Article Preview
Top1. Introduction
Every day quintillions of bytes of data are created from various sources like sensor data pertaining to climate information, census information and agricultural information etc., that get posted to social media sites, the digital pictures and videos, sale and purchase transaction records and cell phone GPS signals etc. to name a few. Generation of huge amount of data from various sources may create lots of difficulties in learning by the system using various classifiers because of redundant features available in data sets (Jain, Duin & Mao, 2000). It is required to reduce the dimension of data sets in order to get significant and informative features for decreasing the cost, storage and process time for better classification and prediction. Various preprocessing techniques have been proposed by different authors time to time to tackle the big data but many of these techniques have become inadequate due to their own limitations. Feature selection is one of the well-known techniques in data preprocessing for data mining, machine learning, pattern recognition, bioinformatics and medical image processing (Duda, Hart & Stork, 1973; Zhu, Ong & Dash, 2007; Yu & Liu, 2004; Hall, 1999; Dash & Liu, 2003). Feature selection is the process of selecting those input features that are most predictive of a desired outcome. It is better than other dimensionality reduction techniques as it preserves the real meaning of the features after reduction.
Feature selection methods (Guyon & Elisseeff, 2003; Saeys, Inza & Larrañaga, 2007; Dash & Liu, 1997) can be categorized as filter, embedded, wrapper, unsupervised, semi- supervised and supervised (Bhatt &Gopal, 2005; Zhu, Ong & Dash, 2007; Liu & Yu, 2005), etc. Feature selection techniques can be divided into two categories, firstly, symbolic method and secondly, numerical method. Symbolic methods consider all features as categorical variables and numerical methods take all the features as real valued variables. If there exist any heterogeneous features, symbolic methods(such as Rough set (Pawlak, 1982; Pawlak, 2012) based feature selection) use a discretization approach (Ching, Wong & Chan, 1995) and convert them as symbolic features which may lead to some sort of assumption and cause information loss. Discretization may damage two types of structures, firstly neighborhood structure and secondly ordered structure in real space. Latter problem was handled by fuzzy rough set model, but it was not able to tackle the former problem. Very few researches have been proposed to deal with neighborhood structure of the data sets.