Self-attention Convolutional Neural Network for Sequential Recommendation

Abstract

Sequential recommendation uses user-interaction history of preferred items to predict which items a user is most likely to interact with in future. To tackle this prediction problem, over the years, many researchers have developed different approaches such as Convolutional Neural Networks (CNN) to identify patterns in interaction history, Self-Attention recommendation systems to find the similarity of items in a sequence with each other to model the connections between items etc. Although these approaches provide promising results, there are still a lot of scope for improvements. One limitation of such approaches for Top-N recommendation is the inability to capture long range dependencies in the sequence of interactions. Another issue is the inability for the network to model hierarchical information. To mitigate such limitations of existing approaches, in this paper, we propose a new approach called Self-Attention Convolutional (SACORec) network which combines both CNNs and Self-Attention. The idea is to take advantage of the effectiveness of various methods while mitigating their limitations. To find the effectiveness of our proposed model, we ran our model on several public data sets and collected a variety of evaluation metrics. The empirical study shows that our proposed model significantly outperforms state-of-art architectures that implement only one paradigm to the sequential recommendation. The efficiency of our model has led us to believe that it will open more pathways for future research in this area.