Optimizing Kernel Functions Using Transfer Learning from Unlabeled Data

Abstract

In this paper, we propose an approach to learn the kernel which uses transferred knowledge from unlabeled data to cope with situations where training examples are scarce. In our approach, unlabeled data has been used to construct an optimized kernel that better generalizes on the target dataset. For the proposed kernel learning algorithm, Fisher Discriminant Analysis (FDA) is used in conjunction with Maximum Mean Discrepancy (MMD) test of statistics to optimize a base kernel using labeled and unlabeled data. Thereafter, the constructed kernel from both labeled and unlabeled datasets is used in SVM to evaluate the results which proved to increase prediction accuracy.