Effective hybrid graph and hypergraph convolution network for collaborative filtering

Li, Xunkai; Guo, Ronghui; Chen, Jianwen; Hu, Youpeng; Qu, Mingli; Bin, Jiang

doi:10.1007/s00521-022-07735-y

Cited by 5 publications

(2 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…HGCN is combined with feature crossover networks in a parallel manner to achieve a balance between feature crossover and excessive smoothing. Li et al [12] proposed an effective hybrid graph and hypergraph convolutional network(EHGCN) recommendation model that uses graphs and hypergraphs to model the correlation between nodes in interaction graphs, achieving multi-level learning. EHGCN also adopts the DenseGCN training framework to optimize the graph convolution strategy from the perspective of graph signal processing…”

Section: A Graph and Hypergraph Convolution Networkmentioning

confidence: 99%

“…2) LightGCN [4]: This model constructs a bipartite graph of user and course interaction, uses a light GCN to learn the feature representation of nodes, and the embeddings of nodes in each layer graph convolved are added as the final embeddings of nodes 3) EHGCN [12]:This is an effective hybrid graph and hypergraph convolution network for collaborative filtering, which uses graph and hypergraph to model the correlation among nodes in the interaction graph for multilevel learning.Then the improved DenseGCN model framework is used to implement multi-layer graph convolution. 4) SGL [14]:This model is a basic graph contrastive learning model.…”

Section: ) Baselinesmentioning

confidence: 99%

See 1 more Smart Citation

Course Recommendation Model Based on Layer Dropout Graph Differential Contrastive Learning

Ouyang,

Long,

Gao

et al. 2024

IEEE Access

View full text Add to dashboard Cite

At present, the course recommendation model of graph collaborative filtering mainly uses bipartite graph modeling to obtain user-course cooperative relationship. However, the bipartite graph lacks the acquisition of user-user and course-course relationship information. In addition, due to the inherent defects of graph convolution, multi-layer graph convolution will cause overfitting problems. Moreover, the existing graph contrastive learning methods to solve the sparsity of recommendation data simply divide nodes into positive and negative pairs, without taking into account that users who have chosen the same course in the recommendation are similar. In contrastive learning, the feature similarity distance of these users should be different.To solve these problems, a course recommendation model based on layer dropout graph differential contrastive learning(DGDCL) is proposed. Specifically, a hybrid graph convolution network of fusion graph and hypergraph is used to obtain both low-order and high-order information. Then, using the layer dropout method to alleviate overfitting in neural network, the multi-layer feature embeddings of graph nodes are randomly dropout. Finally, two different layer drops are used to generate the contrastive views to reduce the additional noise and computational overhead of generating the contrastive views. The prior similarity of users and courses is used to adjust the calculation of the contrastive loss function, and differentiated contrastive learning of graph nodes is realized to make the contrastive learning more suitable for the recommendation model. The experimental results of XuetangX and MOOCCube datasets show that the proposed model is better than the existing model.

show abstract

Section: A Graph and Hypergraph Convolution Networkmentioning

confidence: 99%