Improving automatic edge selection for relational classification

In this paper, we address the problem of edge selection for networked data, that is, given a set of interlinked entities for which many different kinds of links can be defined, how do we select those links that lead to a better classification of the dataset. We evaluate the current approaches to the edge selection problem for relational classification. These approaches are based on defining a metric over the graph that quantifies the goodness of a specific link type. We propose a new metric to achieve this very same goal. Experimental results show that our proposed metric outperforms the existing ones.