Hill-climber based fuzzy-rough feature extraction with an application to cancer classification

Real-world problems are often imprecise and redundant thereby create difficulty in taking decisions accurately. In recent past, rough set theory has been used for predicting potential genes responsible for causing cancer using discrete dataset. But discretization of data makes the dataset inconsistent by loosing information. To overcome this problem, this paper presents an efficient approach to predict the dominant genes using fuzzy-rough boundary region-based feature selection in combination with a heuristic hill-climber search method. But hill-climber search method produces subsets that contain redundant features. This problem is addressed using fuzzy-rough boundary region-based method that finds the reduct by minimizing the total uncertainty degree of the dataset to achieve faster convergence. Hill-climber based fuzzy-rough boundary region generates fuzzy decision reducts, which represent the minimal set of non-redundant features, capable of discerning between all objects. In this work, we attempt to introduce a prediction scheme that combines the proposed filter method with three different rule classifiers such as JRIP, Decision Tree and PART. We demonstrate the performance by two benchmark microarray datasets and the results show that our proposed method significantly reduce the dimensionality while preserving the classification accuracy.