A Social Recommendation System Based on Lightweight Graph Convolutional Network
Through the information acquisition, dissemination and recommendation module of the lightweight graph convolution network, the data sparseness of users and items in the social recommendation system is solved, the efficiency and accuracy of the recommendation system are improved, and it is suitable for a variety of data environments.
Patent Information
- Application Number
- CN202111060105.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-10
- Publication Date
- 2025-07-22
- Estimated Expiration
- 2041-09-10
AI Technical Summary
The existing social recommendation system based on graph neural networks has data sparse problems when learning user and item representations, and nonlinear activation functions and structural transformations increase the difficulty of model training, making it difficult to effectively integrate information in user social graphs and user-item graphs.
A lightweight graph convolution network is used to obtain, disseminate and recommend information, calculate the representation set of users and items iteratively, and integrate user representations on different graphs using a fusion model, combining weighting and operation to improve matching accuracy.
It improves the efficiency and accuracy of the recommendation system, reduces time and space consumption, enhances the effect of user and item representation, and is suitable for a variety of data environments.
Smart Images
Figure CN113961820B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data mining, and specifically to a social recommendation system based on a lightweight graph convolutional network. Background Art
[0002] With the continuous expansion of the scale of e-commerce, the number and variety of commodities have grown rapidly, and customers need to spend a lot of time to find the commodities they want to buy. This process of browsing a large amount of irrelevant information and products will undoubtedly cause continuous loss of consumers submerged in the problem of information overload. To solve these problems, a personalized recommendation system has emerged. The personalized recommendation system is an advanced business intelligence platform based on massive data mining to help e-commerce websites provide completely personalized decision-making support and information services for their customers' shopping. Among them, the personalized recommendation system based on social information utilizes the known user social information to achieve personalized recommendations for users on the website.
[0003] Different from traditional recommendation systems, traditional recommendation systems have the problem of data sparsity: only mining the interaction information between users and items, so the representation ability is also limited by a small amount of user interaction data. The social-based recommendation system strengthens the representation of users and items by mining social network data, thus solving the problem of data sparsity. Social relationships have been proven in practice to be helpful for improving recommendation accuracy.
[0004] At present, a method model based on graph neural network (GNN) can effectively smooth node representations by aggregating different-order representations of adjacent nodes, and this method is currently the most advanced. However, applying the GNN model to a social recommendation system has two major challenges:
[0005] How to accurately learn the representations of users and items from the user social graph and the user-item graph;
[0006] How to integrate the user information extracted from the social graph and the user-item graph.
[0007] In addition, there is still room for improvement in applying the GNN-based model to a social recommendation system, and not all designs play a positive role in the recommendation process: since only the ID information is utilized in learning the representations of users and items, the non-linear activation function and structure transformation increase the difficulty of model training on the user-item interaction graph.
[0008] Therefore, the performance of existing machine learning models in applying to social recommendation systems needs to be improved urgently, and an improved learning model needs to be proposed. Summary of the Invention
[0009] The object of the present invention is to provide a social recommendation system based on a lightweight graph convolutional network, including an information acquisition module, an information propagation module, and an information recommendation module.
[0010] The information acquisition module randomly generates user vector information and item vector information and transmits them to the information propagation module
[0011] The information acquisition module randomly generates user vector information using edge structure information or standard normal distribution and item vector information n and m are the number of users and items respectively, and d is the dimension size of the vector representation.
[0012] The information propagation module performs iterative calculations on the user vector information and item vector information to obtain item representation sets at different levels and user representation sets at different levels k = 1,..., K.
[0013] The information propagation module transmits the item representation sets and user representation sets to the information recommendation module.
[0014] The information propagation module stores a user-item interaction graph, a user social graph, and a lightweight graph convolutional network.
[0015] The steps of obtaining item representation sets at different levels include: inputting the item vector information into the lightweight graph convolutional network, and making the lightweight graph convolutional network iterate the item vector information K times on the user-item interaction graph to obtain item representation sets at different levels
[0016] where the item representation at the k + 1-th layer is as follows:
[0017]
[0018] In the formula, respectively represent the adjacent sets of item i and user u in the user-item interaction graph.
[0019] The steps of obtaining user representation sets at different levels include:
[0020] 1) Inputting the user vector information Input into the lightweight graph convolutional network, and let the lightweight graph convolutional network iterate the user vector information K times on the user social graph to obtain a set of user social representations at different levels
[0021] The user vector information Input into the lightweight graph convolutional network, and let the lightweight graph convolutional network iterate the user vector information K times on the user-item interaction graph to obtain a set of user interest representations at different levels
[0022] Among them, the (k + 1)-th layer user social representation is as follows:
[0023]
[0024] In the formula, represents the set of adjacent users of user u on the social graph. represents the set of adjacent users of user v on the social graph.
[0025] The (k + 1)-th layer user interest representation is as follows:
[0026]
[0027] 2) Fuse the set of user social representations and the set of user interest representations at the same level to obtain a set of user representations
[0028] The fusion result of the (k + 1)-th layer user social representation and user interest representation is as follows:
[0029]
[0030] Among them, the process parameter is as follows:
[0031]
[0032] In the formula, σ(·) is the tanh activation function. The weight matrix W1, the weight matrix
[0033] The information recommendation module determines the matching level between the user and the item according to the set of item representations and the set of user representations If the matching level is greater than the threshold ε, the item is recommended to the user.
[0034] The information recommendation module according to the set of item representations and the user representation set The steps to determine the matching level between a user and an item include:
[0035] a) Calculate the user vector representation E u and the item vector representation E i , that is:
[0036]
[0037]
[0038] where α k is the weight.
[0039] b) Calculate the matching level between user u and item i that is:
[0040]
[0041] where e i and e u respectively represent the row vectors of E i and E u .
[0042] The social recommendation system based on the lightweight graph convolutional network further includes a database for storing information acquisition module, information dissemination module and information recommendation module data.
[0043] The technical effect of the present invention is beyond doubt. The present invention provides a social recommendation system based on a lightweight graph convolutional network. Based on the social relationship data between users and the interaction data between users and items, the vector representations of users and items are iteratively aggregated from neighbor nodes. A lightweight graph convolutional network is used to propagate the representations of users and items on the user-item interaction graph, while the representation of the user is propagated on the social graph. A fusion model is used to fuse the representations of the user on the two graphs, and the representations of users and items obtained from each layer are combined using a weighted sum. In the present invention, the representation of the user is propagated on the two graphs respectively, and the representation of the item is propagated on the user-item interaction graph, making use of the user's social data to enhance the representations of users and items. The fusion model ensures that the two representations of the user can be fully utilized, ensuring that the information learned by the user in the two graphs will not lead to a decline in the representation effect of users and items due to conflicts with each other, thus ensuring an improved recommendation effect. The present invention is based on the social data of users and the interaction data with items, has no special requirements for the data, and has strong universality. Compared with the existing social recommendation models based on GNN, due to the use of a lightweight graph convolutional network in the user-item interaction graph, two inefficient model processing operations: linear transformation and non-linear activation are removed, making the model under this method time and space efficient and having a better recommendation effect, with strong theoretical and practicality. Using lightweight graph convolution operations to obtain the potential information transmitted by users in the social network graph and the user-item graph, two inefficient model processing operations are removed, and it is more efficient in time and space than traditional GNN models.
[0044] The present invention takes into account the user information in the social network graph, and uses the designed fusion model to integrate the user vector information extracted from the user-item graph and the social graph, and uses weighted sum aggregation to combine the outputs of each layer to obtain a more accurate user-item matching value. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 is the algorithm flow chart of the present invention;
[0046] Figure 2 is the experimental comparison graph of the influence of the fusion function on the recommendation effect; Figure 2 (a) is the precision comparison of socialLGN-GCN, socialLGN-GraphSage, and socialLGN for processing the data of lastFM; Figure 2 (b) is the recall comparison of socialLGN-GCN, socialLGN-GraphSage, and socialLGN for processing the data of lastFM; Figure 2(c) Comparison of the evaluation metric NDCG obtained by processing the data of lastFM by socialLGN-GCN, socialLGN-GraphSage, and socialLGN; Figure 2 (d) Comparison of the precision obtained by processing the data of ciao by socialLGN-GCN, socialLGN-GraphSage, and socialLGN; Figure 2 (e) Comparison of the recall Recall obtained by processing the data of ciao by socialLGN-GCN, socialLGN-GraphSage, and socialLGN; Figure 2 (f) Comparison of the evaluation metric NDCG obtained by processing the data of ciao by socialLGN-GCN, socialLGN-GraphSage, and socialLGN. Detailed implementation manner
[0047] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject matter scope of the present invention is limited to the following embodiments. Without departing from the above-mentioned technical idea of the present invention, various substitutions and changes made according to the common general knowledge and customary means in the art shall be included within the protection scope of the present invention.
[0048] Embodiment 1:
[0049] See Figure 1 , a social recommendation system based on a lightweight graph convolutional network, including an information acquisition module, an information propagation module, and an information recommendation module.
[0050] The information acquisition module randomly generates user vector information and item vector information and transmits them to the information propagation module.
[0051] The information acquisition module randomly generates user vector information using edge structure information or standard normal distribution and item vector information n and m are the number of users and items respectively, and d is the dimension size of the vector representation.
[0052] The information propagation module performs iterative calculations on the user vector information and item vector information to obtain item representation sets at different levels and user representation sets at different levels k = 1,..., K.
[0053] The information propagation module transmits the item representation sets And the user representation set Transmit to the information recommendation module.
[0054] The information dissemination module stores a user-item interaction graph, a user social graph, and a lightweight graph convolutional network. The user-item interaction graph is a graph that describes the relationships and information transfer between users and items. The user social graph is a graph that describes the relationships and information transfer between different users.
[0055] The steps to obtain the item representation sets at different levels include: inputting the item vector information into the lightweight graph convolutional network, and making the lightweight graph convolutional network iterate the item vector information K times on the user-item interaction graph to obtain the item representation sets at different levels
[0056] Among them, the item representation at the k+1-th layer is as follows:
[0057]
[0058] In the formula, respectively represent the adjacent sets of item i and user u in the user-item interaction graph.
[0059] The steps to obtain the user representation sets at different levels include:
[0060] 1) Input the user vector information into the lightweight graph convolutional network, and make the lightweight graph convolutional network iterate the user vector information K times on the user social graph to obtain the user social representation sets at different levels
[0061] Input the user vector information into the lightweight graph convolutional network, and make the lightweight graph convolutional network iterate the user vector information K times on the user-item interaction graph to obtain the user interest representation sets at different levels
[0062] Among them, the user social representation at the k+1-th layer is as follows:
[0063]
[0064] In the formula, represents the adjacent user set of user u on the social graph. represents the adjacent user set of user v on the social graph.
[0065] The user interest representation at the k+1-th layer is as follows:
[0066]
[0067] 2) Fuse the set of user social representations at the same level The set of user interest representations to obtain the set of user representations at different levels
[0068] The fusion result of the user social representation and the user interest representation at the k+1-th level is as follows:
[0069]
[0070] Among them, the process parameter is as follows:
[0071]
[0072] In the formula, σ(·) is the tanh activation function. The weight matrix W1, The weight matrix
[0073] The information recommendation module determines the matching level between the user and the item according to the set of item representations and the set of user representations If the matching level is greater than the threshold ε, the item is recommended to the user.
[0074] The steps for the information recommendation module to determine the matching level between the user and the item according to the set of item representations and the set of user representations include:
[0075] a) Calculate the user vector representation E u and the item vector representation E i , that is:
[0076]
[0077]
[0078] In the formula, α k is the weight.
[0079] b) Calculate the matching level between the user u and the item i That is:
[0080]
[0081] In the formula, e i , e u respectively represent E i , E uRow vectors.
[0082] The social recommendation system based on the lightweight graph convolutional network further includes a database for storing information acquisition module, information dissemination module, and information recommendation module data.
[0083] Embodiment 2:
[0084] A method for using a social recommendation system based on a lightweight graph convolutional network is as follows:
[0085] 1) In the vector representation layer, use edge structure information or standard normal distribution to randomly generate a fixed-length vector representation for all users and items and where n and m are the number of users and items respectively, and d is the dimension size of the vector representation.
[0086] 2) In the vector information propagation layer, perform iterative calculations on the vector information of users and items obtained in step 1) and For the representation of an item Iterate K times through the lightweight graph convolutional network on the user-item interaction graph to obtain a set of item representations at different levels k = 1,..., K; for the representation of a user Use the lightweight graph convolutional network to iterate K times on the social graph and the user-item interaction graph respectively. On the social graph, obtain a set of user social representations at different levels k = 1,..., K, on the user-item interaction graph, obtain a set of user interest representations at different levels k = 1,..., K, and then use a fusion model to fuse the two representations of the user at the same level to obtain the final different-level representations of the user k = 1,..., K.
[0087] Utilize the social graph and the user-item interaction graph to propagate the representations of the nodes in the graph, making the representations of their adjacent nodes smooth, and obtaining the final vector representations of all nodes. Specifically, it is divided into two steps, a and b.
[0088] a) For the representation of an item, its iterative formula on the user-item graph is:
[0089]
[0090] where and respectively represent the adjacent sets of item i and user u in the user-item interaction graph.
[0091] For the representation of a user, its iterative formula on the user-item graph is:
[0092]
[0093] The iterative formula on the social graph is:
[0094]
[0095] where represents the set of adjacent users of user u on the social graph, represents the adjacent set of user u on the user-item interaction graph. The two representations obtained by the user spreading features in the two graphs are respectively and
[0096] b) There are the following three aggregation methods to fuse the iterative results of the user at the k-th layer:
[0097] i) GCN Aggregator
[0098]
[0099] ii) GraphSage Aggregator
[0100]
[0101] iii) Graph Aggregator
[0102]
[0103]
[0104] In this embodiment, the new fusion model Graph Aggregator is used to fuse them into the iterative result of the user representation at the k-th layer. Where σ(·) is the tanh activation function, W1, and the three matrices are all trainable.
[0105] 3) In the prediction layer, for the representations of users and items at different levels obtained in step 2), the final representations of each user and item are obtained through weighted summation. For the final representations of the obtained users and items, the matching scores of users to items are obtained by taking the dot product of the corresponding user and item representations.
[0106] The weighted calculation expression of the final vector representations of users and items is:
[0107]
[0108]
[0109] The calculation expression for the final user-item matching score is as follows:
[0110]
[0111] Example 3:
[0112] Refer to Figure 2 , the verification experiment of the social recommendation system based on the lightweight graph convolutional network is as follows:
[0113] 1) Obtain the dataset as shown in Table 1.
[0114] Table 1 Dataset used in the experiment
[0115]
[0116] 2) Use BPR, SBPR, DiffNet, NGCF, LightGCN, and SocialLGN (the system described in Example 1, Light Graph Convolution Network for social recommendation) to process the data in Table 1 respectively, and obtain the experimental comparison values as shown in Table 2.
[0117] Table 2 Comparison of experimental values of each model in the original dataset
[0118]
[0119] Since social recommendation algorithms mainly focus on dealing with the data sparsity problem and improving the recommendation performance of users who have only a small number of connections with items, an experiment defines a user who interacts with less than 20 items as a cold-start user, and constructs a test set containing only cold-start users from the original dataset. Table 3 shows the recommendation effects of each model in this test set:
[0120] Table 3 Comparison of experimental values of each model in the cold-start test set
[0121]
[0122] Table 4 Comparison of experimental values of fusing node values of each layer and only the node values of the last layer
[0123]
Claims
1. A social recommendation system based on a lightweight graph convolutional network, characterized in that: It includes an information acquisition module, an information dissemination module, and an information recommendation module; The information acquisition module randomly generates user vector information and item vector information and transmits it to the information dissemination module; The information dissemination module performs iterative calculations on the user vector information and the item vector information to obtain item representation sets at different levels and user representation sets at different levels k = 1, …, K; The information dissemination module transmits the set of item representations and the set of user representations to the information recommendation module; The information recommendation module determines a matching level between a user and an item according to the item representation set and the user representation set If the matching level is greater than the threshold ε, the item is recommended to the user; The information dissemination module stores a user-item interaction graph, a user social graph, and a lightweight graph convolutional network; Steps for obtaining a set of item representations at different levels include: inputting item vector information into a lightweight graph convolutional network, and making the lightweight graph convolutional network iterate the item vector information K times on the user-item interaction graph to obtain a set of item representations at different levels Among them, the item at the (k + 1)-th layer represents as follows: In the formula, respectively represent the adjacent sets of item i and user u in the user-item interaction graph; Obtain a set of user representations at different levels The steps include: 1) Input the user vector information into the lightweight graph convolutional network, and let the lightweight graph convolutional network iterate the user vector information K times on the user social graph to obtain a set of user social representations at different levels Input the user vector information into the lightweight graph convolutional network, and let the lightweight graph convolutional network iterate the user vector information K times on the user-item interaction graph to obtain a set of user interest representations at different levels Among them, the user social representation of the (k + 1)-th layer is as follows: wherein, represents the set of adjacent users of user u in the social graph; represents the set of adjacent users of user v in the social graph; User interest representation of the (k + 1)-th layer As shown below: 2) Combine the user social representation sets at the same level with the user interest representation sets to obtain user representation sets at different levels The fusion result of the user social representation and user interest representation at the (k + 1)-th layer is as follows: Among them, the process parameters are as follows: where, σ(·) is the tanh activation function; the weight matrix W1, weight matrix 2. The social recommendation system based on a lightweight graph convolutional network according to claim 1, wherein: The information acquisition module uses side structure information or standard normal distribution to randomly generate user vector information and item vector information n and m are the number of users and items respectively, and d is the dimension size of the vector representation.
3. The social recommendation system based on a lightweight graph convolutional network according to claim 1, wherein The information recommendation module determines the matching level between a user and an item according to the set of item representations and the set of user representations The steps of determining the matching level between a user and an item include: 1) Calculate the user vector representation E u and the item vector representation E i , namely: where α k is the weight; 2) Calculate the matching level between user u and item i That is: where e i and e u respectively represent the row vectors of E i and E u .
4. A social recommendation system based on a lightweight graph convolutional network according to claim 1, characterized in that, It also includes a database that stores the data of the information acquisition module, the information dissemination module, and the information recommendation module.