Improving relational classification using link prediction techniques

In this paper, we address the problem of classifying entities belonging to networked datasets. We show that assortativity is positively correlated with classification performance and how we are able to improve classification accuracy by increasing the assortativity of the network. Our method to increase assortativity is based on modifying the weights of the edges using a scoring function. We evaluate the ability of different functions to serve for this purpose. Experimental results show that, for the appropriated functions, classification on networks with modified weights outperforms the classification using the original weights.