Link prediction in temporal networks: Integrating survival analysis and game theory

Abstract Link prediction is an important task in complex network analysis and can be found in many real-world applications such as recommendation systems, information retrieval, and marketing analysis of social networks. This paper focuses on studying the evolution mechanism of real-world temporal networks. Specifically, given a set of temporal links during a fixed time window, how to predict the existence of links at any point in the future. To address this problem, we propose a novel semi-supervised learning framework, which integrates both survival analysis and game theory. First, we carefully define the ϵ-adjacent network sequence, and make use of time stamp on each link to generate the baseline network evolution sequence. Next, to capture the law of network evolution, we employ the Cox Proportional Hazard Model (Cox PHM) to study the relative hazard associated with each temporal link, so as to estimate the coefficients of covariates, which are defined as a set of neighborhood based proximity features. To narrow the area of inquiry, we further propose a game theory based two-way selection mechanism to predict the future network topology. We finally encapsulate these two new technologies in a robust Autonomy-Oriented-Computing (AOC) multi-agent system, and propose a paralleled algorithm to conduct the temporal link prediction task. Extensive experiments were applied to real-world temporal networks to demonstrate both effectiveness and scalability of the proposed approach.