AI2030 Responsible AI Tools Marketplace

Key Takeaways

GRACE addresses a gap in explainable AI for tabular data: While explainability tools have advanced for images and text, most existing solutions don't work well for tabular datasets with high-dimensional features—a common format in business and scientific applications that GRACE specifically targets.
The method uses "contrastive explanations" to answer "why X not Y?": Rather than explaining a prediction in isolation, GRACE generates a minimally-modified alternative scenario showing what would need to change for the model to predict a different outcome, making the reasoning more intuitive and actionable.
User studies demonstrate significant practical improvements: When people used GRACE explanations, they made approximately 60% more accurate decisions compared to users relying on LIME (a popular baseline explainability method), suggesting real-world benefits for decision-making.
The approach is grounded in established philosophical and causal reasoning principles: GRACE combines "explanation by intervention" from causality research with philosophical ideas about contrastive explanations, providing a theoretically sound foundation rather than an ad-hoc solution.
Performance validated across diverse real-world datasets: The method was tested on eleven public datasets varying significantly in scale and domain (5 to 216 features), with improvements measured across multiple dimensions including fidelity, conciseness, information gain, and influence—indicating robustness across different applications.

Abstract

Despite the recent development in the topic of explainable AI/ML for image and text data, the majority of current solutions are not suitable to explain the prediction of neural network models when the datasets are tabular and their features are in high-dimensional vectorized formats. To mitigate this limitation, therefore, we borrow two notable ideas (i.e., "explanation by intervention" from causality and "explanation are contrastive" from philosophy) and propose a novel solution, named as GRACE, that better explains neural network models' predictions for tabular datasets. In particular, given a model's prediction as label X, GRACE intervenes and generates a minimally-modified contrastive sample to be classified as Y, with an intuitive textual explanation, answering the question of "Why X rather than Y?" We carry out comprehensive experiments using eleven public datasets of different scales and domains (e.g., # of features ranges from 5 to 216) and compare GRACE with competing baselines on different measures: fidelity, conciseness, info-gain, and influence. The user-studies show that our generated explanation is not only more intuitive and easy-to-understand but also facilitates end-users to make as much as 60% more accurate post-explanation decisions than that of Lime.

Accepted at the 26th SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2020)

Key Takeaways

GRACE addresses a gap in explainable AI for tabular data: While explainability tools have advanced for images and text, most existing solutions don't work well for tabular datasets with high-dimensional features—a common format in business and scientific applications that GRACE specifically targets.
The method uses "contrastive explanations" to answer "why X not Y?": Rather than explaining a prediction in isolation, GRACE generates a minimally-modified alternative scenario showing what would need to change for the model to predict a different outcome, making the reasoning more intuitive and actionable.
User studies demonstrate significant practical improvements: When people used GRACE explanations, they made approximately 60% more accurate decisions compared to users relying on LIME (a popular baseline explainability method), suggesting real-world benefits for decision-making.
The approach is grounded in established philosophical and causal reasoning principles: GRACE combines "explanation by intervention" from causality research with philosophical ideas about contrastive explanations, providing a theoretically sound foundation rather than an ad-hoc solution.
Performance validated across diverse real-world datasets: The method was tested on eleven public datasets varying significantly in scale and domain (5 to 216 features), with improvements measured across multiple dimensions including fidelity, conciseness, information gain, and influence—indicating robustness across different applications.

Abstract

Accepted at the 26th SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2020)

GRACE: Generating Concise and Informative Contrastive Sample to Explain Neural Network Model's Prediction

Key Takeaways

Abstract

Paper Info

Responsible AI Topics

GRACE: Generating Concise and Informative Contrastive Sample to Explain Neural Network Model's Prediction

Key Takeaways

Abstract

Paper Info

Responsible AI Topics

Links