engleski

An empirical study of classification algorithms when dealing with the problem of class imbalance and other data intrinsic characteristics

Evaluating and comparing the performance and behaviour of different algorithms is a pivotal step when applying machine learning in various application domains. Nevertheless, learning the concepts of real-world problems is a challenging task because of the different intrinsic characteristics that may be present in such datasets. Since not all machine learning algorithms are made equal, these characteristics do not affect their behaviour uniformly. This paper presents a large-scale empirical study of four different types of classifiers in which we try to determine and rank the degrees of correlation between their performance and the level of class imbalance, data rarity, small disjuncts, class overlapping and noise, and provide insight into classifier behaviour when faced with these problems.

class imbalance ; class overlapping ; small disjuncts

nije evidentirano