Abstract:

To address the poor performance of existing sequential recommendation models in data-sparse scenarios and the semantic deviation problem caused by random-perturbation-based data augmentation methods that often ignore item attributes, a side information-driven sequential recommendation enhancement model (ADAS-Rec) is proposed. The model aims to improve the accuracy and robustness of recommendations. Its core mechanism lies in explicitly guiding the data augmentation process using multi-dimensional item attribute information. Specifically, four synergistic operations are designed: attribute similarity substitution, attribute similarity insertion, and attribute-guided cropping and masking. These operations are intended to generate diverse, high-quality augmented sequences that are semantically consistent with users' fine-grained preferences at the attribute level, effectively alleviating the data sparsity problem. On this basis, to purify the sequence signal, a frequency-domain filtering module is integrated to provide a better input for subsequent preference learning by suppressing sequence noise and enhancing key signal features. Finally, a dual attention mechanism, comprising a guided focus layer and a causal self-attention module, is employed to accurately capture the dynamic user preferences and deep temporal dependencies embedded in the enhanced and filtered sequences, thereby generating precise recommendations. Experimental results on three public datasets, Beauty, Sports, and Yelp, demonstrate that the proposed ADAS-Rec model significantly outperforms various state-of-the-art baseline methods on common evaluation metrics such as Recall@10 and NDCG@10.

Key words: sequential recommendation, data augmentation, side information