Sinkhorn Distance Minimization for Adaptive Semi-Supervised Social Network Alignment.

Area of Science:

• Social Network Analysis
• Graph Mining
• Machine Learning

Background:

• Social network alignment links identical users across platforms, crucial for social graph mining.
• Existing supervised methods demand extensive manual labels, impractical for large-scale, cross-platform data.
• Isomorphism and adversarial learning have been used but struggle with unpredictable user behavior and training instability.

Purpose of the Study:

• To propose Meta-SNA, a meta-learning-based model for robust social network alignment.
• To effectively capture both isomorphism and unique identity characteristics across social platforms.
• To address limitations of adversarial learning and improve cross-platform identity linking.

Main Methods:

• Developed Meta-SNA, a meta-learning framework with a shared meta-model and identity-specific adaptors.
• Utilized Sinkhorn distance for distribution closeness measurement, offering an optimal and efficient solution.
• Employed a novel approach to preserve global cross-platform knowledge while learning specific projection functions.

Main Results:

• Meta-SNA effectively captures both isomorphism and unique identity traits.
• The model demonstrates superior performance in social network alignment tasks.
• Sinkhorn distance implementation improved upon limitations of adversarial learning.

Conclusions:

• Meta-SNA offers a superior approach to social network alignment compared to existing methods.
• The meta-learning framework provides a robust solution for cross-platform identity linking.
• The proposed method effectively handles the complexities of social network data.