Interpretability

Interpretability concerns how clearly a model's predictions, structure or dependence on inputs can be understood.

Different questions

Global interpretation asks how the model behaves across a population. Local interpretation asks why a particular prediction changes with particular inputs.

Tools

Coefficients in simple models, partial dependence, permutation importance and local attribution methods can provide different forms of explanation. Their meaning depends on assumptions and correlations among features.

Prediction is not mechanism

A feature that strongly improves prediction is not necessarily a causal biological driver. It may be a proxy for another variable or reflect the structure of the sampled data.