MC-RGN: Residual Graph Neural Networks based on Markov Chain for sequential recommendation

Sequential recommendation aims to predict the next item the user will interact with based on his/her historical interaction sequences. Existing sequential recommendation methods mainly use self-attention to capture the preferences with individual user-item historical interaction sequences, while ignoring the group preferences that are exhibited by all user-item interaction sequences. Besides, like general recommendation, data sparsity remains the main factor detrimentally affecting recommendation performance. To remedy the damage of data sparsity and improve the recommendation performance, we employ Residual connections and Graph Neural Networks based on Markov Chain to propose MC-RGN, which incorporates group preferences into a single user-item historical sequence. To demonstrate the effectiveness of the proposed method, extensive experiments were conducted on four real-world datasets, MovieLens, Sports, Yelp and Beauty. Experimental results showed that the proposed method achieved an average improvement of 14.32% on Hit@10 and 27.17% on NDCG@10 compared to the baseline models. To showcase whether the proposed method alleviates the negative impact of data sparsity on recommendation performance, we compared the recommendation performance decline of the SASRec and the proposed method under various levels of data sparsity. The results indicate that the proposed method notably reduced the damage caused by data sparsity (0.07% vs 20.94%, 2.37% vs 23.70%, 6.51% vs 27.01%, and 22.78% vs 28.42% at data usage ratios of 80%, 60%, 40%, and 20%, respectively). Finally, our proposed model is available via https://github.com/crx1729/MC-RGN.