Human Guided Multi-proportions Topic Model for Rare Event Detection Without Using Labels

Rare events are subtle in their presence but significant in their impact. However, subtle presence coupled with lack of labels makes it hard to detect rare events. There is no existing method that can detect a predefined rare event without using labels at all. In this paper, we propose a mechanism to employ human guidance in a topic modeling setup to avoid labels, where events are modeled as topics. In contrary to the standard topic models, we use multiple proportions over topics to emphasize rare topics. Following Bayesian principle, we introduce expert information to bias the inference mechanism of the proposed topic model. The expert information required is in the order of number of rare events to be detected which is significantly less compared to the labeling requirement of the order of samples in standard methods. The overall process turns out to be a simple mechanism based on Gibbs sampling. We demonstrate the efficacy of the proposed approach on texts as well as images, and found to achieve significant performance.