Abstract/Description

Quality of features determines the maximum achievable accuracy by any arbitrary classifier in pattern classification problem. In this paper, we have proposed an index that can assess the quality of features in discrimination of patterns in different classes. This index is in-sensitive to the complexity of boundary separating different classes if there is no overlap among features of different classes. Proposed index is model free and requires no clustering algorithm to discover the clustering structure present in the feature space. It is only based on the information of local neighborhood of feature vectors in the feature space. This index can be used to predict the classification accuracy and density of feature vectors of a class in the feature space. Implementation of the index is simple and time efficient. Performance of Arif index on different benchmark physiological data sets is found to be in consistent with the reported accuracies in the literature. Hence this index will be very useful in providing prior useful information about the quality of features before designing any classifier.

Session Theme

Data Mining

Session Type

Other

Session Chair

Dr. Sajjad Haider

Start Date

15-8-2009 5:55 PM

End Date

15-8-2009 6:15 PM

Share

COinS
 
Aug 15th, 5:55 PM Aug 15th, 6:15 PM

Data Mining: Assessment of features quality of class discrimination using arif index and its application to physiological datasets

Quality of features determines the maximum achievable accuracy by any arbitrary classifier in pattern classification problem. In this paper, we have proposed an index that can assess the quality of features in discrimination of patterns in different classes. This index is in-sensitive to the complexity of boundary separating different classes if there is no overlap among features of different classes. Proposed index is model free and requires no clustering algorithm to discover the clustering structure present in the feature space. It is only based on the information of local neighborhood of feature vectors in the feature space. This index can be used to predict the classification accuracy and density of feature vectors of a class in the feature space. Implementation of the index is simple and time efficient. Performance of Arif index on different benchmark physiological data sets is found to be in consistent with the reported accuracies in the literature. Hence this index will be very useful in providing prior useful information about the quality of features before designing any classifier.