Prototype and Criticism Models

Description

Prototype and criticism methods summarise a dataset with a small set of real examples. Prototypes are instances chosen to represent the bulk of the data, so that together they match its overall distribution as closely as possible. Criticisms are the instances the prototypes represent worst, which shows what the summary leaves out. In a set of customer transactions, prototypes might be the usual buying patterns and criticisms might be bulk buyers or one-off high-value customers. A reviewer can read a few dozen examples instead of the whole dataset, but the method describes the data only, so how a model performs on the unusual cases needs a separate check.

Example Use Cases

Explainability

A radiology team summarises a medical imaging dataset with prototypes and criticisms. The prototypes show the typical scans that dominate the data, and the criticisms show rare presentations the data covers thinly. The team uses the criticisms to choose cases for a separate check of how accurate the model is on rare presentations.

A credit risk team summarises its borrower records with prototypes (typical low-risk profiles) and criticisms (unusual but legitimate patterns, such as irregular self-employed income). It passes the criticism cases to a separate accuracy check, to see whether the scoring model handles them as well as typical borrowers.

An engineer preparing a visual inspection model for a production line computes prototypes and criticisms for each defect class on the model's image embeddings. The criticisms show unusual captures, such as odd lighting or partial views, which the team adds to the test set so the model is checked on them.

Limitations

Which examples are chosen depends on the similarity measure and its settings, so different reasonable choices give different prototypes and criticisms. The original paper lists the effect of the kernel choice as future work, so there is no settled guidance.
The method needs a similarity score between every pair of examples, so memory and time grow with the square of the dataset size. Large datasets may need to be sampled before the method is run.
The user sets the number of prototypes and criticisms. Too few leaves large parts of the data unrepresented, and too many defeats the purpose of a summary short enough for a person to read.
It describes the data, not the model. Criticisms are the examples the prototypes represent worst, which is not the same as the examples the model gets wrong, so a separate evaluation is needed to learn how the model performs on them.
The original evidence that criticisms help people understand a dataset is a pilot study with four participants. Those who saw prototypes and criticisms were more accurate but slower than those who saw prototypes alone, so the benefit for a given audience is untested.

Resources

Research Papers

Examples are not Enough, Learn to Criticize! Criticism for Interpretability

Been Kim, Rajiv Khanna, and Oluwasanmi O. Koyejo•Dec 5, 2016

Example-based explanations are widely used in the effort to improve the interpretability of highly complex distributions. However, prototypes alone are rarely sufficient to represent the gist of the complexity. In order for users to construct better mental models and understand complex data distributions, we also need criticism to explain what are not captured by prototypes. Motivated by the Bayesian model criticism framework, we develop MMD-critic which efficiently learns prototypes and criticism, designed to aid human interpretability. A human subject pilot study shows that the MMD-critic selects prototypes and criticism that are useful to facilitate human understanding and reasoning. We also evaluate the prototypes selected by MMD-critic via a nearest prototype classifier, showing competitive performance compared to baselines.

Prototype selection for interpretable classification

Jacob Bien and Robert Tibshirani•Feb 27, 2012

Prototype methods seek a minimal subset of samples that can serve as a distillation or condensed view of a data set. As the size of modern data sets grows, being able to present a domain specialist with a short list of "representative" samples chosen from the data set is of increasing interpretative value. While much recent statistical research has been focused on producing sparse-in-the-variables methods, this paper aims at achieving sparsity in the samples. We discuss a method for selecting prototypes in the classification setting (in which the samples fall into known discrete categories). Our method of focus is derived from three basic properties that we believe a good prototype set should satisfy. This intuition is translated into a set cover optimization problem, which we solve approximately using standard approaches. While prototype selection is usually viewed as purely a means toward building an efficient classifier, in this paper we emphasize the inherent value of having a set of prototypical elements. That said, by using the nearest-neighbor rule on the set of prototypes, we can of course discuss our method as a classifier as well.

Software Packages

alibi

Feb 26, 2019

Algorithms for explaining machine learning models

MMD-Critic (BeenKim/MMD-Critic)

Feb 27, 2026

Tutorials

Prototypes and Criticisms — Interpretable Machine Learning (Molnar)

ProtoSelect on Adult Census and CIFAR10 — Alibi Explain

Jul 8, 2024

Documentations

Alibi Explain Documentation

Jan 1, 2019

Related Techniques

Name	Description	Assurance Goals
RuleFit	RuleFit builds a predictive model that a person can read: a short list of if-then rules plus a few ordinary feature weights. It first trains an ensemble of decision trees (many small trees whose outputs are combined) and turns each branch into a candidate rule, such as 'IF age > 50 AND income < £30k'. It then fits a linear model over those rules and the original features using the lasso, a penalty that sets many weights to exactly zero, so only the most useful terms remain. The surviving rules and weights can be read directly, and each term has an importance score, overall and for single predictions. Fitted to real outcomes, the result can replace an opaque model; fitted to an opaque model's predictions, it is a surrogate that approximates that model, and its fit should then be checked.	Explainability Transparency
Diverse Counterfactual Explanations	Diverse Counterfactual Explanations (DiCE) explains one decision of a classifier by generating several altered versions of the input, called counterfactuals, each changed just enough that the model would have given the other outcome, and chosen to differ from one another so that a person sees several distinct ways the result could have been different. Where the contrastive explanation method returns, for one input, a minimal set of features that must be present and a minimal set that must stay absent, DiCE returns several complete alternative inputs at once, within limits the user sets on which features may change and over what range. In the original method all the counterfactuals are optimised together by gradient descent on a loss with three parts: a hinge loss that pushes each across the model's decision threshold, a distance term (each continuous feature's change divided by its median absolute deviation in the training data, plus the share of categorical features changed) that keeps each close to the original, and a diversity term, the determinant of a similarity matrix between them, that rewards them for differing. Afterwards, small changes are reset to their original values for as long as the counterfactual still gets the other outcome, to make it sparser, and counterfactuals that break causal rules supplied by the user, such as gaining a degree without growing older, can be filtered out. The method as published needs a differentiable model such as a neural network; the reference library adds random-sampling, genetic-algorithm and nearest-training-example searches that need only the model's predictions.	Explainability Transparency
Ridge Regression Surrogates	A ridge regression surrogate approximates a complex model with a linear equation fitted to the complex model's own predictions. The complex model is run on a set of inputs, and ridge regression, a linear model with a penalty that keeps coefficients stable when features are correlated, is fitted to reproduce its outputs. Each coefficient then shows roughly how far a feature moves the complex model's output, averaged over those inputs. Fidelity, often reported as R-squared (the share of the model's variation the surrogate reproduces on held-out inputs), shows how far the coefficients can be trusted. They describe the complex model, not the real world.	Explainability Transparency