Learning Model-Agnostic Counterfactual Explanations for Tabular Data (1910.09398v2)

Abstract: Counterfactual explanations can be obtained by identifying the smallest change made to a feature vector to qualitatively influence a prediction; for example, from 'loan rejected' to 'awarded' or from 'high risk of cardiovascular disease' to 'low risk'. Previous approaches often emphasized that counterfactuals should be easily interpretable to humans, motivating sparse solutions with few changes to the feature vectors. However, these approaches would not ensure that the produced counterfactuals be proximate (i.e., not local outliers) and connected to regions with substantial data density (i.e., close to correctly classified observations), two requirements known as counterfactual faithfulness. These requirements are fundamental when making suggestions to individuals that are indeed attainable. Our contribution is twofold. On one hand, we suggest to complement the catalogue of counterfactual quality measures [1] using a criterion to quantify the degree of difficulty for a certain counterfactual suggestion. On the other hand, drawing ideas from the manifold learning literature, we develop a framework that generates attainable counterfactuals. We suggest the counterfactual conditional heterogeneous variational autoencoder (C-CHVAE) to identify attainable counterfactuals that lie within regions of high data density.

Learning Model-Agnostic Counterfactual Explanations for Tabular Data

The paper examines the development and evaluation of a novel framework, dubbed the Counterfactual Conditional Heterogeneous Variational Autoencoder (C-CHVAE), designed to generate counterfactual explanations for tabular data in a model-agnostic manner. This framework aims to improve the process of deriving counterfactual explanations which, in essence, are modifications to input variables that alter the output of a classification model, producing a desired outcome. This concept is particularly relevant in fields where transparency of decision-making is critical, such as finance and healthcare.

Key Contributions

1. Framework for Faithful Counterfactuals: The paper introduces C-CHVAE, drawing heavily from manifold learning literature. This approach leverages the potential of autoencoders to model the data density, generating counterfactuals that are not only close to the original data (proximity) but also consistent with observed data regions of substantial density (connectedness). These two aspects, termed counterfactual faithfulness, have often been neglected in previous methodologies, which might produce solutions that are technically altered as required but are implausible within the data context.
2. Measurement of Suggestion Difficulty: To augment existing tools for evaluating counterfactual quality, the authors propose a new metric that quantifies the difficulty of achieving suggested counterfactuals. This metric utilizes shifts in percentile positions across the data’s cumulative distribution function, providing an intuitive measure of the effort necessary to actualize a counterfactual suggestion.

Methodology and Techniques

The C-CHVAE framework applies a structured methodology to generate counterfactuals through a series of computational steps:

• Latent Space Manipulation: By embedding the input data structure into a lower-dimensional latent space using an autoencoder, the method identifies minimal perturbations needed in this manifold to move a data point across the decision threshold of a classifier.
• Conditional Modelling: The framework handles various data types (heterogeneous data), allowing for realistic data representation in scenarios containing categorical, ordinal, and continuous variables.
• Class-Agnostic Process: The derived counterfactuals do not rely on the inner workings of a specific classifier, which makes the framework applicable to a vast range of machine learning models without modifications.

Empirical Evaluation

The framework has been thoroughly tested using multiple data sets, including synthetic and real-world credit data sets. Results demonstrate the effectiveness of the C-CHVAE in generating counterfactuals that adhere closely to the faithful criteria, with considerable improvements over existing methods in both heterogeneity handling and faithfulness of the explanations. The generated counterfactuals, while providing a stringent sense of data fidelity, tend to exhibit a higher degree of difficulty in terms of CDF shifts—something future work may aim to balance further.

Implications and Speculation on Future Work

The outcomes of this research hold several implications for theoretical advancement and practical applications:

• Theoretical Implications: By embedding the counterfactual search within the latent space approximation of data density, the framework provides a robust method compatible with deep generative models, contributing to the broader discourse on explanation and interpretability in AI systems.
• Practical Applications: The model-agnostic nature provides significant flexibility in deploying the C-CHVAE in varied industries, assisting stakeholders and end-users in understanding and possibly contesting algorithms' outputs, aligning with regulatory needs for transparency and fairness.

Future research may explore optimizing the trade-off between counterfactual attainability in terms of ease (lower CDF shifts) and fidelity to the original data manifold. Moreover, extending the framework to explicitly quantify feature importance in counterfactual scenarios is another promising avenue.

In conclusion, the C-CHVAE method constitutes a substantial step forward in the generation of meaningful and realistic counterfactual explanations, enhancing both the interpretability and accountability of AI-driven decision systems in practical, high-stakes environments.