A Novel Algorithm for Multi-label Classification by Exploring Feature and Label Dissimilarities

Abstract

Selection of appropriate nearest neighbors greatly affects predictive accuracy of nearest neighbor classifier. Feature similarity is often used to decide the set of k nearest neighbors. Predictive accuracy of multi-label kNN could further be enhanced if in addition to the feature similarity, difference in feature values and dissimilarity of the instance labels are also taken into account to decide the set of k nearest neighbors. This paper deals with an algorithm called “ML-FLD” that not only takes into account features similarity of the instances, but also considers feature difference and label dissimilarity in order to decide the k nearest neighbors of a given unseen instance for the prediction of its labels. The algorithm when tested using well-known datasets and checked with the existing well known algorithms, provides better performance in terms of example based metrics such as hamming loss, ranking loss, one error, coverage, average precision, accuracy, F measure as well as label-based metrics like macro-averaged and micro-averaged F measure.