Combining feature wrapper and filter in a novel evolutionary based feature extraction approach

In pattern recognition and data mining problems, a set of measurable features are required to describe the input objects. The quality of these features in describing the input instance has a direct impact on the success and on the accuracy of the system. Feature extraction is a process for deriving fewer new features than the prior input vectors in order to achieve comparable accuracy with lower cost of feature measurement. In this process, increasing the efficiency could also be considered. In other words feature extraction could help us transform the input space into a decision space in which objects are better discriminated in this space. In this study, we present a new approach to feature extraction using the power of evolutionary algorithm family in optimization to find a linear transformation to a decision space with lower cost of computation and a higher level of accuracy. A combination of feature wrapper and filter methods are used as the evaluation criterion. Mutual information is used as a measure of quality of transformed features and the quality of transformation system is evaluated based on predictive accuracy of the classifier which is in use.