Learning from missing feedback: Exemplar versus model-based methods.

Area of Science:

• Cognitive psychology
• Machine learning
• Decision science

Background:

• Real-world feedback is often selective, leading to biased samples favoring positive outcomes.
• Decision-makers face challenges in learning from such biased data and correcting for inherent biases.

Purpose of the Study:

• To investigate how individuals learn from selectively biased feedback.
• To determine if people can correct for selection bias in learning.
• To compare exemplar and model-based learning approaches in handling missing feedback.

Main Methods:

• Described computational problems of classification learning from biased samples.
• Examined exemplar models (imputing missing negative outcomes) and model-based methods (adjusting task representation).
• Conducted three experiments to test reliance on imputation versus Bayesian models for bias correction.

Main Results:

• Participants employed diverse strategies, with many best described by exemplar models (often with imputation).
• An almost equal proportion of participants were best described by Bayesian models.
• Individual strategies showed some stability across tasks but were influenced by task uncertainty.

Conclusions:

• People adaptively handle missing feedback using varied approaches.
• Learners utilize both exemplar-based imputation and model-based Bayesian correction.
• Strategy choice is influenced by task assumptions and context, demonstrating adaptive learning.