Multi-Label Classification Using Noise Reduction Technique

Abstract

In domain of data mining and machine learning, multi-label classification is widely studied research problem. The goal of multi-label classification is to predict the absence or presence certain labels of a particular applications those are associated with different classes. In this paper, IML-Forest method is presentedwith goal of improving the performance of multi-label classification over different types of datasets. IML-Forest is based on existing ML-Forest technique. In this paper the construction of set of hierarchical trees and designed the label transfer mechanism in order to identify multiple relevant labels in hierarchical way is proposedto solve the problem of label dependencies in multi label classification. Basically relevant labels at higher levels of trees capture the more discriminable label concepts; next they will be shifted at lower level nodes. From the hierarchy the relevant labels are further aggregated in order to compute the label dependency and make the classification prediction. The problem with ML-Forest method is that noise considerations not yet addressed as collected multi-label dataset may be noisy and imbalanced. This can degrade the performance of learning and accuracy. Noise reduction method is proposed on multi-label dataset to solve the problem of noisy and imbalanced dataset. In this paper the text noises related to low-level data errors are handled.