Enhancing Temporal Dynamics and Hypergraph Convolutions for Sequential Recommendation

Sequential recommendation systems focus on fore-casting the subsequent item a user may be interested in, based on their previous interactions. While current sequential recom-mendation techniques have achieved notable success, they still face several limitations: (1) Most existing approaches emphasize the sequence of interactions but neglect the actual timestamps between them. (2) Graph neural network-based methods for sequential recommendations often treat item transitions as simple pairwise relationships, overlooking the intricate high-order inter-actions among items. To overcome these problems, we propose an innovative approach called THCSR, which explicitly models timestamp order and examines how different time intervals affect the prediction of the next item. Additionally, the line graph of a hypergraph is leveraged alongside a graph convolutional network to model the intricate higher-order connections between items. Specifically, consider the items within the user's interaction sequence as hyperedges, and hyperedges that share common items combine to form a hypergraph, capturing higher-order correlations between items. By converting the hypergraph into a line graph, we utilize a graph convolutional network to aggregate information and uncover complex item relationships, thereby extracting significant features. Additionally, an embedding mod-ule incorporating a window function is employed to maintain continuous dependencies at similar timestamps. The results from our experiments conducted on two real-world datasets demonstrate that the proposed model significantly outperforms leading baseline methods in recommendation effectiveness.