Effect of Distance Metrics on Multi-label Classification

Abstract

AbstractMulti-label framework is suitable to represent a lot of data in the real world. Various multi-label classifiers extract information from neighbors and utilize it for further process. However, instances in multi-label data are associated with multiple labels. Therefore, in addition to take into account feature similarity, considering labels also for finding neighbors of known instances may be useful to improve performance of multi-label classifier. This paper presents novel algorithm, called MLFLD-MAXP which is an extension of our algorithm MLFLD. It first searches neighbors using feature similarity as well as label dissimilarity and estimates probabilities using labeled data. It then predicts labels for unlabeled data. In case no label is predicted for an instance, a label with maximum probability is associated with such instances. Algorithms such as MLFLD-MAXP need to make use of certain distance metrics such as Euclidean and choice of distance metric may affect the performance of the classifier. This paper studies the impact of various distance metrics. Empirical evaluation shows improved performance of MLFLD-MAXP when compared with other multi-label classification algorithms. Use of Minkowski distance metric leads to superior performance of MLFLD-MAXP as compared to the use of other distance metrics.KeywordsClassificationMulti-labelFeature similarityDistance metricsNearest neighbor