Merging Clusters in Summary Structures for Data Stream Mining based on Fuzzy Similarity Measures

Abstract

Fuzzy Clustering is one of the mining techniques that have been used to extract information from Data Streams.The d-FuzzStream algorithm is a fuzzy version of the Online-Offline Framework, which consists of two steps: an online step, where a summary structure formed by fuzzy microclusters is built and an offline step, where the micro-clusters are clustered in batch mode. The quality of the data summary depends on the criteria used to decide whether an example starts a new micro-cluster or is absorbed by the existing ones; and whether two microclusters became similar enough to be merged. In d-FuzzStream algorithm such decisions are based on concepts of fuzzy dispersion and a distance-based fuzzy clusters similarity. In this paper we investigate the behavior of different fuzzy similarity measures on the decision of merging two fuzzy micro-clusters during the online step. Experiments were run using five synthetic data sets and four fuzzy similarity measures. The results obtained are analyzed and discussed through informative and purity measures.