[Paper Review] Predicting Positive and Negative Links in Online Social Networks

Jure Leskovec, Daniel P. Huttenlocher+1 more|arXiv (Cornell University)|Mar 11, 2010

Complex Network Analysis TechniquesPhysics and Astronomy21 references141 citations

TL;DR

This paper proposes a machine learning framework to predict the sign (positive or negative) of links in online social networks using structural features from local network neighborhoods. It demonstrates that incorporating negative edge information significantly improves prediction accuracy—especially on Wikipedia, where performance jumps from 0.6983 to 0.7114—while revealing that local structural balance and status theories underlie link formation across diverse platforms like Epinions, Slashdot, and Wikipedia.

ABSTRACT

We study online social networks in which relationships can be either positive (indicating relations such as friendship) or negative (indicating relations such as opposition or antagonism). Such a mix of positive and negative links arise in a variety of online settings; we study datasets from Epinions, Slashdot and Wikipedia. We find that the signs of links in the underlying social networks can be predicted with high accuracy, using models that generalize across this diverse range of sites. These models provide insight into some of the fundamental principles that drive the formation of signed links in networks, shedding light on theories of balance and status from social psychology; they also suggest social computing applications by which the attitude of one user toward another can be estimated from evidence provided by their relationships with other members of the surrounding social network.

Motivation & Objective

• To develop a predictive model for the sign of links in online social networks where relationships can be positive (e.g., friendship) or negative (e.g., distrust).
• To investigate whether structural patterns in signed networks—such as balance and status—generalize across diverse online platforms.
• To evaluate whether incorporating negative edge information enhances prediction performance for positive edge existence, even in tasks focused solely on positive links.
• To compare learned models against established social psychology theories of structural balance and status to validate their explanatory power.
• To enable social computing applications that infer unobserved user attitudes from observed relationships in the network vicinity.

Proposed method

• Formulates the edge sign prediction problem as a binary classification task: predict whether a link between two nodes is positive or negative, given all other signed links.
• Computes 16 distinct signed directed triad features to capture local signed neighborhood patterns around each node pair.
• Uses logistic regression models trained on both positive-only and positive-plus-negative edge features to compare predictive performance.
• Introduces a feature set based on 4-path patterns of positive edges (e.g., FFpp, BBpp) and combines them with signed triad features to model structural complexity.
• Employs a cross-validation framework to evaluate model accuracy on real datasets from Epinions, Slashdot, and Wikipedia.
• Compares model performance against random guessing (0.50 accuracy) and evaluates improvement via AUC-like metrics across all datasets.

Experimental results

Research questions

• RQ1 Can the sign of a link in a signed social network be predicted with high accuracy using only local network structure?
• RQ2 Do principles from social psychology—specifically structural balance and status theory—explain the observed patterns of positive and negative links?
• RQ3 Does incorporating information about negative edges improve the prediction of positive edge existence in social networks?
• RQ4 Is there a global status ordering in signed networks, or do they instead form opposing factions as predicted by balance theory?
• RQ5 How do predictive models based on signed network features compare to those relying only on positive edges in terms of performance?

Key findings

• Incorporating negative edge information improves sign prediction accuracy by 3 percentage points on Epinions (0.5911 vs. 0.5612) and Slashdot (0.5953 vs. 0.5579), with a 50% relative improvement in boost over random guessing.
• On Wikipedia, the model using both positive and negative features achieves 0.7114 accuracy, a significant improvement from 0.6983 when using only positive features.
• The results show that negative relationships are not isolated but interdependent with positive ones, and thus should be modeled jointly for better performance.
• There is strong evidence for a global status ordering in all three datasets, suggesting that users implicitly rank others by perceived status, which aligns with status theory.
• In contrast, there is little evidence for a global factional structure (i.e., two opposing groups), indicating that structural balance operates more strongly at a local level than globally.
• The study confirms that models trained on signed network features significantly outperform previous approaches, especially when negative edge signals are included.