Two-Stage Enhancement for Recommendation Systems Based on Contrastive Learning

Knowledge-aware recommendation systems use knowledge graphs as side information to enhance the performance of recommendation systems. The existing methods often employ graph neural networks to process the relational networks and use contrastive learning to obtain more effective node representations. However, persistent challenges from active users’ noisy data and the cold-start problem related to inactive users impact model performance. Recent studies mainly use data augmentation methods to address these issues but only consider the enhancement during the training stage and neglect the pre-training stage. For this reason, we propose a Two-stage Enhancement for Recommendation Systems based on Contrastive Learning (TERSCL). This model primarily integrates two stages of data augmentation methods: one is the pre-training enhancement, where an enhanced adjacency matrix is obtained for subsequent training. The other generates a distinct dropout subgraph during each training round to participate in the subsequent graph encoding and contrastive learning process. Extensive experiments conducted on three benchmark datasets verify that TERSCL can significantly improve recommendation performance and mitigate the cold-start problem.