This paper addresses an issue of training optimization of multi-aspect rating inference. First, to address the issue of author inconsistency rating annotation, this paper proposes two simple approaches to improving the standard rating inference models by optimizing sample selection for training, including tolerance-based selection and ranking-loss-based selection methods. Second, to explore correlations between ratings across a set of aspects, this paper presents an aspect-oriented collaborative filtering technique to improve rating inference models. Experiments on two publicly available English and Chinese restaurant review data sets have demonstrated significant improvements over standard algorithms.