Introduction
Modern social media users typically maintain accounts across multiple platforms, creating a fragmented digital identity. Predicting new connections (social links) and identifying the same user across different services (anchor links) are critical tasks for recommendation systems and cross-platform analysis. Traditionally, these tasks have been treated in isolation. This paper introduces Collective Link Fusion (CLF), a framework that addresses these challenges simultaneously by leveraging the inherent correlations between social and anchor links.
The Problem: Labeling and Isolation
Existing link prediction methods face two primary hurdles. First is the Labeling Dilemma: in social networks, we observe existing links (positive examples), but the absence of a link does not necessarily imply a negative relationship—it might simply be a connection that hasn't formed yet. Standard classifiers often mislabel these "unlabeled" pairs as negatives.
Second is the issue of Isolated Predictions. Social link prediction and anchor link prediction are deeply intertwined. Knowing that two users are friends on Twitter increases the likelihood they are the same pair of users on Foursquare if their structural patterns match. Conversely, identifying an anchor user provides a bridge to transfer social connection data between networks.
The Collective Link Fusion (CLF) Framework
The CLF framework operates in two phases to overcome these hurdles.
Phase 1: PU Link Formation Inference
To solve the labeling dilemma, the authors employ Positive and Unlabeled (PU) Learning. They introduce a "bridging probability" to estimate the likelihood of a link forming ($y=1$) given its current state ($z$). This approach allows the model to extract signals from currently unconnected pairs without assuming they are permanent negatives, effectively turning the "unlabeled" sea into a source of probabilistic information.
Phase 2: Information Propagation
The core of the framework is the construction of Aligned Probabilistic Networks. Edges in these networks are weighted by the formation probabilities from Phase 1. CLF then uses a Collective Random Walk to propagate information across the partially aligned networks.
This propagation happens in two ways:
- Strict Co-existence: Direct transfer of link information between known anchor users.
- Loose Co-existence: Information flows from a source network (e.g., Twitter) to a target network (e.g., Foursquare) even for non-anchor users, provided they are connected to anchor users. This "pulls" relevant structural data across the bridge.
Experimental Results
The framework was evaluated using real-world datasets from Twitter (5,223 users) and Foursquare (5,392 users), with 3,388 verified anchor links connecting them.
CLF demonstrated significant performance gains over traditional baselines, including Multi-network Link Prediction (MLP) and standard Random Walk with Restart (RWR). The model proved particularly robust in "cold-start" scenarios, maintaining high accuracy even when data in the target network was sparse.
As shown in the ROC analysis, CLF consistently achieves a higher True Positive Rate for any given False Positive Rate compared to other methods. In anchor link prediction specifically, CLF’s AUC was often 40% to 50% higher than simpler multi-network models.
Conclusion and Future Work
The CLF framework demonstrates that anchor and social link predictions are mutually enhancing tasks. By treating the problem as a collective inference challenge and utilizing PU learning, CLF provides a powerful tool for cross-platform user modeling. Future research will focus on scaling these collective random walks to massive networks with millions of nodes and extending the framework to handle more than two networks simultaneously.
Summary updated