Detecting emerging and evolving novelties with locally adaptive density ratio estimation

摘要

In today’s dynamic environment, there naturally exist two types of novelties: emerging and evolving. Emerging novelties are represented by concepts which are completely different from previously seen instances, while evolving novelties are characterized by relatively new aspects of existing concepts. Most existing algorithms for novelty detection focus on detecting only one type of novelty, giving little or no attention to the other. In real situations, the challenge is that these two types of novelties are not easily distinguishable and sometimes a truly novel concept does not fit perfectly under one of these categories. In this paper, a locally adaptive kernel density ratio method is proposed to capture the two characteristics in one formula. In specific, the density ratio between new and reference data is used to capture evolving novelties, and at the same time, the locally adaptive kernel is employed into the density ratio objective function to capture emerging novelties based on the local neighborhood structure. The effectiveness and robustness of the proposed method are demonstrated in the detection of novel handwritten digits and a set of benchmark novelty detection tasks. Additionally, we further examine its applicability in detecting novelties in social media data, which are characterized by a mixture of emerging and evolving topics over time.