Establishing Interconnections of Similarity-Based Classifiers for Multi-label Learning with Missing Labels

Sambhav Jain, Reshma Rastogi

Abstract: Similarity-based classification framework is extensively used to address the problem of multi-label learning. Through this research, we establish the connection between similarity-based classification with many popular state-of-the-art multi-label learning models. In fact, we show that the similarity-based classification framework shares homology with Support Vector Machines. Further, we explore the application of the proposed framework to deal with the problem of multilabel learning with missing labels. Our models require only the plugin estimates for sample-sample and label-label similarity, which are coheased into a single term that leads to a parameter-free optimisation problem. Hence, The proposed model design is simple and time-efficient. Further, we have proposed two transductive models and two lazy learners which can be used as per applicability. The models shows competitive performance with other state-of-the-art models across five well-known multilabel datasets.