Homophily of Neighborhood in Graph Relational Classifier

Quality of collective inference relational graph classifier depends on a degree of homophily in a classified graph. If we increase homophily in the graph, the classifier would assign class-membership to the instances with reduced error rate. We propose to substitute traditionally used graph neighborhood method (based on direct neighborhood of vertex) with local graph ranking algorithm (activation spreading), which provides wider set of neighboring vertices and their weights. We demonstrate that our approach increases homophily in the graph by inferring optimal homophily distribution of a binary Simple Relational Classifier in an unweighted graph. We validate this ability also experimentally using the Social Network of the Slovak Companies dataset.