Multi-scale compositionality: identifying the compositional structures of social dynamics using deep learning.

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Network Science

Background:

• Social media platforms exhibit complex temporal dynamics across multiple scales.
• Understanding these dynamics is crucial for applications like anomaly detection and forecasting.
• Existing methods struggle to capture the heterogeneity and scale-specific signatures of social interactions.

Purpose of the Study:

• To develop a novel model for analyzing social media temporal dynamics.
• To identify signatures of social dynamics at different time scales and their interactions.
• To provide a framework for understanding and engineering social media behavior.

Main Methods:

• Proposing the Recursive Convolutional Bayesian Model (RCBM), a deep-learning framework.
• Utilizing specialized convolution operators to exploit the heterogeneity of social dynamics.
• Ensuring runtime and convergence through formal analyses.

Main Results:

• The RCBM outperforms state-of-the-art methods in solution quality and computational efficiency.
• Identified compositional structures accurately characterize social dynamics on Twitter and Yelp datasets.
• Demonstrated the utility of pattern identification for anomaly detection and improved forecasting.

Conclusions:

• The RCBM offers a powerful new approach to understanding complex social media dynamics.
• The identified patterns provide insights into opinion spreading and online content promotion.
• This work advances the field of computational social science with practical applications.