Unsupervised link selection in networks

Abstract

Real-world networks are often noisy, and the existing linkage structure may not be reliable. For example, a link which connects nodes from different communities may affect the group assignment of nodes in a negative way. In this paper, we study a new problem called link selection, which can be seen as the network equivalent of the traditional feature selection problem in machine learning. More specifically, we investigate unsupervised link selection as follows: given a network, it selects a subset of informative links from the original network which enhance the quality of community structures. To achieve this goal, we use Ratio Cut size of a network as the quality measure. The resulting link selection approach can be formulated as a semi-definite programming problem. In order to solve it efficiently, we propose a backward elimination algorithm using sequential optimization. Experiments on benchmark network datasets illustrate the effectiveness of our method.