A Deep Attribute-Induced Collaborative Filtering Recommendation Model with Attribute-Aware Attention and Item-Priority Modelling

The growth of digital applications has increased information overload, causing difficult for users to find appropriate content. Recommendation systems address this issue by generating customized recommendations based on individual user interests. However, most collaborative and hybrid recommendation methods mainly depend on the rating matrix and dot-product-based similarity computation, which captures only low-order linear interactions. Apart from that, when only ratings are available, collaborative filtering struggles to distinguish between limited data availability and genuine cold-start scenarios. To address these limitations, this paper introduces a deep recommendation framework, termed Deep Attribute-Induced Collaborative Filtering (DeepACF). This unified neural recommendation framework is designed to improve personalized recommendations by jointly modeling user information, item information, contextual attributes, and item-priority features within a single architecture. The proposed model consists of input, embedding, attention, and prediction modules. In the embedding layer, user IDs, item IDs, user-side features, item-side features, and contextual attributes are transformed into dense latent representations through dedicated encoders. In addition, a learned item-priority encoder is introduced to capture item age, popularity, cold-item status, and average rating tendency. These encoded representations are further refined through feature-level attention, which allows the model to allocate varying importance to the available features according to the user-item interaction context. The final representations are then fused and passed through a multilayer perceptron to generate the final predicted score. Experimental results on the Movie-Lens datasets show that DeepACF achieves superior rating prediction accuracy, obtaining RMSE/MAE values of 0.8376/0.6643 and 0.8081/0.6418, respectively. For Top-N recommendation, the model achieves competitive ranking performance, with Hit Rate/NDCG values of 0.7801/0.5504 on Movie-Lens 100K and 0.7985/0.6104 on Movie-Lens 1M.