Building Interpretable Machine Learning Models

Transcript

1. Building Interpretable Machine Learning Models Allen Akinkunle allenkunle.me @allenakinkunle 1
2. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 2
3. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 3
4. Introduction ๏ Machine learning algorithms build mathematical models using training data, in order to make predictions or inference. ๏ A model is a mapping ๏ A internal workings of a black-box model is opaque e.g. neural networks ๏ The internal workings of a white-box model is available for inspection e.g. linear regression, logistic regression f : X → Y 4
5. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 5
6. Interpretability ”Interpretability is the degree to which a human can understand the cause of a decision” Miller, Tim. “Explanation in artificial intelligence: Insights from the social sciences.” arXiv Preprint arXiv:1706.07269. (2017) ”Interpretability is the degree to which a human can consistently predict the model’s result” Kim, Been, Rajiv Khanna, and Oluwasanmi O. Koyejo. “Examples are not enough, learn to criticize! Criticism for interpretability.” Advances in Neural Information Processing Systems (2016) 6
7. More interpretable Less interpretable Less complex More complex Model Complexity Model Interpretability 7 Less performant More performant Prediction Performance
8. Why do we care about interpretability? ๏ Ethical and legal reasons ๏ Social acceptance and trust ๏ Auditing and debugging 8
9. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 9
10. Intrinsically Interpretable Models ๏ The easiest way to achieve interpretability is to use machine learning algorithms that produce interpretable models. ๏ Linear regression, Logistic regression, Decision trees are examples. 10
11. 11 Linear regression Logistic regression y = β0 + β1 x1 + … + βp xp + ϵ log ( p(X) 1 − p(X)) = β0 + β1 X1 + ⋯ + βp Xp p(X) = 1 1 + e(−β0 +β1 X1 +⋯+βp Xp )
12. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 12
13. Post-hoc Interpretation Methods ๏ Model-agnostic methods applied to analyse models after training ๏ Global surrogates ๏ Local interpretable model-agnostic explanations (LIME) ๏ Shapley Additive Explanations (SHAP) ๏ Individual Conditional Expectation (ICE) ๏ Permutation Feature Importance 13
14. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 14
15. Local Interpretable Model-agnostic Explanations (LIME) ๏ Local surrogate model used to explain the individual predictions of a black-box model. ๏ Local surrogate model should have local fidelity. 15 Source: https://github.com/marcotcr/lime
16. Local Interpretable Model-agnostic Explanations (LIME) ๏ Mathematically, where is the model being explained is a interpretable model measures how unfaithful is in approximating in the locality defined by measures the complexity ๏ There is a fidelity-interpretability trade-off ξ(x) = arg min g∈G ℒ(f, g, πx ) + Ω(x) f g ∈ G ℒ(f, g, πx ) g f πx Ω(x) 16
17. Caveats ๏ The neighbourhood is really hard to define. Currently uses exponential smoothing kernel. ๏ Sampling could be improved. ๏ The complexity of the explanation model has to be defined in advance. 17
18. Talk Outline ๏ Introduction ๏ Interpretability ๏ Intrinsically Interpretable Models ๏ Post-hoc Interpretation Methods ๏ Local Interpretable Model-Agnostic Explanations (LIME) ๏ Demo 18