Reverse engineering a social agent-based hidden markov model--visage

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Network Science

Background:

• Understanding social group evolution is crucial for analyzing community dynamics.
• Agent-based models and hidden Markov models (HMMs) are used to represent complex systems.
• Learning system dynamics from observational data without full state information is a significant challenge.

Purpose of the Study:

• To develop a machine learning approach for discovering agent dynamics that drive social group evolution.
• To identify micro-laws governing agent actions within a community based on communication data.
• To determine group structure, evolution, and micro-laws without knowledge of semantic content or state transitions.

Main Methods:

• An agent-based hidden Markov model (HMM) was introduced to represent agent dynamics.
• The problem was framed as a mixed optimization problem for model identification.
• A multistage learning process was developed to learn group structure, evolution, and micro-laws from communication data.

Main Results:

• The approach successfully identified agent dynamics and micro-laws from observed communications.
• Experiments on synthetic and real-world data (Enron emails, Movie newsgroups) demonstrated the method's feasibility.
• The model accurately approximated group structure, evolution, and underlying agent behaviors.

Conclusions:

• Machine learning can effectively infer agent dynamics and micro-laws driving social evolution from communication data.
• This method provides insights into the driving forces behind social evolution in communities.
• The approach offers a novel way to analyze social networks and group behaviors without requiring semantic understanding.