The Influence of Inconsistent Data on Cost-Sensitive Classification Using Prism Algorithms: An Empirical Study

Abstract

Cost-sensitive classification is one of the hottest research topics in data mining and machine learning. It is prevalent in many applications, such as Automatic Target Recognition (ATR), medical diagnosis, etc. However, the data in practice may be inconsistent due to dimensional reduction operation or other pre-processing, yet it is not clear how the inconsistent data affects cost-sensitive learning. This paper presents an empirical comparative study using four Prism rule-generating algorithms with J-measure pruning, two of which are proposed in this paper. The most important result of our study is that inconsistent data dose often affects the performance of cost-sensitive Prism classifiers, and in the inconsistent data setting, merely a single Prism classifier’s robustness cannot completely satisfy the requirements of cost-sensitive systems.