Safe semi-supervised learning from risky labeled and unlabeled samples

Abstract

In machine learning community, Safe Semi-Supervised Learning (S3L) has received the growing concerns. A few S3L methods are proposed to safely exploit risky unlabeled samples which may lead to performance degeneration of Semi-Supervised Learning (SSL). However, these methods have the following drawbacks: (1) The risk of unlabeled samples are artificially determined by analyzing different characteristics between SSL and Supervised Learning (SL); (2) They do not consider the negative impact of labeled samples. In this paper, we present l 1 -norm based S3L which can safely exploit the risky labeled and unlabeled samples. In our algorithm, we employ l 1 -norm based loss function to build an optimization problem for S3L. For the purpose of solving the l 1 -norm optimization problem, we introduce an effective iterative optimization strategy. In the iterative process, the samples' weights (including labeled and unlabeled ones) which can reflect the corresponding risk degrees are adaptively estimated. Therefore, it is expected to simultaneously reduce the negative effect of both labeled and unlabeled samples. Finally, we carry out the experiments on benchmark datasets and our algorithm can yield the desired learning performance.