Integrated Sequential Semantics with a Dual-Tower Time-Aware Recommendation Model

Sequence recommendation models aim to predict items a user may be interested in next by analyzing their interaction history sequence. These models often encounter two primary challenges: data sparsity due to short user interaction sequences and cold-start issues due to insufficient user-item rating data. Although user reviews contain rich semantic information that can mitigate these problems, most existing methods do not fully utilize this data. This paper introduces a novel twin-tower time-aware recommendation model (SSDTR) that integrates sequence semantics. The model extracts semantic features of users and products by leveraging a pretrained BERT model to filter reviews from product descriptions and users' historical reviews. Additionally, a sequence-level attention network is employed to extract sequence features from product IDs and filtered review IDs, effectively combining these to address data sparsity and cold-start issues. SSDTR uses a bi-Tower structure to independently model user and product features, capturing deep interaction patterns through multimodal feature fusion to enhance recommendation accuracy and interpretability. An LSTM network is also incorporated to analyze the temporal correlations between user interactions and time nodes for rating prediction. Experiments demonstrate that SSDTR outperforms baseline models in MSE, MAE, and NDCG@10, indicating its effectiveness in reducing data sparsity and improving recommendation performance.