Asymmetric neighborhood aggregated residual graph convolution collaborative filtering recommendation method and device

Abstract

The application relates to a kind of asymmetric neighborhood aggregation residual graph convolution collaborative filtering recommendation method and equipment.The method constructs the two-part graph of user and item according to user feature information, item feature information and its interaction data, the feature information of user and item is one-hot encoding, then using linear embedding layer is mapped to initial node embedding vector;Node embedding vector is carried out feature propagation in improved graph convolution layer, the weighted average of node embedding vector of all layers is obtained, and the final embedding vector of user and item is obtained;Inner product operation is carried out on the final embedding vector in the prediction layer, and the score of target user to target item is obtained.The application uses a new feature propagation mechanism to update the node embedding vector of user and item, fully captures the high-order interaction relationship between user and item, not only relieves the over-smoothing effect, but also fully utilizes the low-order node features of user and item, thereby effectively improving the accuracy of recommendation.

Description

Technical Field

[0001] This invention relates to the field of personalized recommendation technology, and in particular to a method and apparatus for residual graph convolutional collaborative filtering recommendation based on asymmetric neighborhood aggregation. Background Technology

[0002] In the field of personalized recommendation, collaborative filtering models are widely used, but their performance is limited by the sparsity of the data. Graph-based collaborative filtering models, however, demonstrate superior recommendation performance by modeling user-item interaction data as a bipartite graph and then using graph neural networks to learn user interests. Graph convolutional networks can effectively learn the structural properties and node feature information of graphs, making them well-suited for graph-based collaborative filtering models and mitigating the data sparsity problem to some extent.

[0003] Despite the excellent performance of collaborative filtering models based on graph convolutional networks (GCNNs) in recommendations, the following issues remain: 1. The propagation effect of low-order node features of users or items is weakened in GCNNs, resulting in insufficient ability of existing models to extract interaction information. 2. In real-world scenarios, GCNNs exhibit over-smoothing effects at deeper layers, making it impossible for existing models to model deep interactions between users and items. 3. Existing models generally employ symmetric normalization in neighborhood aggregation of graph convolutions, thereby reducing recommendation performance.

[0004] For example, NGCF and LightGCN are two representative graph collaborative filtering methods. They update the node embedding vectors of each layer by stacking multiple graph convolutional layers and finally merge the embedding vectors of all layers for recommendation. However, their feature propagation mechanisms do not fully utilize the feature information of low-level nodes, resulting in insufficient ability of the model to extract interaction information. Furthermore, since graph convolutional networks are essentially low-pass filters, deeper nodes become increasingly difficult to distinguish as the number of layers increases, causing the model to lose its ability to model deep interactions. Finally, extensive experiments on LightGCN for asymmetric neighborhood aggregation revealed that asymmetric normalization in neighborhood aggregation degrades model performance. Summary of the Invention

[0005] To address the problems of the aforementioned graph convolutional network collaborative filtering model, this invention proposes a residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation. The aim is to enable this method to more fully explore the higher-order connectivity relationships between users and items, and to model deeper interactions, thereby learning high-quality node representations and ultimately improving recommendation performance.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A residual graph convolutional collaborative filtering recommendation method with asymmetric neighborhood aggregation includes the following steps:

[0008] S1. Construct a bipartite graph of user-item relationships based on user feature information, item feature information, and their interaction data, including an adjacency matrix and an asymmetric normalized graph Laplace matrix.

[0009] S2. Perform one-hot encoding on the feature information of users and items, then use a linear embedding layer to map it to the initial node embedding vector, and then construct the node embedding matrix.

[0010] S3. Based on the graph Laplacian matrix obtained in step S1, construct an improved graph convolutional layer for feature propagation to obtain a more refined node embedding vector.

[0011] S4. Perform a weighted average of the node embedding vectors of all layers to obtain the final embedding vectors of users and items;

[0012] S5. In the prediction layer, the inner product operation is performed on the final embedded vector to obtain the target user's rating of the target item. The loss function is used to continuously optimize the model parameters, thereby optimizing the user's rating matrix of the item and predicting the user's preference for the item.

[0013] The specific process of step S1 is as follows:

[0014] S1.1 User feature information is represented as a user node set U = {u1, u2, u3, ..., u...} m The item feature information is represented as an item node set I = {i1, i2, i3, ..., i}. n};Let R ui This is a record of interactions between user u and item i. If an interaction occurs, then R... ui If it is 1, otherwise R ui The value is 0, thus obtaining the user-item interaction matrix. And calculate the adjacency matrix A according to equation (1):

[0015]

[0016] S1.2. Constructing a graph Laplacian using asymmetric normalization in the neighborhood aggregation of graph convolution.

[0017] matrix It is calculated according to formula (2):

[0018]

[0019] Where D is the degree matrix, D ii =∑ j A ijp is the normalization coefficient, used to control the normalization strength in neighborhood aggregation.

[0020] The specific process of step S2 is as follows:

[0021] S2.1 The embedding layer performs one-hot encoding on the initial user feature information and item feature information, mapping it to a low-dimensional dense embedding vector; the user feature information is mapped to... Item feature information mapped to in d is the dimension of the embedding vector;

[0022] S2.2 Construct a matrix using the vector representation described above:

[0023]

[0024] in This represents the initial node embedding matrix for users and items, where m and n represent the number of users and items, respectively.

[0025] Step S3 uses the propagation of user u in the l-th graph convolutional layer as an example to illustrate the propagation process of the node embedding vector in the graph convolutional layer:

[0026] During initial propagation, in the user's low-order embedding vector and higher-order embedding vectors A residual connection is established between them, and its formal representation is shown in equation (4):

[0027]

[0028] in Denotes the asymmetric normalized graph Laplace norm. N represents the embedding vector of item i in layer l-1. u and N i Let represent the number of neighbor nodes for user u and item i, respectively;

[0029] The high-order embedding vector generated after the initial propagation The vector of user u is further refined in layer l by propagating it again in the convolutional layer l. This alleviates the oversmoothing effect in deep graph convolutional networks, and its formal representation is shown in Equation (5):

[0030]

[0031] In equation (5), α is a hyperparameter used to eliminate information redundancy and prevent overfitting;

[0032] Similarly, the more refined embedding vector of item i in layer l Its formal representation is shown in equation (6):

[0033]

[0034] In formula (6) This represents the embedding vector of user u in layer l-1;

[0035] The specific process of step S4 is as follows: Through feature propagation in step S3, refined node embedding vectors of user u and item i at different layers are obtained. Then, these node embedding vectors are weighted and averaged to obtain the final embedding vector, as shown below:

[0036]

[0037] In equation (9), This represents the final embedding vector of the target user u. This represents the final embedding vector of the target item i.

[0038] The specific process of step S5 is as follows:

[0039] S5.1, For the final embedding vectors of target user u and target item i generated in step S4

[0040] and The rating of target user u for target item i is calculated using equation (10), thus obtaining the rating matrix of all users for the item.

[0041]

[0042] S5.2. Establish the scoring loss function L according to equation (11). bpr :

[0043]

[0044] in These are all the training data, i∈N u This refers to items that have interacted with user u. Let represent negative samples obtained from the set of items that have no interaction with user u; λ is the regularization coefficient. This indicates that l2 regularization is applied to the initial embedding, where δ(·) is the Sigmoid function;

[0045] S5.3. Apply the mini-batch ADMA optimizer to the above loss function L. bpr Optimize to make L bpr The algorithm converges to its minimum value, thus yielding the optimal rating matrix for items by users. And based on the optimal rating matrix Item recommendations can be made to predict user preferences for items.

[0046] To facilitate batch processing, user u and item i utilize the graph Laplacian matrix generated in step S1. After converting to matrix form, feature propagation is performed, and the formal representation is as follows:

[0047]

[0048]

[0049] Where E (l) E represents the higher-order embedding matrix generated during the initial propagation at layer l. (l) It is E (l) The refined embedding matrix generated after propagation again at layer l.

[0050] An electronic device includes at least one processor and a memory communicatively connected to the processor, the memory storing a program executable by the at least one processor, which, when executed by the processor, enables the implementation of the above-described method.

[0051] A computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the above-described method.

[0052] The beneficial effects of this invention are:

[0053] 1. This invention utilizes a novel feature propagation mechanism to update the node embedding vectors of users and items, fully capturing the high-order interaction relationships between users and items. This not only alleviates the over-smoothing effect but also makes full use of the low-order node features of users and items, thereby effectively improving the accuracy of recommendations.

[0054] 2. In order to achieve batch processing, this invention generates embedding vector matrices for users and items based on the graph Laplacian matrix during feature propagation. This allows for graph learning even with few features, has good scalability, and can be fused with more diverse data sources, such as user attributes and item attributes, thereby effectively alleviating the data sparsity problem in collaborative filtering.

[0055] 3. This invention employs asymmetric normalization to construct a graph Laplacian matrix for neighborhood aggregation in graph convolution, further improving recommendation performance. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the method of the present invention;

[0057] Figure 2 This is a schematic diagram of a model structure provided by the present invention;

[0058] Figure 3 These are experimental results of the LightGCN-p variant model on the Movielens-1M dataset;

[0059] Figure 4 These are experimental results of the LightGCN-p variant model on the Gowalla dataset;

[0060] Figure 5 These are experimental results of the LightGCN-p variant model on the Yelp2018 dataset. Detailed Implementation

[0061] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0062] A residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation, such as... Figure 1 As shown, it includes the following steps:

[0063] S1. Construct a bipartite graph of user-item relationships based on user characteristic information (ID), item characteristic information (ID), and their interaction data. This graph includes an adjacency matrix and an asymmetric normalized graph Laplacian matrix. The specific process is as follows:

[0064] First, obtain user feature information and item feature information. User feature information is represented as a user node set U = {u1, u2, u3, ..., u...} m}, where m represents the number of users, and the item feature information is represented as the item node set I = {i1, i2, i3, ..., i}. n Let R be the number of items, n, and the adjacency matrix be A. ui This is a record of interactions between user u and item i. If an interaction occurs, then R... ui If it is 1, otherwise R ui The value is 0, thus obtaining the user-item interaction matrix. And calculate the adjacency matrix A according to equation (1):

[0065]

[0066] Then, in the neighborhood aggregation of graph convolution, asymmetric normalization is used to construct the graph Laplacian matrix. It is calculated according to formula (2):

[0067]

[0068] Where D is the degree matrix, D ii =∑ j A ij p is the normalization coefficient, used to control the normalization strength in neighborhood aggregation.

[0069] S2. Perform one-hot encoding on the feature information (ID) of users and items, then map it to the initial node embedding vector using a linear embedding layer, and finally construct the node embedding matrix. The process is as follows:

[0070] First, after one-hot encoding of the initial user and item feature information in the embedding layer, it is mapped into a low-dimensional dense embedding vector; the user feature information is mapped to... Item feature information mapped to in d is the dimension of the embedding vector;

[0071] Then, construct a matrix using the vector representation described above:

[0072]

[0073] in This represents the initial node embedding matrix for users and items, where m and n represent the number of users and items, respectively.

[0074] S3. Based on the graph Laplacian matrix obtained in step S1, construct an improved graph convolutional layer for feature propagation to obtain a more refined node embedding vector.

[0075] This invention utilizes an improved graph convolutional propagation mechanism in the residual propagation layer to perform feature propagation, extracting more user-item interaction information, thereby achieving more accurate user interest modeling and enhancing recommendation performance.

[0076] For the node embedding vector obtained in step S2 and We learn complex high-order connectivity relationships between user nodes and item nodes through improved graph convolutional layers, and continuously update node embedding vectors. and This yields more refined node embedding vectors for each layer.

[0077] In this invention, the propagation of the node embedding vector in the graph convolutional layer is completed in two steps. Taking the propagation of user u in the l-th graph convolutional layer as an example, the propagation process of the node embedding vector in the graph convolutional layer is illustrated as follows:

[0078] During initial propagation, in the user's low-order embedding vector and higher-order embedding vectors Residual connections are established between them to fully utilize the low-order embedding vectors of user nodes, and their formal representation is shown in Equation (4):

[0079]

[0080] in Denotes the asymmetric normalized graph Laplace norm. N represents the embedding vector of item i in layer l-1. u and N i Let represent the number of neighbor nodes for user u and item i, respectively.

[0081] The high-order embedding vector generated after the initial propagation The vector of user u is further refined in layer l by propagating it again in the convolutional layer l. This alleviates the oversmoothing effect in deep graph convolutional networks, and its formal representation is shown in Equation (5):

[0082]

[0083] In equation (5), α is a hyperparameter used to eliminate information redundancy and prevent overfitting; generally, the value of α is 0.1 or 0.2.

[0084] Similarly, the more refined embedding vector of item i in layer l Its formal representation is shown in equation (6):

[0085]

[0086] In formula (6) This represents the embedding vector of user u in layer l-1;

[0087] To facilitate batch processing, user u and item i utilize the graph Laplacian matrix generated in step S1. After converting to matrix form, feature propagation is performed, and the formal representation is as follows:

[0088]

[0089]

[0090] Where E (l′) E represents the higher-order embedding matrix generated during the initial propagation at layer l. (l) It is E (l′) The refined embedding matrix generated after propagation again at layer l.

[0091] S4. Take a weighted average of the node embedding vectors from all layers to obtain the final embedding vectors for users and items. The specific process is as follows: Through feature propagation in step S3, refined node embedding vectors for user u and item i at different layers are obtained. Then, these node embedding vectors are weighted and averaged to obtain the final embedding vector, as shown below:

[0092]

[0093] In equation (9), This represents the final embedding vector of the target user u. This represents the final embedding vector of the target item i.

[0094] S5. In the prediction layer, perform an inner product operation on the final embedded vector to obtain the target user's rating of the target item. Then, continuously optimize the model parameters using the loss function, thereby optimizing the user's rating matrix for the item and predicting the user's preference for the item. The specific process is as follows:

[0095] S5.1, For the final embedding vectors of target user u and target item i generated in S4 and The rating of target user u for target item i is calculated using equation (10), thus obtaining the rating matrix of all users for the item.

[0096]

[0097] S5.2. Establish the scoring loss function L according to equation (11). bpr :

[0098]

[0099] in These are all the training data, i∈N u This refers to items that have interacted with user u. Let represent negative samples obtained from the set of items that have no interaction with user u; λ is the regularization coefficient. This indicates that l2 regularization is applied to the initial embedding, and δ(·) is the Sigmoid function.

[0100] S5.3. Apply the mini-batch ADMA optimizer to the above loss function L. bpr Optimize to make L bpr The algorithm converges to its minimum value, thus yielding the optimal rating matrix for items by users. And based on the optimal rating matrix Item recommendations can be made to predict user preferences for items.

[0101] The advantages of the method of the present invention are illustrated below with comparative experiments:

[0102] First, extensive asymmetric neighborhood experiments were conducted on LightGCN. Specifically, the matrix form of the feature propagation mechanism of LightGCN and LightGCN-p variant models is shown below:

[0103] E (l) =D -12 AD -12 E (l-1) ,

[0104] E (l) =D -p AD -(1-p) E (l-1) ,

[0105] Experimental results of the LightGCN-p variant model on three datasets: Movielens-1M, Gowalla, and Yelp2018, as follows: Figure 3-5 As shown in the figure, the experimental results of the LightGCN-p variant model show that the use of symmetric normalization in neighborhood aggregation does indeed reduce the performance of existing models.

[0106] Then, to verify the effectiveness of the proposed method (ARGCCF), experiments were conducted on three real-world datasets: Movielens-1M (i.e., ML-1M), Gowalla, and Yelp 2018. These datasets differ in scene themes, sizes, and sparsity, but are all publicly available.

[0107] The statistical information for the three datasets is as follows:

[0108]

[0109] This invention employs two commonly used evaluation metrics in graph-based collaborative filtering: Recall and NDCG (Normalized Discount Cumulative Gain). Recall@k is a commonly used metric for measuring recall in recommendation tasks, representing the proportion of the user's k recommended positive samples out of the entire positive sample set. NDCG@k is also a commonly used metric for measuring the quality of the recommendation list's ranking; the higher the score, the better the ranking of positive samples.

[0110] This invention was compared with six other methods: MF, NeuMF, GC-MC, LR-GCCF, NGCF, and LightGCN.

[0111] The recommendation performance of the proposed method and the comparison method on three datasets is shown in the table below:

[0112]

[0113] Analysis of the experimental results shows that the ARGCCF method performs best on the three datasets, achieving the highest recall and ndcg scores. The scores obtained by the proposed method are achieved with 6 graph convolutional layers, demonstrating the effectiveness of the proposed embedding propagation mechanism. This mechanism not only fully utilizes the low-order node feature information of users and items but also alleviates the over-smoothing effect, further improving the model's recommendation performance and enhancing the user experience.

[0114] In summary, unlike the symmetric neighborhood aggregation method commonly used in graph convolutional networks, the method provided in this invention constructs an asymmetric graph Laplacian matrix in step S1 by changing the exponent size of the degree matrix, thereby further improving recommendation performance. Furthermore, in step S3, an improved feature propagation mechanism extracts more user-item interaction information, resulting not only in high-quality user and item representations but also mitigating the over-smoothing effect in graph convolution. Experiments demonstrate the effectiveness of the method provided in this invention and its improved recommendation performance.

Claims

1. A residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation, characterized in that... Includes the following steps: S1. Construct a bipartite graph of user-item relationships based on user feature information, item feature information, and their interaction data, including an adjacency matrix and an asymmetric normalized graph Laplace matrix. S2. Perform one-hot encoding on the feature information of users and items, then use a linear embedding layer to map it into an initial node embedding vector, and then construct a node embedding matrix. S3. Based on the graph Laplacian matrix obtained in step S1, construct an improved graph convolutional layer for feature propagation to obtain a more refined node embedding vector. Among them, users In the The propagation process of the convolutional layer is as follows: During initial propagation, in the user's low-order embedding vector and higher-order embedding vectors A residual connection is established between them, and its formal representation is shown in equation (4): （4）， in Denotes the asymmetric normalized graph Laplace norm. express Layered items Embedded vector, and Representing users respectively and items The number of neighboring nodes, p It is the normalization coefficient; The higher-order embedding vector generated after the initial propagation In the The graph convolutional layer propagates again to obtain the user's... exist More refined embedding vectors in the layer This alleviates the oversmoothing effect in deep graph convolutional networks, and its formal representation is shown in equation (5): （5）， In formula (5) These are hyperparameters used to eliminate information redundancy and prevent overfitting; Similarly, items exist More refined embedding vectors in the layer Its formal representation is shown in equation (6): （6）， In formula (6) Indicates user exist Layer embedding vector; S4. Perform a weighted average of the node embedding vectors of all layers to obtain the final embedding vectors of users and items; S5. In the prediction layer, the inner product operation is performed on the final embedded vector to obtain the target user's rating of the target item. The loss function is used to continuously optimize the model parameters, thereby optimizing the user's rating matrix of the item and predicting the user's preference for the item.

2. The residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation according to claim 1, characterized in that: The specific process of step S1 is as follows: S1.1 User feature information is represented as a set of user nodes. Item feature information is represented as a set of item nodes. ;make For users With items The interaction records between them; if there is any interaction behavior, then... =1, otherwise The value is 0, thus obtaining the user-item interaction matrix. And calculate the adjacency matrix according to equation (1). A : （1）； S1.

2. Asymmetric normalization is used to construct the graph Laplacian matrix in the neighborhood aggregation of graph convolution. It is calculated according to formula (2): （2）， in D It is a degree matrix. , p It is the normalization coefficient, used to control the normalization strength in neighborhood aggregation.

3. The residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation according to claim 1, characterized in that: The specific process of step S2 is as follows: S2.1 The embedding layer performs one-hot encoding on the initial user feature information and item feature information, mapping it to a low-dimensional dense embedding vector; the user feature information is mapped to... Item feature information is mapped to ;in , , d It is the dimension of the embedded vector; S2.2 Construct a matrix using the vector representation described above: （3）， in This represents the initial node embedding matrix for users and items. m and n These represent the number of users and the number of items, respectively.

4. The residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation according to claim 1, characterized in that: The specific process of step S4 is as follows: Through feature propagation in step S3, the user is obtained. and items The node embedding vectors are refined at different layers, and then these node embedding vectors are weighted and averaged to obtain the final embedding vector, as shown below: （9）， In equation (9), Indicates target user The final embedding vector, Indicates the target item The final embedding vector.

5. The residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation according to claim 1, characterized in that: The specific process of step S5 is as follows: S5.1 For the target user generated in step S4 and target items The final embedding vector and The target user is calculated using equation (10). For target items The rating values ​​are used to obtain a rating matrix of all users for the item. : （10）； S5.

2. Establish the scoring loss function according to equation (11). : （11）， in It is all the training data. Indicates to users Interactive items, Indicates to users Negative samples obtained from a set of items with no interaction; It is the regularization coefficient. Indicates implementation of the initial embedding Regularization, It is the Sigmoid function; S5.

3. Apply the mini-batch ADMA optimizer to the above loss function. Optimize to make The algorithm converges to its minimum value, thus yielding the optimal rating matrix for items by users. And based on the optimal scoring matrix Item recommendations can be made to predict user preferences for items.

6. The residual graph convolutional collaborative filtering recommendation method based on asymmetric neighborhood aggregation according to claim 1, characterized in that: To facilitate batch processing, users and items Using the graph Laplacian matrix generated in step S1 After converting to matrix form, feature propagation is performed, and the formal representation is as follows: （7）， （8）， in Indicates the first The high-order embedding matrix generated during the initial propagation of the layer, yes In the The refined embedding matrix generated after the layer propagates again.

7. An electronic device, characterized in that: It includes at least one processor and a memory communicatively connected to the processor, the memory storing a program executable by at least one processor, which, when executed by the processor, enables the implementation of the method as described in any one of claims 1-6.

Images