Two-phase network generation towards within-network classifiers evaluation

Within-network classifiers have been widely used to predict unknown data in networks. In order to evaluate the performance of existing classifiers, it is essential to generate synthetic networks with various properties. However, conventional network generation methods become ineffective under this scenario, since they are unable to produce node labels, exert topological constraints, or provide stable generation performance. In this paper, we propose a novel network generation method for evaluating within-network classifiers, which consists of two generation phases. In the first phase of topology generation, network topology can be obtained by incorporating any existing topology generation models. In the second phase of label generation, we model the problem as a multi-objective optimization. Specifically, we prove that generating node labels over an existing topology conforming homophily constraint is NP-hard, and devise a genetic algorithm based strategy for node label generation. Extensive experiments demonstrate that our method can produce synthetic networks with stable properties, and ensure that the network topology is fixed and label parameters take effect independently, thus making it sufficient for evaluating the sensitivity of classifiers against different parameters.