GPS: Factorized group preference-based similarity models for sparse sequential recommendation

One of the key tasks for recommender systems is the prediction of personalized sequential behavior. There are two primary means of modeling sequential patterns and long-term user preferences: Markov chains and matrix factorization, respectively. Together, they provide a unified approach to predicting user actions. In spite of their strengths in tackling dense data, however, these methods struggle with the sparsity issues often present in real-world datasets. In approaching this problem, we propose combining similarity-based methods (demonstrably helpful for sequentially unaware item recommendation) with Markov chains to offer individualized sequential recommendations. This approach, called GPS (a factorized group preference-based similarity model), further leverages the idea of group preference along with user preference to introduce a greater array of interactions between users—which in turn eases the problem of data sparsity and cold users and cuts down on the assumption of a strong independency within various factors. By applying our method to a range of large, real-world datasets, we demonstrate quantitatively that GPS outperforms several state-of-the-art methods, particularly in cases with sparse datasets. Regarding qualitative findings, GPS also grasps personalized interactions and can provide recommendations that are both on-target and meaningful.