Modification of Semi-supervised Algorithm Based on Gaussian Random Fields and Harmonic Functions

Abstract

In this paper we propose an improvement for a semi-supervised learning algorithm based on Gaussian random fields and harmonic functions. Semi-supervised learning based on Gaussian random fields and harmonic functions is a graph-based semi-supervised learning method that uses data point similarity to connect unlabeled data points with labeled data points, thus achieving label propagation. The proposed improvement concerns the way of determining similarity between two points by using a hybrid RBF-kNN kernel. This improvement makes the algorithm more resilient to noise and makes label propagation more locality-aware. The proposed improvement was tested on five synthetic datasets. Results indicate that there is no improvement for datasets with big margin between classes, however in datasets with low margin proposed approach with hybrid kernel outperforms existing algorithms with a simple kernel.