Constrained Active Learning for Anchor Link Prediction Across Multiple Heterogeneous Social Networks.

Area of Science:

• Social Network Analysis
• Machine Learning
• Data Mining

Background:

• Users often maintain multiple social media accounts across diverse platforms.
• Identifying anchor links between these accounts is essential for cross-network applications like recommendation and data transfer.
• Supervised models for anchor link prediction require abundant labeled data, which is often unavailable and costly to obtain.

Purpose of the Study:

• To address the challenge of limited labeled data in anchor link prediction.
• To introduce and evaluate active learning strategies tailored for the unique constraints of anchor link discovery.
• To develop novel methods that leverage unlabeled data effectively for improved cross-network user identification.

Main Methods:

• Formulating the active learning based anchor link prediction problem with a one-to-one constraint.
• Defining new information gain measures specific to the constraints of anchor link prediction.
• Developing and implementing several constrained active learning methods for anchor link prediction.
• Conducting extensive experiments on real-world social network datasets.

Main Results:

• The proposed Mean-entropy-based Constrained Active Learning (MC) method demonstrated superior performance.
• The MC method significantly outperformed state-of-the-art anchor link prediction techniques.
• The active learning approach effectively utilized unlabeled data, overcoming limitations of traditional supervised methods.

Conclusions:

• Constrained active learning is a promising approach for anchor link prediction in scenarios with limited labeled data.
• The developed information gain measures and MC method provide effective solutions for cross-network user identification.
• This research offers a more efficient and practical way to discover user connections across heterogeneous social networks.