MOAFS: A Massive Online Analysis library for feature selection in data streams

Abstract

Each feature selection algorithm performs efficiently depending on different circumstances, such as data dimensionality (low, medium, high or ultra), speed rate, attribute type (nominal or numerical), number of classes, among others. Therefore, to perform different experiments on some of the most relevant feature selection algorithms proposed for data streams classification problems, the Massive Online Analysis Feature Selection (MOAFS) was created. MOAFS is a library for the Massive Online Analysis (MOA) (Bifet, Holmes, Kirkby, & Pfahringer, 2010) framework, and it is based on the MOAReduction (Ramírez-Gallego, Krawczyk, García, Woźniak, & Herrera, 2017) extension. It contains seven feature selection algorithms to be used as dimensionality reduction techniques in data streams classification problems, especially in the text-domain field, since they are not directly available on MOA. MOAFS includes incremental versions of information-based filter algorithms such as Information Gain and Gain Ratio (Quinlan, 1986), Symmetrical Uncertainty used by the Fast Correlation-Based Filter (Yu & Liu, 2003), Chi-squared (Pearson, 1992) and Cramers V-Test (Cramer, 1946), as well as wrapper algorithms such as Online Feature Selection (Wang, Zhao, Hoi, & Jin, 2014) and Extremal Feature Selection (Carvalho & Cohen, 2006). Therefore, MOAFS is a package for MOA to perform feature selection in data streams classification problems.