Incremental Learning of Fuzzy Decision Trees for Streaming Data Classification

Data stream analysis is growing in popularity in the last years since several application domains require to continuously and quickly analyse data produced by sensors with the aim of, for instance, reacting immediately when problems arise, or detecting new trends. The specificity of these domains imposes strict temporal constraints on machine learning algorithms to be used for mining useful insights. The Hoeffding Decision Tree (HDT) is a well-known classification algorithm for efficient streaming data classification. In this paper, with the aim of improving HDT accuracy and capability of handling noisy data, we exploit the learning procedure proposed in HDT for adapting a recently proposed fuzzy decision tree to cope with streaming data classification problems. We tested the fuzzy approach on a benchmark dataset for the on-line learning of data stream classification models. Results show that, during the on-line learning process, the fuzzy approach outperforms HDT in terms of accuracy