LIME

Preface

A fitted model can predict accurately while giving little indication of how it reached one particular prediction. This is common with random forests, gradient boosted trees, support vector machines, and neural networks. It matters in biological work because a prediction about one tumour, patient, or cell is often followed immediately by a second question: which measurements made the model behave this way for this sample?

LIME, short for Local Interpretable Model Agnostic Explanations, answers that question by studying the fitted model close to one observation. It creates artificial observations, obtains predictions from the original model, and fits a simpler model that approximates those predictions nearby. The result is an explanation of the fitted model’s local behaviour. It is not a model of the biological mechanism and it is not automatically a statement about the population as a whole.

Start with Understanding LIME.