Explanatory Model Analysis. Explore, Explain and Examine Predictive Models

✍ Scribed by Przemyslaw Biecek, Tomasz Burzykowski

• Publisher: CRC Press
• Year: 2021
• Tongue: English
• Leaves: 328
• Series: CHAPMAN & HALL/CRC DATA SCIENCE SERIES
• Category: Library

✦ Table of Contents

Cover
Half Title
Title Page
Copyright Page
Dedication
Table of Contents
Part I Introduction
1 Introduction
1.1 The aim of the book
1.2 A bit of philosophy: three laws of model explanation
1.3 Terminology
1.4 Black-box models and glass-box models
1.5 Model-agnostic and model-specific approach
1.6 The structure of the book
1.7 What is included in this book and what is not
1.8 Acknowledgements
2 Model Development
2.1 Introduction
2.2 Model-development process
2.3 Notation
2.4 Data understanding
2.5 Model assembly (fitting)
2.6 Model audit
3 Do-it-yourself
3.1 Do-it-yourself with R
3.1.1 What to install?
3.1.2 How to work with DALEX?
3.1.3 How to work with archivist?
3.2 Do-it-yourself with Python
3.2.1 What to install?
3.2.2 How to work with dalex?
3.2.3 Code snippets for Python
4 Datasets and Models
4.1 Sinking of the RMS Titanic
4.1.1 Data exploration
4.2 Models for RMS Titanic, snippets for R
4.2.1 Logistic regression model
4.2.2 Random forest model
4.2.3 Gradient boosting model
4.2.4 Support vector machine model
4.2.5 Models’ predictions
4.2.6 Models’ explainers
4.2.7 List of model-objects
4.3 Models for RMS Titanic, snippets for Python
4.3.1 Logistic regression model
4.3.2 Random forest model
4.3.3 Gradient boosting model
4.3.4 Support vector machine model
4.3.5 Models’ predictions
4.3.6 Models’ explainers
4.4 Apartment prices
4.4.1 Data exploration
4.5 Models for apartment prices, snippets for R
4.5.1 Linear regression model
4.5.2 Random forest model
4.5.3 Support vector machine model
4.5.4 Models’ predictions
4.5.5 Models’ explainers
4.5.6 List of model-objects
4.6 Models for apartment prices, snippets for Python
4.6.1 Linear regression model
4.6.2 Random forest model
4.6.3 Support vector machine model
4.6.4 Models’ predictions
4.6.5 Models’ explainers
Part II Instance Level
5 Introduction to Instance-level Exploration
6 Break-down Plots for Additive Attributions
6.1 Introduction
6.2 Intuition
6.3 Method
6.3.1 Break-down for linear models
6.3.2 Break-down for a general case
6.4 Example: Titanic data
6.5 Pros and cons
6.6 Code snippets for R
6.6.1 Basic use of the predict_parts() function
6.6.2 Advanced use of the predict_parts() function
6.7 Code snippets for Python
7 Break-down Plots for Interactions
7.1 Intuition
7.2 Method
7.3 Example: Titanic data
7.4 Pros and cons
7.5 Code snippets for R
7.6 Code snippets for Python
8 Shapley Additive Explanations (SHAP) for Average Attributions
8.1 Intuition
8.2 Method
8.3 Example: Titanic data
8.4 Pros and cons
8.5 Code snippets for R
8.6 Code snippets for Python
9 Local Interpretable Model-agnostic Explanations (LIME)
9.1 Introduction
9.2 Intuition
9.3 Method
9.3.1 Interpretable data representation
9.3.2 Sampling around the instance of interest
9.3.3 Fitting the glass-box model
9.4 Example: Titanic data
9.5 Pros and cons
9.6 Code snippets for R
9.6.1 The lime package
9.6.2 The localModel package
9.6.3 The iml package
9.7 Code snippets for Python
10 Ceteris-paribus Profiles
10.1 Introduction
10.2 Intuition
10.3 Method
10.4 Example: Titanic data
10.5 Pros and cons
10.6 Code snippets for R
10.6.1 Basic use of the predict_profile() function
10.6.2 Advanced use of the predict_profile() function
10.6.3 Comparison of models (champion-challenger)
10.7 Code snippets for Python
11 Ceteris-paribus Oscillations
11.1 Introduction
11.2 Intuition
11.3 Method
11.4 Example: Titanic data
11.5 Pros and cons
11.6 Code snippets for R
11.6.1 Basic use of the predict_parts() function
11.6.2 Advanced use of the predict_parts() function
11.7 Code snippets for Python
12 Local-diagnostics Plots
12.1 Introduction
12.2 Intuition
12.3 Method
12.3.1 Nearest neighbors
12.3.2 Local-fidelity plot
12.3.3 Local-stability plot
12.4 Example: Titanic
12.5 Pros and cons
12.6 Code snippets for R
12.7 Code snippets for Python
13 Summary of Instance-level Exploration
13.1 Introduction
13.2 Number of explanatory variables in the model
13.2.1 Low to medium number of explanatory variables
13.2.2 Medium to a large number of explanatory variables
13.2.3 Very large number of explanatory variables
13.3 Correlated explanatory variables
13.4 Models with interactions
13.5 Sparse explanations
13.6 Additional uses of model exploration and explanation
13.7 Comparison of models (champion-challenger analysis)
Part III Dataset Level
14 Introduction to Dataset-level Exploration
15 Model-performance Measures
15.1 Introduction
15.2 Intuition
15.3 Method
15.3.1 Continuous dependent variable
15.3.1.1 Goodness-of-fit
15.3.1.2 Goodness-of-prediction
15.3.2 Binary dependent variable
15.3.2.1 Goodness-of-fit
15.3.2.2 Goodness-of-prediction
15.3.3 Categorical dependent variable
15.3.3.1 Goodness-of-fit
15.3.3.2 Goodness-of-prediction
15.3.4 Count dependent variable
15.4 Example
15.4.1 Apartment prices
15.4.2 Titanic data
15.5 Pros and cons
15.6 Code snippets for R
15.7 Code snippets for Python
16 Variable-importance Measures
16.1 Introduction
16.2 Intuition
16.3 Method
16.4 Example: Titanic data
16.5 Pros and cons
16.6 Code snippets for R
16.7 Code snippets for Python
17 Partial-dependence Profiles
17.1 Introduction
17.2 Intuition
17.3 Method
17.3.1 Partial-dependence profiles
17.3.2 Clustered partial-dependence profiles
17.3.3 Grouped partial-dependence profiles
17.3.4 Contrastive partial-dependence profiles
17.4 Example: apartment-prices data
17.4.1 Partial-dependence profiles
17.4.2 Clustered partial-dependence profiles
17.4.3 Grouped partial-dependence profiles
17.4.4 Contrastive partial-dependence profiles
17.5 Pros and cons
17.6 Code snippets for R
17.6.1 Partial-dependence profiles
17.6.2 Clustered partial-dependence profiles
17.6.3 Grouped partial-dependence profiles
17.6.4 Contrastive partial-dependence profiles
17.7 Code snippets for Python
17.7.1 Grouped partial-dependence profiles
17.7.2 Contrastive partial-dependence profiles
18 Local-dependence and Accumulated-local Profiles
18.1 Introduction
18.2 Intuition
18.3 Method
18.3.1 Local-dependence profile
18.3.2 Accumulated-local profile
18.3.3 Dependence profiles for a model with interaction and correlated explanatory variables: an example
18.4 Example: apartment-prices data
18.5 Pros and cons
18.6 Code snippets for R
18.7 Code snippets for Python
19 Residual-diagnostics Plots
19.1 Introduction
19.2 Intuition
19.3 Method
19.4 Example: apartment-prices data
19.5 Pros and cons
19.6 Code snippets for R
19.7 Code snippets for Python
20 Summary of Dataset-level Exploration
20.1 Introduction
20.2 Exploration on training/testing data
20.3 Correlated explanatory variables
20.4 Comparison of models (champion-challenger analysis)
Part IV Use-cases
21 FIFA 19
21.1 Introduction
21.2 Data preparation
21.2.1 Code snippets for R
21.2.2 Code snippets for Python
21.3 Data understanding
21.4 Model assembly
21.4.1 Code snippets for R
21.4.2 Code snippets for Python
21.5 Model audit
21.5.1 Code snippets for R
21.5.2 Code snippets for Python
21.6 Model understanding (dataset-level explanations)
21.6.1 Code snippets for R
21.6.2 Code snippets for Python
21.7 Instance-level explanations
21.7.1 Robert Lewandowski
21.7.2 Code snippets for R
21.7.3 Code snippets for Python
21.7.4 CR7
21.7.5 Wojciech Szczęsny
21.7.6 Lionel Messi
22 Reproducibility
22.1 Package versions for R
22.2 Package versions for Python
Bibliography
Index