Possibilistic Approach For Novelty Detection In Data Streams

In many real-world applications data arrive continuously, in the form of streams. Such data can be used for the acquisition of knowledge by machine learning methods. In data streams learning, novelty detection is a relevant topic, which aims to identify the emergence of a new concept or a drift in the known concept in real time. Most approaches in the literature that focus on the novelty detection problem, make assumptions that limit the method usefulness. For instance, some methods are designed lying on the supposition that labeled data will be available at some time in the stream, while others restrict the proposed algorithm to one-class problems. Some recent approaches aim to overcome the limitations mentioned, considering multiclass problems and unlabeled data streams. In addition, there are also proposals that explore concepts of fuzzy set theory to add more flexibility to the learning process, although restricted to labeled streams. In this paper, we propose a method for novelty detection in data streams called Possibilistic Fuzzy multiclass Novelty Detector for data streams (PFuzzND). Our methods generates models based on a proposed summarization structures named SPFMiC (Supervised Possibilistic Fuzzy Micro-Cluster), which provides flexible class boundaries, allowing the identification of different types of novel information, i.e, novel classes, extension of classes or outliers more efficiently. Experiments show that our approach is promising in dealing with the changes in data streams and presents improvements in comparison to the non-fuzzy version.