Data preprocessing and enhanced label dependent learner for multilabel learning

Abstract

In recent years the need of multiple response variables analysis becomes vital in real world scenario Multi label classification plays a very important role in the task of multi label learning The multi label learning can be categorised into two ways Problem Transformation PT method and Algorithm Adaptation AA method PT methods transform the multi label problem into several single label problems and then conventional single label base learners are applied over the multiple single label problems In AA methods the conventional data mining learners are upgraded to adopt the multi label problem directly The goal of multi label learning is to predict a subset of targets for an instance from a test dataset In multi label learning the relationships among multiple targets are compulsory as the interactions of target variables might be useful for the accurate prediction The label space can contain high degree of information among them and ignorance of such information may greatly degrade the performance of the learning algorithms Most of the conventional multi label learning algorithms can handle these hindrances to some extent to improve the learning performance On other hand imbalanced nature of the multi label dataset might penalize learners performance hence balancing label distribution and incorporating label information can help the multi label learner for better classification in terms of both accuracy and efficiency Incorporating information among the labels and balancing the class distributions are the two important challenges in multi label and the solving steps for above two issues can help better classifier performance in terms of low hamming loss and high accuracy newline