Discovering social groups via latent structure learning

Area of Science:

• Cognitive Science
• Social Psychology
• Computational Modeling

Background:

• Human societies universally feature social coalitions, but the mechanisms for learning and representing group boundaries remain poorly understood.
• Existing models often rely on explicit category labels or pairwise similarity, potentially overlooking more complex relational learning processes.

Purpose of the Study:

• To investigate how individuals infer social group structures beyond explicit labels and similarity.
• To examine the role of relational information in forming social group representations.
• To test whether inferred latent groups influence behavior even when contradicted by explicit group labels.

Main Methods:

• Developed a computational model of latent structure learning to predict social group inference.
• Conducted four experiments involving consequential social learning (e.g., political values) in controlled environments.
• Manipulated the presence and salience of explicit group labels versus inferred latent structures.

Main Results:

• Evidence suggests people integrate relational information (how others relate to each other) to infer group boundaries.
• Inferred latent social groups significantly influenced individual choices and attributions about other agents.
• The influence of latent groups persisted even when explicit labels contradicted the inferred structure.

Conclusions:

• Social group representations are dynamically formed through learning latent relational structures.
• This mechanism provides a unified account for social learning and group formation, potentially explaining phenomena like balance theory.