Sampling-based Self-Supervised Augmented Contrast Learning for Movie Recommendation Systems

As the number of video users continues to grow, film and television have become one of the indispensable media in people’s lives. With this comes a huge explosion of information data, thus recommendation systems become one of the effective ways to solve information overload. We hope to research a better recommendation algorithm to further help provide more interesting movies to our users. In recent years, contrast learning has become a self-supervised learning paradigm that shows some effectiveness without labels. Graph neural networks have also been widely used in past recommender systems, but there are still problems of large overlap of interaction information, weak node representation, and long-tail distribution of user behavior. To solve the above problems, this paper proposes a movie recommendation system (SSACL) based on sampling-based self-supervised augmented contrast learning. Specifically, we learn node embeddings by constructing heterogeneous network views and meta-path views and project them into a multi-layer perceptron for contrast learning, while we employ graph augmentation methods such as resampling and edge discarding to assist in the recommendation task. Our aim is to leverage node-level information and the application of contrast learning to the recommendation task. Our empirical and ablation studies on two real datasets demonstrate that the SSACL model outperforms the current mainstream approaches.