Mixed-curvature Knowledge-Enhanced Graph Contrastive Learning for Recommendation.
Yihao Zhang,Junlin Zhu,Ruizhen Chen,Weiwen Liao,Yulin Wang,Wei Zhou
DOI: https://doi.org/10.1016/j.eswa.2023.121569
IF: 8.5
2024-01-01
Expert Systems with Applications
Abstract:Contrastive learning has recently triggered a series of valuable studies in the recommendation field, as it can extract supervised information from large-scale unsupervised data, reducing interference from unrelated entities. However, in real scenarios, supervised data is often sparse and challenging to fully leverage, inevitably leading to decreased accuracy in recommendation algorithms based on contrastive learning. Although knowledge graphs can offer abundant facts and serve as a rich source for supervised signals, achieving data augmentation effects, knowledge graph embedding heightens sensitivity to changes in graph node information. Moreover, the Euclidean space is not entirely suitable for accommodating varying degrees of graph structure expansion. To fill this research gap, we propose a mixed-curvature knowledge-enhanced graph contrastive learning for recommendation (MKGCL). Specifically, we design a knowledge-enhanced approach to generate knowledge-enhanced global views, which combines user–item interaction views with knowledge-augmented semantic views, resulting in more reliable and comprehensive regulatory signals. In particular, based on the characteristics of the curvature space, we construct multiple product manifolds of a single curvature space, thereby constructing a comprehensive mixed-curvature space, which ensures better scalability for the expansion of the graph structure. Extensive experiments on four benchmark datasets, results show that MKGCL significantly outperforms other state-of-the-art algorithms in comparison.