Abordagem fuzzy para detecção de novidade em fluxo contínuo de dados

Abstract

In recent years, we have witnessed the advent of computational systems capable of generating an immense amount of data in a short time period. These applications can be found in areas such as sensor networks, financial markets and computer networks. Systems that produce data incessantly, creating a continuous Data Stream (DS), can be infinite in size and can mutate in its statistical distribution over time. These DS can be used as sources for the automatic acquisition of useful knowledge by machine learning methods. However, the infinite and mutable nature of these data sets can essentially cause new concepts to emerge, which are examples that differ significantly from the examples learned by the model. Occurrences of this behavior in real-world applications may be credit card fraud or computer network intrusions. In this way, the task of detecting these examples, known as novelty detection, stands out as an important research topic. In general, classical methods for detecting novelty are not able to deal with the particularities of DS. Thus, different approaches have been proposed in order to create adaptable models that can accomplish this task. However, the unpredictable characteristics of DS's create difficulties in the learning process, encouraging the search for a more flexible learning. The integration of fuzzy set theory concepts is a timely way of making DS learning more adaptable to imprecisions. Recently, there have been proposals for machine learning models in DS based on fuzzy sets theory with the objective of collaborating for the flexibility and adaptability of the knowledge learned in DS's. Nonetheless, in the context of novelty detection the proposed approaches are few and limited to the domains of study. This paper presents a proposal for a fuzzy approach to detecting novelty in DS investigating techniques for detection of novelty in DS and machine learning models in DS based on fuzzy set theory. The analysis of the results, showed that the proposals favor the novelty detection task, facilitating the identification of discrepant data through the representation and treatment of imprecise data.