Improving recommendation quality through outlier removal

Rating data collected by recommendation systems contain noise caused by human uncertainty and malicious attacks. Existing outlier removal approaches usually aim at detecting noise inserted into ground-truth ratings. However, in real applications, the ground-truth of the training data are unavailable, or even unimportant for the prediction task. In this paper, we propose an efficient and effective outlier removal algorithm to improve the quality of the training data. The noise is modeled by the mixture of Gaussian distribution, which can approximate any continuous distribution. First, we employ the expectation-maximization algorithm to calculate the low-rank matrices, whose product forms the recovered ratings. Second, we compare the original and recovered ratings to solicit suspected outliers. This process is repeated a number of times, and ratings that are suspected enough times will be treated as outliers. To validate the effectiveness of our algorithm, we compared the prediction quality of four popular recommendation algorithms. Results showed that several measures on the algorithms were improved with the new training data.