A Game-Theoretic Approach to Graph Transduction: An Experimental Study

Abstract

An important class of problems in machine learning is based on semi-supervised process to look for consistent labeling for a set of objects. The crucial aspect which features this topic is how to spread the available knowledge on the data to infer a proper answer for the unknown one. In this dissertation we use an important graph based model of semi-supervised learning known as Graph Transduction. Our research begins from a novel solution in Game Theory which models the problem just a noncooperative game: the Graph Transduction Game (GTG). We study this algorithm for the specific instance of visual world, leading the analysis over the main information of visual similarity between images and measures of similarity projected to categories. Finally we introduce the implications arisen on large-scale classification, guessing a possible solution generalizing the current version of GTG, which can exploit of category similarities actively in the learning.