Evaluation of novelty detection algorithms for multi-label data streams classification

Abstract

Data streams (DS) are potentially unlimited sequences, generated continuously, non-stationary and, in many cases, at high speed. There are several studies currently investigating the application of Novelty Detection (ND) techniques in DS. This is because ND is an important task, mainly because new concepts may appear, disappear or evolve over time. Most of the work found in the novelty detection literature, presents novelty detection as a multi-class classification task, however for Multi-Label Classification (MLC) problems, ND has not yet been exploited. Still, there is a lack of multi-label data stream classification works and, none of these works, deals with important DS classification restrictions like scarcity of labeled data and concept evolution. Another challenge is how to evaluate the performance of these techniques, mainly because of the unknown examples and unsupervised phase, which generates novelty patterns without an association with the true classes. In this research, we plan to develop a new evaluation methodology, based on the proposed in de Faria et al. (2015), for multi-label data stream classification with novelty detection methods. (AU)