Contrastive Learning and Multi-Choice Negative Sampling Recommendation

Most existing recommendation models that directly model user interests on user-item interaction data usually ignore the natural noise present in the interaction data, leading to bias in the model's learning of user preferences during data propagation and aggregation. In addition, the currently adopted negative sampling strategy does not consider the relationship between the prediction scores of positive samples and the degree of difficulty of negative samples, and is unable to adaptively select a suitable negative sample for each positive sample, leading to a decrease in the model recommendation performance. In order to solve the above problems, this paper proposes a Contrastive Learning and Multi-choice Negative Sampling Recommendation. Firstly, an improved topology-aware pruning strategy is used to process the user-item bipartite graph, which uses the topology information of the graph to remove noise and improve the accuracy of model prediction. In addition, a new multivariate selective negative sampling module is designed, which ensures that each positive sample selects a negative sample of appropriate hardness through two sampling principles, improving the model embedding space representation capability, which in turn leads to improved model recommendation accuracy. Experimental results on the Urban-Book and Yelp2018 datasets show that the proposed algorithm significantly improves all the metrics compared to the state-of-the-art model, which proves the effectiveness and sophistication of the algorithm in different scenarios.

@article{Xue2024,
  title     = {Contrastive Learning and Multi-Choice Negative Sampling Recommendation},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {5},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Yun Xue and Xiaodong Cai and Sheng Fang and Li Zhou},
  doi       = {10.14569/IJACSA.2024.0150592},
  url       = {https://doi.org/10.14569/IJACSA.2024.0150592}
}