Constrained Local Regularized Transducer for Multi-Component Category Classification

Abstract

Transductive learning is proposed to incorporate both labeled and unlabeled examples into the learning process. Several methods have been developed and show encouraging performance. However, people may meet complicated classification tasks in real world applications, where one category contains multiple components. Traditional transductive learning algorithms are not very effective in such settings. In this paper, we propose a novel transductive learning approach called constrained local regularized transducer(CLRT) for multi-component category classification. CLRT is based on the local separable assumption that it is possible to build a linear predictor in one small area. We implement the assumption by minimizing a unified objective function, which can be optimized globally. Experiment results validate that CLRT can achieve satisfied performance robustly and efficiently.