Isolating critical data points from boundary region with feature selection

Abstract

Immense databases may contain critical instances or chunks-a small heap of records or instances which has domain specific information. These chunks of information are useful in future decision making for improving classification accuracy for labeling of critical, unlabeled instances by reducing false positives and false negatives. Classification process may be assessed based on efficiency and effectiveness. Efficiency is concerned with the time to process the records by reducing attributes in the data set and effectiveness is the improvement in classification accuracy using crucial information. This work focuses on reducing the attributes in the large databases, put forwards an innovative procedure for computing criticality which isolates critical instances from the boundary region and are validated using real-world data set. This work also uses different attribute reduction technique used for fetching the critical instances to reduce the computational time. Results of the experiments show that only subsets of instances are isolated as critical nuggets. It is found that use of attribute reduction technique decreases the computational time. The data set with reduced attributes does not affect the classification accuracy and produces the same result as with the original data set. It also reveals that these critical records helps in improving classification accuracy substantially along with reduced computational time and are validated using real-life data sets.