Sequential Recommendation Model Enhanced by Attribute Information-Driven Mechanism

To address the issues that existing sequential recommendation models perform poorly in data-sparse scenarios, and that most current data augmentation methods adopt random perturbation strategies while neglecting item attribute information, thereby causing semantic shift, this paper proposes an attribute-driven sequential recommendation enhancement model named ADAS-Rec. Aiming to enhance the accuracy and robustness of recommendation results, the core idea of the model is to impose explicit constraints on the data augmentation process using multi-dimensional item attribute information. To this end, four collaborative augmentation operators are specifically designed: attribute-similar substitution, attribute-similar insertion, attribute-guided cropping, and attribute-guided masking. Relying on these operators, the model can generate high-quality augmented sequences with diverse content and attribute semantics that align with users’ fine-grained preferences, thereby effectively alleviating the data sparsity problem. To further optimize the quality of sequential representations, the model incorporates a frequency-domain filtering module, which removes redundant noise from sequences and highlights key features, providing higher-quality input data for user preference modeling. Finally, the model adopts a dual-attention mechanism that integrates a guided fusion layer and a causal self-attention module, enabling precise mining of users’ dynamic preferences and deep temporal dependencies within the augmented and denoised sequences, and completes recommendation predictions based on these preferences and dependencies. 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 across evaluation metrics such as R@10 and NDCG@10.