Limitations of exemplar models of multi-attribute probabilistic inference

Area of Science:

• Cognitive Psychology
• Decision Making
• Machine Learning Models

Background:

• Categorization and prediction tasks often involve learning complex relationships between object attributes and criterion variables.
• Exemplar models and heuristic models (e.g., take-the-best, weighted-additive) offer different explanations for human decision-making in such tasks.

Purpose of the Study:

• To contrast the predictive accuracy of exemplar models of categorization against classic models.
• To investigate human sensitivity to attribute interactions in prediction tasks under varying training conditions.
• To analyze response times to infer underlying decision-making strategies.

Main Methods:

• Participants learned to predict a criterion variable from pairs of objects with binary attributes.
• The study manipulated attribute interactions influencing the criterion variable.
• Response times were analyzed alongside prediction accuracy under typical and extended training scenarios.

Main Results:

• Under typical training, participants showed limited sensitivity to attribute interactions, challenging exemplar model predictions.
• With extended training, participants eventually learned these interactions.
• Response time analysis suggested participants recoded attributes into configural cues and applied hierarchical rules, further challenging exemplar models.

Conclusions:

• Human observers initially struggle with complex attribute interactions in prediction tasks, contrary to exemplar model expectations.
• Extended training facilitates learning, but decision strategies appear to shift towards emergent configural cues and rule-based processing.
• Findings suggest a dynamic adaptation in decision-making strategies, favoring heuristic-like processing over pure exemplar-based reasoning.