HOSLIM: Higher-Order Sparse LInear Method for Top-N Recommender Systems

Abstract

Current top-N recommendation methods compute the recommendations by taking into account only relations between pairs of items, thus leading to potential unused information when higher-order relations between the items exist. Past attempts to incorporate the higher-order information were done in the context of neighborhood-based methods. However, in many datasets, they did not lead to significant improvements in the recommendation quality. We developed a top-N recommendation method that revisits the issue of higher-order relations, in the context of the model-based Sparse LInear Method (SLIM). The approach followed (Higher-Order Sparse LInear Method, or HOSLIM) learns two sparse aggregation coefficient matrices S and S′ that capture the item-item and itemset-item similarities, respectively. Matrix S′ allows HOSLIM to capture higher-order relations, whose complexity is determined by the length of the itemset. Following the spirit of SLIM, matrices S and S′ are estimated using an elastic net formulation, which promotes model sparsity. We conducted extensive experiments which show that higher-order interactions exist in real datasets and when incorporated in the HOSLIM framework, the recommendations made are improved. The experimental results show that the greater the presence of higher-order relations, the more substantial the improvement in recommendation quality is, over the best existing methods. In addition, our experiments show that the performance of HOSLIM remains good when we select S′ such that its number of nonzeros is comparable to S, which reduces the time required to compute the recommendations.