Multilabel user classification using the community structure of online networks.

Area of Science:

• Computer Science
• Network Analysis
• Machine Learning

Background:

• Online social platforms generate vast amounts of networked data.
• User classification is crucial for understanding user behavior and tailoring services.
• Existing methods often struggle with the complexity and scale of social network data.

Purpose of the Study:

• To propose a novel framework for semi-supervised, multi-label user classification on social networks.
• To introduce an effective algorithm for learning sparse user embeddings from graph structures.
• To enhance classification performance by integrating unsupervised community detection with supervised feature weighting.

Main Methods:

• Developed Approximate Regularized Commute-Time Embedding (ARCTE) for sparse graph embedding.
• Utilized an improved personalized PageRank algorithm for local graph structure analysis.
• Implemented supervised community feature weighting to prioritize predictive communities.

Main Results:

• ARCTE effectively projects users onto a latent space by extracting local communities.
• Supervised weighting significantly boosts the importance of predictive communities.
• Extensive comparative studies show ARCTE outperforms existing graph embedding methods.

Conclusions:

• ARCTE offers a significant advancement in semi-supervised user classification for networked data.
• The method achieves up to 35% relative improvement in F1-score compared to leading competitors.
• ARCTE provides a robust and scalable solution for user classification in online social platforms.