On Social Network-Based Algorithms for Data Stream Clustering

Abstract

Extracting useful patterns from data is a challenging task that has been extensively investigated by both machine learning researchers and practitioners for many decades. This task becomes even more problematic when data is presented as a potentially unbounded sequence, the so-called data streams. Albeit most of the research on data stream mining focuses on supervised learning, the assumption that labels are available for learning is unverifiable in most streaming scenarios. Thus, several data stream clustering algorithms were proposed in the last decades to extract meaningful patterns from streams. In this study, we present three recent data stream clustering algorithms based on insights from social networks’ theory that exhibit competitive results against the state of the art. The main distinctive characteristics of these algorithms are the following: (1) they do not rely on a hyper-parameter to define the number of clusters to be found; and (2) they do not require batch processing during the offline steps. These algorithms are detailed and compared against existing works on the area, showing their efficiency in clustering quality, processing time, and memory usage.