Dynamic stream clustering using ants

Abstract

Data stream mining is the process of extracting knowledge from continuous sequences of data. It differs from conventional data mining in that a stream is potentially unbounded, data points arrive online and each data point can be examined only once. Furthermore, in non-stationary environments the statistical properties of the data can change over time. This paper presents a bio-inspired approach to clustering non-stationary data streams. The proposed algorithm, Ant-Colony Stream Clustering (ACSC), is based on the concept of artificial ants which identify clusters as nests of micro-clusters in dense areas of the data. Micro-clusters are N-dimensional spheres with a maximum radius ε. In ACSC the ε-neighbourhood, crucial in density clustering, is adaptive and doesn’t require expert, data-dependent tuning. The algorithm uses the sliding window model and summary statistics for each window are stored offline. Experimental results over real and synthetic datasets show that the clustering quality of ACSC is comparable or favourable to leading stream-clustering algorithms while requiring fewer parameters and considerably less computation.