Evolving clustering algorithm based on mixture of typicalities for stream data mining

摘要

Many applications have been producing streaming data nowadays, which motivates techniques to extract knowledge from such sources. In this sense, the development of data stream clustering algorithms has gained an increasing interest. However, the application of these algorithms in real systems remains a challenge, since data streams often come from non-stationary environments, which can affect the choice of a proper set of model parameters for fitting the data or finding a correct number of clusters. This work proposes an evolving clustering algorithm based on a mixture of typicalities. It is based on the TEDA framework and divide the clustering problem into two subproblems: micro-clusters and macro-clusters. Experimental results with benchmarking data sets showed that the proposed methodology can provide good results for clustering data and estimating its density even in the presence of events that can affect data distribution parameters, such as concept drifts. In addition, the model parameters were robust in relation to the state-of-the-art algorithms.