Rating personalization improves accuracy: A proportion-based baseline estimate model for collaborative recommendation

Abstract

Baseline estimate is an important latent factor for recommendations. The current baseline estimate model is widely used by characterizing both items and users. However, it doesn’t consider different users’ rating criterions and results in predictions may be out of recommendation’s rating range. In this paper, we propose a novel baseline estimate model to improve the current performance, named PBEModel (Proportion-based Baseline Estimate Model), which uses rating proportions to compute the rating personalization. The PBEModel is modeled as a piecewise function according to different rating personalization. In order to verify this new baseline estimate, we apply it into SVD++, and propose a novel SVD++ model named PBESVD++. Experiments based on six real datasets show that the proposed PBEModel is rational and more accurate than current baseline estimate model, and the PBESVD++ has relatively higher prediction accuracy than SVD++.