| Popis: |
Social and sequential recommendations employing bidirectional attention architecture represent a notable advancement in deep learning, enhancing recommender system performance. This breakthrough facilitates the representation learning of interactions, both forward and backward, concerning dynamic user preferences influenced at the social and sequential levels. Despite previous research efforts to accurately model user preference changes, they encounter two primary shortcomings: 1) an insufficient to account for user preference shifts using unidirectional approaches, and 2) a disregard for the impact of social factors on user preference changes, leading to suboptimal outcomes. In this study, we propose a novel framework, the Bidirectional Hyperbolic Graph Capsule Co-Attention Network (BiHGCA), addressing the challenges posed by social and sequential dynamics in user preference changes. This model integrates a bidirectional hyperbolic graph attention network and a bidirectional capsule attention network with a newly designed gate, T-BiGRUDense, which stands for Time-aware Bidirectional Gated Recurrent Unit and Dense neural network. The bidirectional hyperbolic graph attention network aims to capture shifts in user preferences at the social level. Simultaneously, the bidirectional capsule attention network is tailored to model user preference dynamics at the sequential level. Moreover, T-BiGRUDense is innovatively crafted to merge the collaborative attention signals from both the social and sequential levels, enhancing next movie recommendation predictions. Empirical evaluations conducted on movie benchmark datasets illustrate the superiority of BiHGCA over existing state-of-the-art methods in delivering Top-N recommendations and its effectiveness across various levels of data sparsity. |
