Adaptive hybrid methods for Feature selection based on Aggregation of Information gain and Clustering methods

The growing abundance of information necessitates the need for appropriate methods for organization and evaluation. Mining data for information and extracting conclusions has been a fertile field of research. However data mining needs methods to preprocess the data. Feature selection is a growing field of interest about selecting proper information from information repositories. The aim of this paper is to highlight the need for feature selection methods in data mining encompassing the best characteristics of the data. In recent times there has been interest in developing hybrid feature selection methods combining the characteristics of various filter and wrapper methods. The proposed method advocates an adaptive aggregation strategy using a) the gain ratio for candidate features and b) clustering methods to find the distribution of candidate features. The underlying principle of the strategy is that the best individual features need not constitute the best sub-set of features representing the problem. A given feature might provide more information when present with certain other feature(s) than when considered by itself. The adaptive method has been implemented for the datasets from the UCI repository and results correlated. The conclusions show that the proposed method shows encouraging results.