A Comparative Study of Simple Online Learning Strategies for Streaming Data

Abstract

Since several years ago, the analysis of data streams has attracted considerably the attention in various research fields, such as databases systems and data mining. The continuous increase in volume of data and the high speed that they arrive to the systems challenge the computing systems to store, process and transmit. Furthermore, it has caused the development of new online learning strategies capable to predict the behavior of the streaming data. This paper compares three very simple learning methods applied to static data streams when we use the 1-Nearest Neighbor classifier, a linear discriminant, a quadratic classifier, a decision tree, and the Na¨ıve Bayes classifier. The three strategies have been taken from the literature. One of them includes a time-weighted strategy to remove obsolete objects from the reference set. The experiments were carried out on twelve real data sets. The aim of this experimental study is to establish the most suitable online learning model according to the performance of each classifier