Contribution to Relational Classification with Homophily Assumption

Relational classification is a set of methods employing relations between instances in a dataset as well as their attributes. Homophily is a phenomenon present in graphs which capture real-world data, e.g., social connections between humans. Homophily is defined as following: related (neighbouring) vertices are more likely to share similarities (e.g., the same class, attribute value) as non-related instances. Contemporary relational classifiers implicitly require homophily to be present in a graph (so called homophily assumption), however these methods are unable to determine the homophily of each node and take benefit of this information. Our work is at first dedicated to classification of relational classifiers. Next, impact of homophily assumption on particular branches of relational classifiers is analyses and then homophily measures are defined. According to this analysis, two new relational classifiers are designed. First method belongs to simple relational methods and employs local graph ranking in order to redefine neighbourhood function, second method is collective inference based and involves information exchange moderation. Both methods are capable to increase the quality of class assignment in networked data due to their capability to employ and measure homophily in a graph.