Book Image

Interpretable Machine Learning with Python - Second Edition

By : Serg Masís
4 (4)
Book Image

Interpretable Machine Learning with Python - Second Edition

4 (4)
By: Serg Masís

Overview of this book

Interpretable Machine Learning with Python, Second Edition, brings to light the key concepts of interpreting machine learning models by analyzing real-world data, providing you with a wide range of skills and tools to decipher the results of even the most complex models. Build your interpretability toolkit with several use cases, from flight delay prediction to waste classification to COMPAS risk assessment scores. This book is full of useful techniques, introducing them to the right use case. Learn traditional methods, such as feature importance and partial dependence plots to integrated gradients for NLP interpretations and gradient-based attribution methods, such as saliency maps. In addition to the step-by-step code, you’ll get hands-on with tuning models and training data for interpretability by reducing complexity, mitigating bias, placing guardrails, and enhancing reliability. By the end of the book, you’ll be confident in tackling interpretability challenges with black-box models using tabular, language, image, and time series data.

Related Content you might be interested in

Current Title:

Small book image
Interpretable Machine Learning with Python - Second Edition
Serg Masís
Oct, 2023 | 606 pages
Small book image
Applied Machine Learning Explainability Techniques
Aditya Bhattacharya
Jul, 2022 | 306 pages
Small book image
Deep Learning and XAI Techniques for Anomaly Detection
Cher Simon
Jan, 2023 | 218 pages
Small book image
Responsible AI in the Enterprise
Adnan Masood | Heather Dawe
Jul, 2023 | 318 pages
Table of Contents (17 chapters)
Preface
Preface
Who this book is for
What this book covers
To get the most out of this book
Get in touch
1
Interpretation, Interpretability, and Explainability; and Why Does It All Matter?
Interpretation, Interpretability, and Explainability; and Why Does It All Matter?
Technical requirements
What is machine learning interpretation?
Understanding the difference between interpretability and explainability
A business case for interpretability
Summary
Image sources
Dataset sources
Further reading
Free Chapter
2
Key Concepts of Interpretability
Key Concepts of Interpretability
Technical requirements
The mission
The approach
Preparations
Interpretation method types and scopes
Appreciating what hinders machine learning interpretability
Mission accomplished
Summary
Further reading
3
Interpretation Challenges
Interpretation Challenges
Technical requirements
The mission
The approach
The preparations
Loading the libraries
Reviewing traditional model interpretation methods
Understanding limitations of traditional model interpretation methods
Studying intrinsically interpretable (white-box) models
Recognizing the trade-off between performance and interpretability
Discovering newer interpretable (glass-box) models
Mission accomplished
Summary
Dataset sources
Further reading
4
Global Model-Agnostic Interpretation Methods
Global Model-Agnostic Interpretation Methods
Technical requirements
The mission
The approach
The preparations
Model training and evaluation
What is feature importance?
Assessing feature importance with model-agnostic methods
Visualize global explanations
Feature summary explanations
Feature interactions
Summary
Further reading
5
Local Model-Agnostic Interpretation Methods
Local Model-Agnostic Interpretation Methods
Technical requirements
The mission
The approach
The preparations
Leveraging SHAP’s KernelExplainer for local interpretations with SHAP values
Employing LIME
Using LIME for NLP
Trying SHAP for NLP
Comparing SHAP with LIME
Mission accomplished
Summary
Dataset sources
Further reading
6
Anchors and Counterfactual Explanations
Anchors and Counterfactual Explanations
Technical requirements
The mission
The approach
The preparations
Understanding anchor explanations
Exploring counterfactual explanations
Mission accomplished
Summary
Dataset sources
Further reading
7
Visualizing Convolutional Neural Networks
Visualizing Convolutional Neural Networks
Technical requirements
The mission
The approach
Preparations
Visualizing the learning process with activation-based methods
Evaluating misclassifications with gradient-based attribution methods
Understanding classifications with perturbation-based attribution methods
Mission accomplished
Summary
Further reading
8
Interpreting NLP Transformers
Interpreting NLP Transformers
Technical requirements
The mission
The approach
The preparations
Visualizing attention with BertViz
Interpreting token attributions with integrated gradients
LIME, counterfactuals, and other possibilities with the LIT
Mission accomplished
Summary
Further reading
9
Interpretation Methods for Multivariate Forecasting and Sensitivity Analysis
Interpretation Methods for Multivariate Forecasting and Sensitivity Analysis
Technical requirements
The mission
The approach
The preparation
Assessing time series models with traditional interpretation methods
Generating LSTM attributions with integrated gradients
Computing global and local attributions with SHAP’s KernelExplainer
Identifying influential features with factor prioritization
Quantifying uncertainty and cost sensitivity with factor fixing
Mission accomplished
Summary
Dataset and image sources
Further reading
10
Feature Selection and Engineering for Interpretability
Feature Selection and Engineering for Interpretability
Technical requirements
The mission
The approach
The preparations
Understanding the effect of irrelevant features
Reviewing filter-based feature selection methods
Exploring embedded feature selection methods
Discovering wrapper, hybrid, and advanced feature selection methods
Considering feature engineering
Mission accomplished
Summary
Dataset sources
Further reading
11
Bias Mitigation and Causal Inference Methods
Bias Mitigation and Causal Inference Methods
Technical requirements
The mission
The approach
The preparations
Detecting bias
Mitigating bias
Creating a causal model
Understanding heterogeneous treatment effects
Testing estimate robustness
Mission accomplished
Summary
Dataset sources
Further reading
12
Monotonic Constraints and Model Tuning for Interpretability
Monotonic Constraints and Model Tuning for Interpretability
Technical requirements
The mission
The approach
The preparations
Placing guardrails with feature engineering
Tuning models for interpretability
Implementing model constraints
Mission accomplished
Summary
Dataset sources
Further reading
13
Adversarial Robustness
Adversarial Robustness
Technical requirements
The mission
The approach
The preparations
Learning about evasion attacks
Defending against targeted attacks with preprocessing
Shielding against any evasion attack by adversarial training of a robust classifier
Evaluating adversarial robustness
Mission accomplished
Summary
Dataset sources
Further reading
14
What’s Next for Machine Learning Interpretability?
What’s Next for Machine Learning Interpretability?
Understanding the current landscape of ML interpretability
Speculating on the future of ML interpretability
Summary
Further reading
15
Other Books You May Enjoy
Other Books You May Enjoy
16
Index
Index

Defending against targeted attacks with preprocessing

There are five broad categories of adversarial defenses:

  • Preprocessing: changing the model’s inputs so that they are harder to attack.
  • Training: training a new robust model that is designed to overcome attacks.
  • Detection: detecting attacks. For instance, you can train a model to detect adversarial examples.
  • Transformer: modifying model architecture and training so that it’s more robust – this may include techniques such as distillation, input filters, neuron pruning, and unlearning.
  • Postprocessing: changing model outputs to overcome production inference or model extraction attacks.

Only the first four defenses work with evasion attacks, and in this chapter, we will only cover the first two: preprocessing and adversarial training. FGSM and C&W can be defended easily with either of these, but an AP is tougher to defend against, so it might require a stronger detection...

Previous Section
End of Section 7
Next Section