Interpretable machine learning for psychological research: Opportunities and pitfalls.

Area of Science:

• Psychology
• Computer Science
• Data Science

Background:

• Machine learning (ML) is increasingly used for prediction in psychology.
• Psychological research requires understanding predictor relevance, effects, and interactions, not just predictions.
• Many ML methods, such as random forests and neural networks, are considered 'black boxes' due to limited interpretability.

Purpose of the Study:

• To present and illustrate interpretation techniques for black box ML models in psychology.
• To explore the opportunities and potential pitfalls of using these interpretation techniques.
• To demonstrate how interpretable ML can aid in describing complex psychological relationships.

Main Methods:

• Review and presentation of various interpretation techniques for ML models.
• Illustration using two popular black box ML methods: random forests and neural networks.
• Use of simulated didactic examples and an empirical dataset to demonstrate interpretation and objectify visualizations.

Main Results:

• Interpretation techniques can help describe how ML models make predictions.
• Correlated predictors can impact the interpretation of predictor relevance and effect shapes.
• Detection of interaction effects may be challenging and depends on specific data settings.

Conclusions:

• Interpretable machine learning techniques, when critically applied, can be useful for psychological research.
• These methods can aid in understanding complex relationships between variables.
• Researchers must be aware of potential misinterpretations, especially concerning predictor correlations and interactions.