Effective Semi-supervised Learning Based on Local Correlation

Abstract

Traditionally, the manipulation of unlabeled instances is solely based on prediction of the existing model, which is vulnerable to ill-posed training set, especially when the labeled instances are limited or imbalanced. To address this issue, this paper investigate the local correlation based on the entire data distribution, which is leveraged as informative guidance to ameliorate the negative influence of biased model. To formulate the self-expressive property between instances within a limited vicinity, we develop the sparse self-expressive representation learning method based on column-wise sparse matrix optimization. Optimization algorithm is presented via alternating iteration. Then we further propose a novel framework, named semi-supervised learning based on local correlation, to effectively integrate the explicit prior knowledge and the implicit data distribution. In this way, the individual prediction from the learning model is refined by collective representation, and the pseudo-labeled instances are selected more effectively to augment the semi-supervised learning performance. Experimental results on multiple classification tasks indicate the effectiveness of the proposed algorithm.