liveBook · Manning

Fidelity-interpretability tradeoff

As explained in Ensemble Methods for Machine Learning

The fidelity-interpretability tradeoff is the tension between accurately approximating a model and keeping the explanatory model simple enough to understand.

Explaining a black-box prediction requires balancing two goals: matching the model accurately and keeping the explanation simple enough to understand. Surrogate models make this tension explicit by approximating black-box behavior while remaining interpretable. section 9.4.1 section 9.6

The tradeoff appears operationally in surrogate-model design:

  • Decision-tree surrogates measure fidelity with the R2 score and interpretability through tree size, where fewer leaf nodes generally improve interpretability. section 9.4.1
  • LIME generates perturbed examples near the instance being explained, weights them by distance, and fits a linear surrogate using weighted mean squared error. section 9.4.1
  • Linear surrogates become less interpretable as they use more features, so selecting a compact representation limits complexity while preserving local fidelity. section 9.4.1

Local explanation for an XGBoost prediction

LIME can explain an XGBoost prediction by constructing a local surrogate dataset, accounting for categorical features, and fitting a linear explanation around one test instance. For test example 3104, negative feature contributions appear to the left of zero and positive contributions to the right.

How Ensemble Methods for Machine Learning covers it

How deep it goes
Defines fidelity and interpretability measures and explains local weighted fitting, but stops short of modeling feature interactions.
Examples
Uses a decision-tree surrogate, perturbed local examples, and an LIME explanation of XGBoost test example 3104.
What it assumes
Assumes a black-box machine-learning model and an individual prediction to explain.
What you can do after
Apply LIME to approximate an individual black-box prediction with a weighted, interpretable linear surrogate.
How it connects
Connects surrogate models with decision trees, linear models, LIME, global interpretability, and local interpretability.
Cover of Ensemble Methods for Machine Learning
liveBookmanning.com
Read Ensemble Methods for Machine Learning on liveBook Get the book at manning.com

This explainer was generated by Manning's knowledge graph from the text of Ensemble Methods for Machine Learning. The section links point to the parts of the book it draws on.