Feature Selection for Handling Concept Drift in the Data Stream Classification

Abstract

With the advance in both hardware and software technologies, streaming data is ubiquitous today, and it is often a challenging task to store, analyze and visualize such rapid large volumes of data. One of difficult problems in the data stream domain is the data streams classification problem. The traditional classification algorithms have to be adapted to run in a streaming environment because of the underlying resource constraints in terms of memory and running time. There are at least three hard aspects in the data streams classification: large length, concept drift and feature selection. Concept drift is a common attribute of data streams that occurs as a result of changes in the underlying concepts. Feature selection has been extensively studied from a conventional mining perspective, but it is a much more challenging problem in the data stream domain. The concept drift and large length make impossible applying classical feature selection methods in the learning procedure. This paper proposes a new Bayesian framework to feature selection in data streams pattern recognition problem. We suggest a hierarchical probabilistic model with sparse regularization for estimation of decision rule parameters. The proposed approach gives a strong Bayesian formulation of the shrinkage criterion for predictor selection. Experimental results show that the proposed framework outperforms other methods of concept drift analysis.