CRL: Collaborative Representation Learning by Coordinating Topic Modeling and Network Embeddings.

Area of Science:

• Computer Science
• Data Science
• Network Analysis

Background:

• Network Representation Learning (NRL) excels at tasks like vertex classification and link prediction.
• Social networks often contain rich textual data, usable for representation learning.
• Existing NRL methods often overlook global text structures, focusing on local network neighborhoods.

Purpose of the Study:

• To propose a unified model, Collaborative Representation Learning (CRL), integrating global text and local network information.
• To collaboratively model topics and learn network embeddings.
• To enhance topic discovery and network representation learning.

Main Methods:

• Developed a matrix factorization (MF) based unified model (CRL).
• Incorporated Fletcher-Reeves (FR) MF optimization (AFR) for efficient parameter learning.
• Evaluated CRL on topic coherence and vertex classification using real-world datasets.

Main Results:

• CRL improves topic discovery performance compared to baseline topic models.
• CRL learns superior network representations over state-of-the-art context-aware NRL models.
• The AFR optimization method achieves convergence in few iterations.

Conclusions:

• CRL effectively leverages complementary global and local information for network analysis.
• The proposed model offers enhanced performance in both topic modeling and network embedding tasks.
• This approach advances the field of network representation learning by integrating diverse data modalities.