Link Prediction in Complex Networks by Supervised Rank Aggregation

Abstract

Link prediction is a central task in the field of dynamic complex network analysis. A major trend in this area consists of applying a dyadic topological approach. Most of existing approaches apply machine learning algorithms where the link prediction problem is converted into a binary classification task. In this work, we propose a new dyadic topological link prediction approach applying supervised social choice algorithm. Given a training graph observed over a period [t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> , t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> '], this interval is divided into two sub-intervals: the learning interval and the labeling one. For each unlinked couple of vertices in the learning interval, a topological feature vector is computed. The labeling interval is used to fix the class of each example (e.g. linking, not-linking). Instead of learning a classification model as it is the case when applying machine learning approaches, we use these data to learn weights to associate to each computed feature based on the ability of each attribute to predict observed links. These weights are then used within weighted/supervised computational social choice algorithms to predict new links at time t > t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> '. Two weighting schemes are experimented. We introduce weighted social choice rules by modifying classical voting approaches, namely: the Borda rule and the Kemeny aggregation rule. We also introduce our own concept of finding weights. We have implemented our approach on an academic coauthoring dataset (DBLP dataset). The preliminary results have been quite good, so we are working further on experimentation.