Transductive learning via validation risk minimization for semi-supervised learning

Transductive learning aims to recover the labels of unlabeled samples in a dataset with several labeled samples. The empirical risk minimization principle is commonly adopted in transductive learning. In this paper, the validation risk minimization principle is proposed and applied in transductive learning. The key idea is to introduce the validation strategies of supervised learning into transductive learning, leading to two learning frameworks with a general learning machine. Generally, these frameworks result in bilevel programming problems. Further, by using the least squares support vector machine as the learning machine in the two frameworks, we propose two transductive learning models which degenerate the bilevel programming problems into linear or quadratic integer programming problems. New algorithms are respectively designed to solve these problems efficiently. Preliminary experiments on synthetic, benchmark, and real-world datasets show that the proposed models achieve competitive transductive performance.