[Research Insight] Predicting User Roles across Social Networks via Feature Transformation
Predicting User Roles in Social Networks Using Transfer Learning with Feature Transformation
This paper proposes a transductive transfer learning framework for predicting user roles (e.g., Administrator, Bot, Moderator) in unlabeled social networks. The core method utilizes feature transformation techniques to map local structural network features into a global, dataset-independent feature space, allowing models trained on one network to be applied effectively to another.
TL;DR
How do you identify the "Admins" or "Bots" in a brand-new social network where no one is labeled? This paper introduces a transductive transfer learning approach that extracts node-level structural features (like how many people you talk to) and transforms them into a universal feature space. By normalizing Power-law distributions and PageRank values, the authors show we can train a classifier on Wikipedia and successfully find influencers in an enterprise forum.
The "Broken" Assumption of Network Analysis
In standard machine learning, we assume our training data and test data come from the same "distribution." In social networks, this assumption is almost always false. A "high degree" user in a small 500-person community might only have 10 connections, while a "high degree" user in Wikipedia might have 10,000.
If you train a model on Wikipedia to find administrators, it will look for massive degree counts. Apply that model to a smaller corporate forum, and it will find... no one. The structural features are "domain-specific" because they depend on the size and density of the network.
Methodology: Bridging the Gap with Power-Law Transformations
The authors propose a robust pipeline to make features comparable across networks, regardless of scale.
1. Feature Extraction & Aggregation
Instead of looking at user names or profiles, the model looks purely at the topology:
- Base Features: Degree, Indegree, Outdegree, Local Clustering Coefficient, and PageRank.
- Recursive Aggregation: A node is defined by its neighbors. The model uses "Feature Aggregation" (Algorithm 1) to iteratively include the average features of a node's neighbors, capturing the "local flavor" of the network hierarchy.
2. The Core Innovation: Feature Transformation
To solve the "Apple vs. Oranges" comparison problem, they apply specific mathematical transformations:
- Power-law Degree Transformation: Since social network degrees follow a distribution, they map any power-law distribution to a standard one (). This aligns the "tails" of different networks.
- Normalized PageRank: Standard PageRank decreases as the network grows. They use a normalized version that is independent of network size.
Fig 1. The Transfer Learning Pipeline: Moving from Domain-Specific Networks to a Common Feature Space.
Experiments & Real-World Validation
The authors tested their approach on two major datasets:
- Wiki-talk: 28 languages of Wikipedia talk pages.
- Boards.ie to Software AG: Transferring knowledge from an Irish public forum to a private corporate community.
Key Findings:
- Power-law Transformation works: As shown in the figure below, the degree distributions of different networks (which are originally separate) perfectly overlap after transformation, allowing the classifier to "understand" them in the same way.
- Superior Performance: The "All" transformation method consistently achieved higher ROC-AUC compared to "None" or simple "Quantile" transformations.
Fig 2. (a) Distributions before transformation (separated) vs. (b) Distributions after transformation (aligned).
Critical Insight & Conclusion
The brilliance of this paper lies in its simplicity and physical intuition. It recognizes that while social networks vary in size, the "physics" of human interaction (modeled by Power-laws and Centrality) remains consistent.
Takeaway for Practitioners: If you are building tools to moderate communities or identify influencers, don't just rely on raw numbers. By applying structural transformations, you can leverage "Global Knowledge" from massive datasets like Wikipedia to solve "Local Problems" in smaller, unlabelled networks.
Limitations: The method relies on the assumption that the target network follows a Power-law. While common, very niche or highly regulated networks might deviate, requiring different transformation priors.
Main Reference: Jun Sun, Jérôme Kunegis, and Steffen Staab. "Predicting User Roles in Social Networks using Transfer Learning with Feature Transformation."