Posterior self-information based uncertainty measurement for data classification and learning

Identifying reliable information from the information ocean is a natural talent of human being, which can be hardly formalized by machines. In this paper, we show how to measure the uncertainty of each sample in terms of their similarity. For applications under complex uncertainty, it is desirable that we provide a systematic way for users to identify what are reliable information. In our approach, Posterior self-information for different kinds of data sets is defined according to the Shannon entropy theory. To improve the performance of three learning algorithms under uncertainty, this method is introduced to them by weighting their loss function. On both simulation data set and several public data sets, we find our method leads to significant improvements when uncertainty is high, and preserves original performance when uncertainty can be ignored.