StreamXM: An Adaptive Partitional Clustering Solution for Evolving Data Streams

A challenge imposed by continuously arriving data streams is to analyze them and to modify the models that explain them as new data arrives. We propose StreamXM, a stream clustering technique that does not require an arbitrary selection of number of clusters, repeated and expensive heuristics or in-depth prior knowledge of the data to create an informed clustering that relates to the data. It allows a clustering that can adapt its number of classes to those present in the underlying distribution. In this paper, we propose two different variants of StreamXM and compare them against a current, state-of-the-art technique, StreamKM. We evaluate our proposed techniques using both synthetic and real world datasets. From our results, we show StreamXM and StreamKM run in similar time and with similar accuracy when running with similar numbers of clusters. We show our algorithms can provide superior stream clustering if true clusters are not known or if emerging or disappearing concepts will exist within the data stream.