projects, and items into 64-dimensional vector representations, and the user vector is The item vector is
[0008] The user initialized in S1 and the item vector Input the Light graph convolutional network (LGC) to get the user and Projects vector;
[0009] S3, input the initial user vector and item vector into the lightweight graph convolution module (LGC) to calculate the user and Projects vector; and use the information propagation rules of this layer to update the vector representation of users and items;
[0010] S4, by repeating steps S2 and S3 multiple times, multiple user and item vector representations are obtained, and the user and item vectors are weighted and combined respectively to obtain the user and item vectors of this stage;
[0011] S5, calculate the two sets of project and user vectors obtained in S4 to get the final user and project vector representation. The final user vector is The final item vector is
[0012] S6, performing a dot product operation on the user and item vectors obtained in S5 to obtain the user's interaction probability with the item, and returning the recommendation result.
[0013] Furthermore, for each user e in S1 u ∈R d (project, relationship i ,e r ∈R d ) assigns an initial embedding vector: The initialization vector is 64-dimensional.
[0014] Further, step S2 specifically includes:
[0015] S2.1, construct the user-item bipartite graph using the user-item interaction matrix;
[0016] Establishment of user-project interaction matrix: Monitor user activities on the platform through user behavior logs, user feedback surveys, user purchase history, and user social media data methods, including clicks, browsing, purchases, ratings, comments, sharing, and preferences, and organize various interactive behaviors or feedback data between users and projects; mark positive interactive behaviors between user projects as 1, and non-positive interactive behaviors as 0, to form a user-project interaction matrix;
[0017] According to the above user-item interaction matrix, a user-item bipartite graph is constructed; specifically, based on the user-item interaction matrix, user nodes and item nodes are created; each user and item is represented as a node in the graph, ensuring that each node has a unique identifier; next, edges are added to the bipartite graph according to the user-item interaction matrix; generally, if there is interaction between a user and an item, an edge is added between them;
[0018] In this graph, users and items form two non-overlapping node sets respectively; each user node corresponds to a row of the matrix, and each item node corresponds to a column of the matrix; then, for users and items with an interaction record of 1 in the user-item interaction matrix, an edge with a weight of 1 is added between the two.
[0019] S2.2, obtain the number of project nodes around a certain user through the bipartite graph Obtain the number of user nodes around a project through a bipartite graph
[0020] S2.3, using the results of step 2.2, update the vector representation of users and items, specifically including:
[0021] pass Update the vector representation of the user;
[0022] pass Update the vector representation of the item;
[0023] Multiple user vectors are passed through Get user vector
[0024] Pass multiple items to the vector Get user vector
[0025] Furthermore, S3 will initialize the user in step 1 and Projects vector, and the relationship initialized in step 1 The vector is input into the knowledge-aware layer of attention and graph convolutional networks (KAGC) to obtain the user and Projects vector.
[0026] Further, in S4, the above steps S2 and S3 are repeated L times to obtain the first group of users and Projects Vector and the second group of users and Projects Vector, calculate the first set of vectors above to get Vector Sum Vector, the second set of vectors is calculated Vector Sum L is a natural number greater than 0.
[0027] Further, in S5, the two sets of user and item vectors obtained in S4 are combined to obtain user and Projects vector.
[0028] Further, S3 specifically includes the following steps:
[0029] S3.1, input the knowledge graph vector into the graph neural network to obtain the attention representation, which specifically includes:
[0030] The user Relationship with initialization Performing dot product operation yields
[0031] The project Relationship with initialization Performing dot product operation yields
[0032] The above and pass Weighted
[0033] Will After normalization, we get
[0034] pass Get the neighbor vector representation of the item;
[0035] S3.2, the user vector is processed to obtain the self-attention representation, which specifically includes:
[0036] By User Vector, use the following formula to calculate three vectors, W q , W k , W v is the feature transformation matrix:
[0037]
[0038] By Q u With K u The vector dot product of
[0039] Will After normalization, we get
[0040] pass Get the user's self-attention vector representation;
[0041] S3.3, the item vector is updated through the attention representation in step 3.1, and the user vector is updated through the self-attention representation in step 3.2, specifically including:
[0042] pass Get the vector representation of the item pass Get the vector representation of the user
[0043] Further, step S4 specifically includes:
[0044] Using step S2, multiple user vectors are obtained
[0045] Use step S2 to get multiple item vectors
[0046] Multiple user vectors are passed through Get user vector
[0047] Pass multiple items to the vector Get user vector
[0048] Using step S3, multiple user vectors are obtained
[0049] Using step S3, multiple item vectors are obtained
[0050] Multiple user vectors are passed through Get user vector
[0051] Pass multiple items to the vector Get user vector
[0052] Further, in step S5;
[0053]
[0054] In step S6;
[0055]
[0056] Represents the interaction probability between a user and an item.
[0057] Beneficial effects of the present invention:
[0058] 1. The proposed method consistently produces the best recommendation performance on the MovieLens-20M, Last-FM, and Book-Crossing datasets. On the MovieLens-20M dataset, the recall@20 and normalized discounted cumulative gain@20 are improved by 6.65% and 3.63% respectively compared to the latest and best performing baseline LighterKGCN method. On the Last-FM dataset, they are improved by 4.73% and 19.43% respectively, and on the Book-Crossing dataset, they are improved by 14.65% and 18.20% respectively. Recall is a measure of how well a recommendation system can incorporate user interests into the recommendation list when these interests are known. An increase in recall indicates that the system can more comprehensively cover items that users may be interested in, reducing the number of cases where user interests are missed, which is very important to ensure that users do not miss content that they may like.
[0059] 2. Normalized Cumulative Loss Gain is a metric for evaluating the ranking quality of a recommendation system, taking into account the ranking and relevance of items in the recommendation list. Normalized Cumulative Loss Gain Improvement means that the system not only provides items with high relevance, but also ranks them higher, ensuring that users can more easily see and interact with these items, which helps improve user satisfaction and click-through rate.
[0060] 3. The simultaneous improvement of recall rate and normalized discounted cumulative gain indicates that the recommendation performance of the method of the present invention has achieved a good balance: the method of the present invention not only covers user interests as much as possible, but also presents these interests in a more relevant way, providing more attractive personalized recommendations, which helps to improve user experience and increase user trust in the recommendation system, thereby promoting user participation and interaction.
[0061] 4. The method of the present invention is good at incorporating complementary information in the item knowledge graph into the item vector representation, and deeply understands the unique requirements and potential interests of different users; therefore, it can produce better recommendations, thereby improving the performance of the recommendation system. At the same time, because the method incorporates the item knowledge graph and attention mechanism, it can well alleviate the problem of data sparsity and insufficient explanation of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 This is a comparison chart of model performance under different training set ratios;
[0063] Figure 2 A schematic diagram of a flow chart of a recommendation method based on a parallel graph neural network of the present invention;
[0064] Figure 3 A schematic diagram of the architecture of a recommendation method based on a parallel graph neural network of the present invention;
[0065] Figure 4 This is a comparative experimental data table. DETAILED DESCRIPTION
[0066] Example 1 is as shown in the attached Figure 2 As shown:
[0067] A recommendation method based on a parallel graph neural network includes the following steps:
[0068] S1, for each user e u ∈R d (project, relationship i ,e r ∈R d ) assigns an initial embedding vector: The initialization vector is 64-dimensional.
[0069] S2, the user initialized in S1 and the item vector Input the Light graph convolutional network (LGC) to get the user and Projects Vectors, including:
[0070] S2.1, construct the user-item bipartite graph using the user-item interaction matrix;
[0071] Establishment of user-project interaction matrix: Monitor user activities on the platform through user behavior logs, user feedback surveys, user purchase history, and user social media data methods, including clicks, browsing, purchases, ratings, comments, sharing, and preferences, and organize various interactive behaviors or feedback data between users and projects; mark positive interactive behaviors between user projects as 1, and non-positive interactive behaviors as 0, to form a user-project interaction matrix;
[0072] According to the above user-item interaction matrix, a user-item bipartite graph is constructed; specifically, based on the user-item interaction matrix, user nodes and item nodes are created; each user and item is represented as a node in the graph, ensuring that each node has a unique identifier; next, edges are added to the bipartite graph according to the user-item interaction matrix; generally, if there is an interaction between a user and an item, an edge is added between them; for example, if user A rates item X, an edge is added between the user A node and the item X node. For each edge, a weight can be assigned, which reflects the strength or importance of the interaction;
[0073] In this graph, users and items form two non-overlapping node sets respectively; each user node corresponds to a row of the matrix, and each item node corresponds to a column of the matrix; then, for users and items with an interaction record of 1 in the user-item interaction matrix, an edge with a weight of 1 is added between the two.
[0074] S2.2, obtain the number of project nodes around a certain user through the bipartite graph Obtain the number of user nodes around a project through a bipartite graph
[0075] S2.3, using the results of step 2.2, update the vector representation of users and items, specifically including:
[0076] pass Update the vector representation of the user;
[0077] pass Update the vector representation of the item;
[0078] Multiple user vectors are passed through Get user vector
[0079] Pass multiple items to the vector Get user vector
[0080] S3, the user initialized in step 1 and Projects vector, and the relationship initialized in step 1 The vector is input into the knowledge-aware layer of attention and graph convolutional networks (KAGC) to obtain the user and Projects Vector. Specifically include:
[0081] S3.1, input the knowledge graph vector into the graph neural network to obtain the attention representation, which includes:
[0082] The user Relationship with initialization Performing dot product operation yields
[0083] The project Relationship with initialization Performing dot product operation yields
[0084] The above and pass Weighted
[0085] Will After normalization, we get
[0086] pass Get the neighbor vector representation of the item;
[0087] S3.2, the user vector is processed to obtain the self-attention representation, which specifically includes:
[0088] By User Vector, use the following formula to calculate three vectors, W q , W k , W v is the feature transformation matrix:
[0089]
[0090] By Q u With K u The vector dot product of
[0091] Will After normalization, we get
[0092] pass Get the user's self-attention vector representation;
[0093] S3.3, the item vector is updated through the attention representation in step 3.1, and the user vector is updated through the self-attention representation in step 3.2, specifically including:
[0094] pass Get the vector representation of the item
[0095] pass Get the vector representation of the user
[0096] S4, by repeating steps S2 and S3 multiple times, a total of L times, obtain the first group of users and Projects Vector and the second group of users and Projects Vector, calculate the first set of vectors above to get Vector Sum Vector, the second set of vectors is calculated Vector Sum L is a natural number greater than 0. Specifically including:
[0097] Using step S2, multiple user vectors are obtained
[0098] Use step S2 to get multiple item vectors
[0099] Multiple user vectors are passed through Get user vector
[0100] Pass multiple items to the vector Get user vector
[0101] Using step S3, multiple user vectors are obtained
[0102] Using step S3 to obtain multiple item vectors
[0103] Multiple user vectors are passed through Get user vector
[0104] Pass multiple items to the vector Get user vector
[0105] S5, calculate the two sets of project and user vectors obtained in S4 to get the final user and project vector representation. The final user vector is The final item vector is
[0106]
[0107] S6, performing a dot product operation on the user and item vectors obtained in S5 to obtain the user's interaction probability with the item, and returning the recommendation result;
[0108]
[0109] Represents the interaction probability between a user and an item.
[0110] Comparative experiment as attached Figure 4 As shown:
[0111] As shown in Table 1, the method of the present invention always produces the best recommendation performance on the MovieLens-20M dataset, the Last-FM dataset, and the Book-Crossing dataset. On the MovieLens-20M dataset, the recall @20 and the normalized discounted cumulative gain @20 are improved by 6.65% and 3.63% respectively compared with the latest and best baseline LighterKGCN method. On the Last-FM dataset, they are improved by 4.73% and 19.43% respectively, and on the Book-Crossing dataset, they are improved by 14.65% and 18.20% respectively. Recall is a measure of the extent to which a recommendation system can incorporate user interests into the recommendation list when these interests are known. An increase in recall indicates that the system can more comprehensively cover items that users may be interested in, reducing the number of cases where user interests are missed, which is very important to ensure that users do not miss content that they may like.
[0112] Normalized Cumulative Loss Gain is a metric for evaluating the ranking quality of a recommendation system, taking into account the ranking and relevance of items in the recommendation list. Normalized Cumulative Loss Gain Lift indicates that the system not only provides items with high relevance, but also ranks them higher, ensuring that users are more likely to see and interact with these items, which helps improve user satisfaction and click-through rate.
[0113] Therefore, the simultaneous improvement of recall rate and normalized discounted cumulative gain indicates that the recommendation performance of the method of the present invention has achieved a good balance: the method of the present invention not only covers user interests as much as possible, but also presents these interests in a more relevant way, providing more attractive personalized recommendations, which helps to improve user experience and increase user trust in the recommendation system, thereby promoting user participation and interaction.
[0114] This shows that the proposed method is good at incorporating complementary information in the item knowledge graph into the item vector representation and gaining a deep understanding of the unique requirements and potential interests of different users; therefore, it can generate better recommendations, thereby improving the performance of the recommendation system. At the same time, since the method incorporates the item knowledge graph and attention mechanism, it can effectively alleviate the data sparsity problem and the problem of insufficient explanation of the recommendation system.
[0115] The training set ratio represents the ratio of the training set to the original training set. The ratio of the training set to the original training set is changed from 1.0 to 0.2, and the number of test sets is always kept fixed, as shown in the attached figure. Figure 1 shown.
[0116] The experimental results show that the recommendation method of the present invention can maintain excellent performance in cold start conditions. In book and music recommendations, the recommendation method of the present invention shows the best performance in all sparsities, indicating that the model can effectively alleviate the cold start problem of the recommendation system.
[0117] Example 2
[0118] As attached Figure 3 As shown, the steps of recommending in the recommendation method based on parallel graph neural network include:
[0119] 1. Input user information, project information, and knowledge graph information into the initialization vector layer, initialize users, projects, and items into 64-dimensional vector representations, and the user vector is The item vector is
[0120] 2. Input the initial user vector and item vector into the attention and graph convolutional network module (KAGC) to calculate the user and Projects vector, and use the information propagation rules of this layer to update the vector representation of users and items;
[0121] 3. Input the initial user vector and item vector into the lightweight graph convolution module (LGC) to calculate the user and Projects vector; and use the information propagation rules of this layer to update the vector representation of users and items;
[0122] 4. By repeating steps 2 and 3 multiple times, multiple user and item vector representations are obtained, and then the user-item vector representations of different layers are aggregated to obtain the final user and item vector representation. The final user vector is The final item vector is
[0123]
[0124] 5. Get the final recommendation prediction through vector dot product operation
[0125] 6. Train and optimize the model through the BPR loss function.
Claims
1. A recommendation method based on parallel graph neural network, characterized in that: The following steps are involved: S1, input user information, project information, and knowledge graph information into the initialization vector layer, initialize users, projects, and items into 64-dimensional vector representations, and the user vector is The item vector is The user initialized in S1 and the item vector Input lightweight graph convolution module to get user and Projects vector; S3, input the initial user vector and item vector into the lightweight graph convolution module (LGC) to calculate the user and Projects vector; and use the information propagation rules of this layer to update the vector representation of users and items; S4, by repeating steps S2 and S3 multiple times, multiple user and item vector representations are obtained, and the user and item vectors are weighted and combined respectively to obtain the user and item vectors of this stage; S5, calculate the two sets of project and user vectors obtained in S4 to get the final user and project vector representation. The final user vector is The final item vector is S6, performing a dot product operation on the user and item vectors obtained in S5 to obtain the user's interaction probability with the item, and returning the recommendation result.
2. The recommendation method based on parallel graph neural network according to claim 1, characterized in that: In S1, for each user e u ∈R d (project, relationship i ,e r ∈R d ) assigns an initial embedding vector: The initialization vector is 64-dimensional.
3. The recommendation method based on parallel graph neural network according to claim 2, characterized in that: The S2 step specifically includes: S2.1, construct the user-item bipartite graph using the user-item interaction matrix; Establishment of user-project interaction matrix: Monitor user activities on the platform through user behavior logs, user feedback surveys, user purchase history, and user social media data methods, including clicks, browsing, purchases, ratings, comments, sharing, and preferences, and organize various interactive behaviors or feedback data between users and projects; mark positive interactive behaviors between user projects as 1, and non-positive interactive behaviors as 0, to form a user-project interaction matrix; According to the above user-item interaction matrix, a user-item bipartite graph is constructed; specifically, based on the user-item interaction matrix, user nodes and item nodes are created; each user and item is represented as a node in the graph, ensuring that each node has a unique identifier; next, edges are added to the bipartite graph according to the user-item interaction matrix; generally, if there is interaction between a user and an item, an edge is added between them; In this graph, users and items form two non-overlapping node sets respectively; each user node corresponds to a row of the matrix, and each item node corresponds to a column of the matrix; then, for users and items with an interaction record of 1 in the user-item interaction matrix, an edge with a weight of 1 is added between the two. S2.2, obtain the number of project nodes around a certain user through the bipartite graph Obtain the number of user nodes around a project through a bipartite graph S2.3, using the results of step 2.2, update the vector representation of users and items, specifically including: pass Update the vector representation of the user; pass Update the vector representation of the item; Multiple user vectors are passed through Get user vector Pass multiple items to the vector Get user vector 4. The recommendation method based on parallel graph neural network according to claim 3, characterized in that: The user initialized in step 1 is added to S3 and Projects vector, and the relationship initialized in step 1 The vector input attention and graph convolutional network module (Knowledge-aware layer of attention and graph convolutionalnetworks, KAGC) obtains the user and Projects vector.
5. The recommendation method based on parallel graph neural network according to claim 4, characterized in that: In S4, the above steps S2 and S3 are repeated L times to obtain the first group of users. and Projects Vector and the second group of users and Projects Vector, calculate the first set of vectors above to get Vector Sum Vector, the second set of vectors is calculated Vector Sum L is a natural number greater than 0.
6. The recommendation method based on parallel graph neural network according to claim 5, characterized in that: In S5, the two groups of user and item vectors obtained in S4 are combined to obtain the user and Projects vector.
7. The recommendation method based on parallel graph neural network according to claim 6, characterized in that: S3 specifically includes the following steps: S3.1, input the knowledge graph vector into the graph neural network to obtain the attention representation, which includes: The user Relationship with initialization Performing dot product operation yields The project Relationship with initialization Performing dot product operation yields The above and pass Weighted Will After normalization, we get pass Get the neighbor vector representation of the item; S3.2, the user vector is processed to obtain the self-attention representation, which specifically includes: By User Vector, use the following formula to calculate three vectors, W q , W k , W v is the feature transformation matrix: By Q u With K u The vector dot product of Will After normalization, we get pass Get the user's self-attention vector representation; S3.3, the item vector is updated through the attention representation in step 3.1, and the user vector is updated through the self-attention representation in step 3.2, specifically including: pass Get the vector representation of the item pass Get the vector representation of the user 8. The recommendation method based on parallel graph neural network according to claim 7, characterized in that: Step S4 specifically includes: Using step S2, multiple user vectors are obtained Use step S2 to get multiple item vectors Multiple user vectors are passed through Get user vector Pass multiple items to the vector Get user vector Using step S3, multiple user vectors are obtained Using step S3, multiple item vectors are obtained Multiple user vectors are passed through Get user vector Pass multiple items to the vector Get user vector 9. The recommendation method based on parallel graph neural network according to claim 8, characterized in that: In step S5; In step S6; Represents the interaction probability between a user and an item.