Sequence-aware similarity learning for next-item recommendation

Sequence-aware next-item recommendation has recently been studied because of the noteworthy usefulness of the sequential information integrated into recommendation algorithms. Following the development thread of sequential recommendation methods, especially the factored Markov chains segment, and based on a well-designed fusing similarity model with factored high-order Markov chains (i.e., Fossil), we propose a novel and generic similarity learning framework for next-item recommendation called sequence-aware factored mixed similarity model (S-FMSM), which contains two variants with pairwise preference learning and pointwise preference learning. Unlike the baseline methods that model the general representation and the sequential representation in two divided factorization components, we use a factored mixed similarity model that unites the general similarity and the sequential relationship between two successive items for their sequential representation learning. Experiments on six datasets show that our newly introduced general similarity can notably improve the results of the recommended ranking lists. Furthermore, a study on tuning the prior trade-off parameter indicates the importance of the general similarity on different datasets. We adjust the number of latent dimensions and try different similarity measurements, which showcases that our S-FMSM is universal and gives us more insight on it.