Authors: Saurabh Singh; Madhavi Sinha
Addresses: Department of Computer Science and Engineering, Birla Institute of Technology, Jaipur, India ' Department of Computer Science and Engineering, Birla Institute of Technology, Jaipur, India
Abstract: Feature weighting has attracted lots of researchers in pattern classification and data mining community. To select an appropriate subset of features, original feature set is first weighted according to its importance in determining its class label, and then using a threshold, a subset of features is obtained. Some adequate research works have been reported, which used feature weights for clustering by k-means algorithm. Few such algorithms have been developed, but as far as best of our knowledge, weighted pattern vectors have not been used by any pattern recognition algorithm especially in biometrics. In this paper, we formulate a unique pattern specific feature based weighting method, which weights the features of an individual pattern vector on the basis of their discriminative power to recognise that pattern. Proposed approach yields FAR = 0% at FRR = 0.2 for Euclidian distance and 0.02% and 0.3 respectively for cosine similarity measure.
Keywords: feature selection; feature weighting; pattern recognition; machine learning; keystroke dynamics.
International Journal of Applied Pattern Recognition, 2018 Vol.5 No.1, pp.1 - 10
Received: 28 Apr 2017
Accepted: 22 Dec 2017
Published online: 09 Mar 2018 *