Multi Class Semi-Supervised Classification with Graph Construction Based on Adaptive Metric Learning

Abstract

This paper proposes a graph based Semi-Supervised Learning (SSL) approach by constructing a graph using a metric learning technique. It is important for SSL with a graph to calculate a good distance metric, which is crucial for many high-dimensional data sets, such as image classification. In this paper, we construct the similarity affinity matrix (graph) with the metric optimized by using Adaptive Metric Learning (AML) which performs clustering and distance metric learning simultaneously. Experimental results on real-world datasets show that the proposed algorithm is significantly better than graph based SSL algorithms in terms of classification accuracy, and AML gives a good distance metric to calculate the similarity of the graph. In eight benchmark datasets, 1 to 11 percent is attributed to the improvement of classification accuracy of state of the art graph based approaches.