Bayesian flexible mixture model with rating conversion on multi-criteria recommender system

Abstract

PurposeMulti-criteria recommender systems (MC-RSs) allow users to express their preference in multiple aspects. Bayesian flexible mixture model (BFMM) is a model-based RS which extends FMM from single-criterion to MC. However, results of BFMM have a preference on different rating pattern problem. In single-criterion, FMM with decoupled normalization and W’s transposed function try to solve this problem. However, these techniques are applied to each criterion separately. Then, the relationship among criteria will be lost. This paper aims to solve different rating pattern problems and loss of the relationship between criteria.Design/methodology/approachThe proposed method is combining between BFMM and rating conversion. First, mean and variance normalization is applied to make MC ratings of an active user and a neighbor lying on the same plane. After that, a pattern of each user is extracted using principal component analysis (PCA). Next, the pattern is used to convert neighbors’ MC ratings to the active user aspect. After that, converted MC ratings of neighbors are aggregated to be overall ratings using multiple linear regression (MLR). Finally, overall rating of the active user toward the target item is predicted using weighted average on the derived neighbors’ overall ratings where the similarity from BFMM acts as a weight.FindingsThe experimental results show that the proposed method where all criteria ratings are converted simultaneously can improve the performance of recommendation.Originality/valueThe proposed method predicts overall rating of the active user by converting MC ratings of each neighbor to the active user aspect at the same time, which can reduce the loss of the relationship between criteria.