A risk degree-based safe semi-supervised learning algorithm

Abstract

Semi-supervised learning has attracted much attention in machine learning field over the past decades and a number of algorithms are proposed to improve the performance by exploiting unlabeled data. However, unlabeled data may hurt performance of semi-supervised learning in some cases. It is instinctively expected to design a reasonable strategy to safety exploit unlabeled data. To address the problem, we introduce a safe semi-supervised learning by analyzing the different characteristics of unlabeled data in supervised and semi-supervised learning. Our intuition is that unlabeled data may be often risky in semi-supervised setting and the risk degree are different. Hence, we assign different risk degree to unlabeled data and the risk degree serve as a sieve to determine the exploiting way of unlabeled data. The unlabeled data with high risk should be exploited by supervised learning and the other should be used for semi-supervised learning. In particular, we utilize kernel minimum squared error (KMSE) and Laplacian regularized KMSE for supervised and semi-supervised learning, respectively. Experimental results on several benchmark datasets illustrate the performance of our algorithm is never inferior to that of KMSE and indicate the effectiveness and efficiency of our algorithm.