Discriminative Link Prediction using Local Links, Node Features and Community Structure

TL;DR

This paper introduces a discriminative link prediction algorithm that combines community-level link density estimates with local node feature similarities, demonstrating significant accuracy improvements over existing methods.

Contribution

The paper presents a novel link prediction method integrating community densities and local features, along with a rigorous benchmarking protocol for evaluation.

Findings

Significant accuracy improvements over Adamic-Adar and random walk methods

Community-based link density estimates enhance prediction quality

The proposed method is the most robust among tested algorithms

Abstract

A link prediction (LP) algorithm is given a graph, and has to rank, for each node, other nodes that are candidates for new linkage. LP is strongly motivated by social search and recommendation applications. LP techniques often focus on global properties (graph conductance, hitting or commute times, Katz score) or local properties (Adamic-Adar and many variations, or node feature vectors), but rarely combine these signals. Furthermore, neither of these extremes exploit link densities at the intermediate level of communities. In this paper we describe a discriminative LP algorithm that exploits two new signals. First, a co-clustering algorithm provides community level link density estimates, which are used to qualify observed links with a surprise value. Second, links in the immediate neighborhood of the link to be predicted are not interpreted at face value, but through a local model of node feature similarities. These signals are combined into a discriminative link predictor. We evaluate the new predictor using five diverse data sets that are standard in the literature. We report on significant accuracy boosts compared to standard LP methods (including Adamic-Adar and random walk). Apart from the new predictor, another contribution is a rigorous protocol for benchmarking and reporting LP algorithms, which reveals the regions of strengths and weaknesses of all the predictors studied here, and establishes the new proposal as the most robust.