Interpretable Counterfactual Explanations Guided by Prototypes

By Arnaud Van Looveren, Janis Klaise. 2019

474 citations recorded by the source corpus.

We propose a fast, model agnostic method for finding interpretable counterfactual explanations of classifier predictions by using class prototypes. We show that class prototypes, obtained using either an encoder or through class specific k-d trees, significantly speed up the the search for counterfactual instances and result in more interpretable explanations. We introduce two novel metrics to quantitatively evaluate local interpretability at the instance level. We use these metrics to illustrate the effectiveness of our method on an image and tabular dataset, respectively MNIST and Breast Cancer Wisconsin (Diagnostic). The method also eliminates the computational bottleneck that arises because of numerical gradient evaluation for $\textit{black box}$ models.

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

A survey of algorithmic recourse: definitions, formulations, solutions, and prospects (2020, methodology): “Whilethisis notgenerallyaccepted [142, 173], one can formulate this requirement as an additional constraint, whereby, e.g., || 𝜹 || 0 ≤ 𝑠 , or || 𝜽 || 0 ≤ 𝑠 .”
Preserving Causal Constraints in Counterfactual Explanations for Machine Learning Classifiers (2019, methodology): “[20] use an auto-encoder loss term to generate counterfactuals that follow the data distribution.”
Generative causal explanations of black-box classifiers (2020, unknown): “[32] Arnaud Van Looveren and Janis Klaise.”
Multi-Objective Counterfactual Explanations (2020, unknown): “te predictions. A model-specic approach was proposed by Wachter et al. [35], who also introduced and formalized the concept of counterfactuals in predictive modeling. Like many model-specic methods [15,20,24,28,33] their approach is limited to dierentiable models. The approach of Tolomei et al. [32]”
Benchmarking and survey of explanation methods for black box models (2021, methodology): “Guided Prototypes, Interpretable Counterfactual Explanations Guided by Prototypes (guidedproto) (Looveren and Klaise 2021) proposes a local post-hoc model-specific explainer that perturbs the input image by using a loss functionL = cL pred+βL1+L2 optimized using gradient descent.”
A Survey on the Explainability of Supervised Machine Learning (2020, background): “Counterfactual instances described by Looveren and Klaise (2019) use a simple loss function and sparsely follow the explanation method of Wachter et al. (2018).”
Counterfactual Explanations and Algorithmic Recourses for Machine Learning: A Review (2020, unknown): “Black-box approaches use gradient-free optimization algorithms such as Nelder-Mead [137], growing spheres [210], FISTA [88, 332], ASP [35], or genetic algorithms [75, 208, 298] to solve the optimization problem.”
Sequential Explanations with Mental Model-Based Policies (2020, background): “…functionality and notation of interpretability, such as saliency maps (Mon-tavon et al., 2017; Binder et al., 2016; Smilkov et al., 2017), example-based (Kim et al., 2016; Dhurandhar et al., 2018; Van Looveren & Klaise, 2019; Koh & Liang, 2017), and surrogate models (Ribeiro et al., 2016; 2018).”
Globally-Consistent Rule-Based Summary-Explanations for Machine Learning Models: Application to Credit-Risk Evaluation (2019, unknown): “They further use this tree to derive counterfactual explanations in the form of minimal changes to the observation that result in an opposite prediction (for additional examples of counterfactual explanations see also work of Dhurandhar et al. 2019, Van Looveren and Klaise 2021, White and Garcez 201”
Explaining NLP Models via Minimal Contrastive Editing (MiCE) (2020, background): “, 2017), has previously been explored in fields such as computer vision (Hendricks et al., 2018; Goyal et al., 2019; Looveren and Klaise, 2019), but studied less in NLP.”
A comparison of instance-level counterfactual explanation algorithms for behavioral and textual data: SEDC, LIME-C and SHAP-C (2019, background): “Other proposed algorithms in the literature focus on relatively low-dimensional data with a mixed set of continuous and categorical variables [53, 45] or on image data where counterfactuals are computed by means of pixel perturbations [21, 28].”
SurvNAM: The machine learning survival model explanation (2021, background): “Due to intuitive and human-friendly explanations provided by this family of methods, several its modifications has been proposed [35, 36, 37, 38].”
Explainable AI: A Review of Machine Learning Interpretability Methods (2020, background): “[54] alibi PH L Agnostic tab img 17 2019”
Counterfactual Explanations for Machine Learning: A Review (2020, methodology): “Adherence to the data manifold has been addressed using several different approaches, like training VAEs on the data distribution [29, 61, 108, 82], constraining the distance of a counterfactual from the k nearest training datapoints [26, 63], directly sampling points from the latent space of a VAE…”
FOCUS: Flexible Optimizable Counterfactual Explanations for Tree Ensembles (2019, methodology): “Fidelity is commonly used to evaluate XAI methods that are based on approximations of a givenmodel: it is ameasure of the agreement between the original modelM and its approximation M̃: fid(M̃, 𝑋 )= 1 𝑁 𝑁∑︁ 𝑛=1 1 [ 𝑦𝑥 (𝑛) =arg max 𝑦′ M̃(𝑦′ |𝑥 (𝑛) ) ] , (26) where 𝑦𝑥 = arg max𝑦 M(𝑦 | �”
Towards Robust and Reliable Algorithmic Recourse (2021, background): “Several approaches in recent literature tackled the problem of providing recourses by generating local (instance level) counterfactual explanations 2 [31, 26, 12, 21, 18].”
Human-Centered Explainable AI (XAI): From Algorithms to User Experiences (2021, background): “How to Be That (a different prediction) • Highlight feature(s) that if changed (increased, decreased, absent, or present) could alter the prediction to the alternative outcome, with minimum effort required∗ • Or show examples with minimum differences but had the alternative outcome† CEM∗ [26], Count”
The Use and Misuse of Counterfactuals in Ethical Machine Learning (2021, background): “In machine learning practice, there are several technical ways to generate and evaluate counterfactuals, such as featurebased explanations, prototype explanations, example-based explanations, or causal explanations [19, 32, 35, 38, 46, 50, 51].”