Towards Better Top-K Recommendation with Attention-Aware Collaborative Matrix Factorization

One of the key issues in top-K recommendation systems is how to evaluate the relevance of candidate items effectively. Conventional recommendation systems typically employ collaborative filtering techniques to compute the similarity or relevance between items. Though great success is being made by matrix factorization, as one of those most representative techniques, is claimed to be less effective in capturing the inherent sequential information between the interactions of user and item. Meanwhile, the self-attention mechanism specializes in handling sequence data, which makes it a wise option to compensate for such information loss in the above case. Until now, there has been little focus on combining these two techniques for the top- K recommendation systems. To this end, this paper introduces a hybrid deep model called Attention-Aware Collaborative Matrix Factorization (AA-CMF), which integrates the self-attention mechanism, neural networks, and matrix factorization to provide a better recommendation. By jointly learning user and item features in a unified framework, it enables a refined prediction of user preferences. More importantly, the model has the strong capability to capture the sequential relationship and dependencies between users and items, as well as among different items, thus boosting the recommendation performance. Extensive experiments were carried out on three real-world datasets, and the findings clearly indicate that AA-CMF surpasses state-of-the-art methods across multiple evaluation metrics. The empirical results also confirmed the effectiveness of the proposed model on sparse datasets.