Awesome

Responsible Machine Learning

With Great Power Comes Great Responsibility. Voltaire (well, maybe)

How to develop machine learning models in a responsible manner? There are several topics worth considering:

Effective. Is the model good enough? Models with low performance should not be used because they can do more harm than good. Communicate the performance of the model in a language that the user understands. Remember that the models will work on a different dataset than the training one. Make sure to assess the performance on the target dataset.
Transparent. Does the user know what influences model predictions? Interpretability and explainability is important. If the model decisions affect us directly or indirectly, we should know where these decisions come from and how they can be changed.
Fair. Does the model discriminate on the basis of gender, age, race or other sensitive attribute? Direct or indirect? It should not! Discrimination can come in many faces. The model may give lower scores, may have lower performance, or may be based on different variables for the protected population.
Secure. Do not let your model be hacked. Every complex system has its vulnerabilities. Seek them out and fix them. Some users may use various tricks to pull model predictions onto their site.
Confidential. Models are often built on sensitive data. Make sure that the data does not leak, so that sensitive attributes are not shared with unauthorized persons. Also beware of model leaks.
Reproducible. Usually the model development process consists of many steps. Make sure that they are completely reproducible and thus can be verified one by one.

Collection of tools for Visual Exploration, Explanation and Debugging of Predictive Models

It takes a village to raise a <del>child</del> model.

The way how we do predictive modeling is very ineffective. We spend way too much time on manual time-consuming and easy to automate activities like data cleaning and exploration, crisp modeling, model validation. We should be focusing more on model understanding, productisation and communication.

Here are gathered tools that can be used to make out work more efficient through the whole model lifecycle. The unified grammar beyond DrWhy.AI universe is described in the Explanatory Model Analysis: Explore, Explain and Examine Predictive Models book.

Lifecycle for Predictive Models

The DrWhy is based on an unified Model Development Process inspired by RUP. Find an overview in the diagram below.

The DrWhy.AI family

Packages in the DrWhy.AI family of models may be divided into four classes.

Model adapters. Predictive models created with different tools have different structures, and different interfaces. Model adapters create uniform wrappers. This way other packages may operate on models in an unified way. DALEX is a lightweight package with generic interface. DALEXtra is a package with extensions for heavyweight interfaces like scikitlearn, h2o, mlr.
Model agnostic explainers. These packages implement specific methods for model exploration. They can be applied to a single model or they can compare different models. ingredients implements variable specific techniques like Ceteris Paribus, Partial Dependency, Permutation based Feature Importance. iBreakDown implements techniques for variable attribution, like Break Down or SHAPley values. auditor implements techniques for model validation, residual diagnostic and performance diagnostic.
Model specific explainers. These packages implement model specific techniques. randomForestExplainer implements techniques for exploration of randomForest models. EIX implements techniques for exploration of gbm and xgboost models. cr19 implements techniques for exploration of survival models.
Automated exploration. These packages combine series of model exploration techniques and produce an automated report of website for model exploration. modelStudio implements a dashboard generator for local and global interactive model exploration. modelDown implements a HTML website generator for global model cross comparison.

Here is a more detailed overview.

DALEX <img src="https://modeloriented.github.io/DALEX/reference/figures/logo.png" align="right" width="100"/>

The DALEX package (Descriptive mAchine Learning EXplanations) helps to understand how complex models are working. The main function explain creates a wrapper around a predictive model. Wrapped models may then be explored and compared with a collection of local and global explainers. Recent developments from the area of Interpretable Machine Learning/eXplainable Artificial Intelligence.

DALEX wraps methods from other packages, i.e. 'pdp' (Greenwell 2017) doi:10.32614/RJ-2017-016, 'ALEPlot' (Apley 2018) arXiv:1612.08468, 'factorMerger' (Sitko and Biecek 2017) arXiv:1709.04412, 'breakDown' package (Staniak and Biecek 2018) doi:10.32614/RJ-2018-072, (Fisher at al. 2018) arXiv:1801.01489.

Vignettes:

General introduction: Survival on the RMS Titanic

DALEXtra <img src="https://github.com/ModelOriented/DALEXtra/blob/master/man/figures/logo.png?raw=true" align="right" width="100"/>

The DALEXtra package is an extension pack for DALEX package. This package provides easy to use connectors for models created with scikitlearn, keras, H2O, mljar and mlr.

Vignettes:

General introduction: DALEX with scikitlearn models

survex <img src="https://raw.githubusercontent.com/ModelOriented/survex/main/man/figures/survex.png" align="right" width="100"/>

The survex package provides model-agnostic explanations for machine learning survival models. It is based on the DALEX package.

Due to a functional type of prediction, either in the form of survival function or cumulative hazard function, standard model-agnostic explanations cannot be applied directly to survival analysis machine learning models. The survex package contains implementations of explanation methods specific to survival analysis, as well as extensions of existing ones for classification or regression.

Vignettes:

ingredients <img src="https://modeloriented.github.io/ingredients/reference/figures/logo.png" align="right" width="100"/>

The ingredients package is a collection of tools for assessment of feature importance and feature effects.

Key functions: feature_importance() for assessment of global level feature importance, ceteris_paribus() for calculation of the Ceteris Paribus / What-If Profiles, partial_dependency() for Partial Dependency Plots, conditional_dependency() for Conditional Dependency Plots also called M Plots, accumulated_dependency() for Accumulated Local Effects Plots, cluster_profiles() for aggregation of Ceteris Paribus Profiles, generic print() and plot() for better usability of selected explainers, generic plotD3() for interactive D3 based explanations, and generic describe() for explanations in natural language.

Vignettes:

iBreakDown <img src="https://modeloriented.github.io/iBreakDown/reference/figures/logo.png" align="right" width="100"/>

The iBreakDown package is a model agnostic tool for explanation of predictions from black boxes ML models. Break Down Table shows contributions of every variable to a final prediction. Break Down Plot presents variable contributions in a concise graphical way. SHAP (Shapley Additive Attributions) values are calculated as average from random Break Down profiles. This package works for binary classifiers as well as regression models.

iBreakDown is a successor of the breakDown package. It is faster (complexity O(p) instead of O(p^2)). It supports interactions and interactive explainers with D3.js plots.

Vignettes:

auditor <img src="https://modeloriented.github.io/auditor/reference/figures/logo.png" align="right" width="100"/>

The auditor package is a tool for model-agnostic validation. Implemented techniques facilitate assessing and comparing the goodness of fit and performance of models. In addition, they may be used for the analysis of the similarity of residuals and for the identification of outliers and influential observations. The examination is carried out by diagnostic scores and visual verification. Due to the flexible and consistent grammar, it is simple to validate models of any classes.

Learn more:

fairmodels <img src="https://modeloriented.github.io/fairmodels/reference/figures/logo.png" align="right" width="100"/>

Flexible tool for bias detection, visualization, and mitigation. Use models explained with DALEX and calculate fairness classification metrics based on confusion matrices using fairness_check() or try newly developed module for regression models using fairness_check_regression(). R package fairmodels allows to compare and gain information about various machine learning models. Mitigate bias with various pre-processing and post-processing techniques. Make sure your models are classifying protected groups similarly.

Learn more:

fairmodels website
Compas recidivism data use case: Basic tutorial
Bias mitigation techniques on Adult data: Advanced tutorial

vivo <img src="https://modeloriented.github.io/vivo/reference/figures/logo.png" align="right" width="100"/>

The vivo package helps to calculate instance level variable importance (measure of local sensitivity). The importance measure is based on Ceteris Paribus profiles and can be calculated in eight variants. Select the variant that suits your needs by setting parameters: absolute_deviation, point and density.

Learn more:

Intuition for Ceteris Paribus Oscillations

randomForestExplainer <img src="https://modeloriented.github.io/randomForestExplainer/reference/figures/logo.png" align="right" width="100"/>

The randomForestExplainer package helps to understand what is happening inside a Random Forest model. This package helps to explore main effects and pairwise interactions, depth distribution, conditional responses and feature importance.

Learn more:

xspliner

The xspliner package is a collection of tools for training interpretable surrogate ML models. The package helps to build simple, interpretable models that inherits informations provided by more complicated ones - resulting model may be treated as explanation of provided black box, that was supplied prior to the algorithm. Provided functionality offers graphical and statistical evaluation both for overall model and its components.

shapper <img src="https://github.com/ModelOriented/shapper/blob/master/man/figures/logo.png?raw=true" align="right" width="100"/>

The shapper is an R wrapper of SHAP python library. It accesses python implementation through reticulate connector.

drifter

The drifter is an R package that identifies concept drift in model structure or in data structure.

Machine learning models are often fitted and validated on historical data under an assumption that data are stationary. The most popular techniques for validation (k-fold cross-validation, repeated cross-validation, and so on) test models on data with the same distribution as training data.

Yet, in many practical applications, deployed models are working in a changing environment. After some time, due to changes in the environment, model performance may degenerate, as model may be less reliable.

Concept drift refers to the change in the data distribution or in the relationships between variables over time. Think about model for energy consumption for a school, over time the school may be equipped with larger number of devices of with more power-efficient devices that may affect the model performance.

EIX

The EIX package implements set of techniques to explore and explain XGBoost and LightGBM models. Main functions of this package cover various variable importance measures and well as functions for identification of interactions between variables.

Learn more:

Cheatsheet

modelStudio <img src="https://modeloriented.github.io/modelStudio/reference/figures/logo.png" align="right" width="100"/>

The modelStudio package automates the explanatory analysis of machine learning predictive models. It generates advanced interactive model explanations in the form of a serverless HTML site with only one line of code. This tool is model-agnostic, therefore compatible with most of the black-box predictive models and frameworks (e.g. mlr/mlr3, xgboost, caret, h2o, parsnip, tidymodels, scikit-learn, lightgbm, keras/tensorflow).

The main modelStudio() function computes various (instance and model-level) explanations and produces a customisable dashboard, which consists of multiple panels for plots with their short descriptions. Easily save the dashboard and share it with others. Tools for Explanatory Model Analysis unite with tools for Exploratory Data Analysis to give a broad overview of the model behavior.

Learn more:

arenar

Arena is an interactive tool that allows you to explore and compare any model regardless of its internal structure.

The arenar package can be run in two modes - live (R runs in the background and calculates all necessary explanations) and serverless (all necessary explanations are calculated earlier).

Using the Arena is trivially simple. Examples with different levels of advancement are available:

modelDown

ThemodelDown package generates a website with HTML summaries for predictive models. Is uses DALEX explainers to compute and plot summaries of how given models behave. We can see how well models behave (Model Performance, Auditor), how much each variable contributes to predictions (Variable Response) and which variables are the most important for a given model (Variable Importance). We can also compare Concept Drift for pairs of models (Drifter). Additionally, data available on the website can be easily recreated in current R session (using archivist package).

Learn more:

rSAFE <img src="https://modeloriented.github.io/rSAFE/reference/figures/logo.png" align="right" width="100"/>

The rSAFE package is a model agnostic tool for making an interpretable white-box model more accurate using alternative black-box model called surrogate model. Based on the complicated model, such as neural network or random forest, new features are being extracted and then used in the process of fitting a simpler interpretable model, improving its overall performance.

Learn more:

EloML <img src="https://modeloriented.github.io/EloML/reference/figures/logo.png" align="right" width="100"/>

The EloML package provides Elo rating system for machine learning models. Elo Predictive Power (EPP) score helps to assess model performance based Elo ranking system.

Learn more:

arxiv

archivist

The archivist package automate serialization and deserialization of R objects. Objects are stored with additional metadata to facilitate reproducibility and governance of data science projects.

Everything that exists in R is an object. archivist is an R package that stores copies of all objects along with their metadata. It helps to manage and recreate objects with final or partial results from data analysis. Use archivist to record every result, to share these results with future you or with others, to search through repository of objects created in the past but needed now.

Learn more:

Tools that are useful during the model lifetime. MI2 stands for our internal tools.

1. Data preparation

dataMaid; A Suite of Checks for Identification of Potential Errors in a Data Frame as Part of the Data Screening Process
inspectdf; A collection of utilities for columnwise summary, comparison and visualisation of data frames.
validate; Professional data validation for the R environment
errorlocate; Find and replace erroneous fields in data using validation rules
ggplot2; System for declaratively creating graphics, based on The Grammar of Graphics.

2. Data understanding

Model Agnostic Variable Importance Scores. Surrogate learning = Train an elastic model and measure feature importance in such model. See DALEX, Model Class Reliance MCR
vip Variable importance plots
SAFE Surrogate learning = Train an elastic model and extract feature transformations.
xspliner Using surrogate black-boxes to train interpretable spline based additive models
factorMerger Set of tools for factors merging paper
ingredients Set of tools for model level feature effects and feature importance.

4. Model assembly

mlr Machine Learning in R paper
caret Classification And Regression Training paper

5. Model audit

auditor model verification, validation, and error analysis vigniette
DALEX Descriptive mAchine Learning EXplanations
iml; interpretable machine learning R package
randomForestExplainer A set of tools to understand what is happening inside a Random Forest
survxai Explanations for survival models paper

6. Model delivery

iBreakDown, pyBreakDown Model Agnostic Explainers for Individual Predictions (with interactions)
ceterisParibus, pyCeterisParibus, ceterisParibusD3 and ingredients Ceteris Paribus Plots (What-If plots) for explanations of a single observation
localModel and live LIME-like explanations with interpretable features based on Ceteris Paribus curves.
lime; Local Interpretable Model-Agnostic Explanations (R port of original Python package)
shapper An R wrapper of SHAP python library
modelDown modelDown generates a website with HTML summaries for predictive models
modelStudio modelStudio generates an interactive dashboard with D3 based summaries for predictive models
drifter Concept Drift and Concept Shift Detection for Predictive Models
archivist A set of tools for datasets and plots archiving paper

Family of Model Explainers

images/DrWhyAI.png

Architecture of DrWhy

DrWhy works on fully trained predictive models. Models can be created with any tool.

DrWhy uses DALEX package to wrap model with additional metadata required for explanations, like validation data, predict function etc.

Explainers for predictive models can be created with model agnostic or model specific functions implemented in various packages.

Hype_Cycle