[Paper Review] Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives

Amit Dhurandhar, Pin‐Yu Chen+5 more|arXiv (Cornell University)|Feb 21, 2018

Explainable Artificial Intelligence (XAI)Computer Science174 citations

TL;DR

The paper introduces Contrastive Explanations Method (CEM) that identifies pertinent positives (what must be present) and pertinent negatives (what must be absent) to justify a neural network’s classification, with optional autoencoder-aided realism.

ABSTRACT

In this paper we propose a novel method that provides contrastive explanations justifying the classification of an input by a black box classifier such as a deep neural network. Given an input we find what should be %necessarily and minimally and sufficiently present (viz. important object pixels in an image) to justify its classification and analogously what should be minimally and necessarily \\emph{absent} (viz. certain background pixels). We argue that such explanations are natural for humans and are used commonly in domains such as health care and criminology. What is minimally but critically \\emph{absent} is an important part of an explanation, which to the best of our knowledge, has not been explicitly identified by current explanation methods that explain predictions of neural networks. We validate our approach on three real datasets obtained from diverse domains; namely, a handwritten digits dataset MNIST, a large procurement fraud dataset and a brain activity strength dataset. In all three cases, we witness the power of our approach in generating precise explanations that are also easy for human experts to understand and evaluate.

Motivation & Objective

• Motivate explanations that are contrastive and human-understandable by identifying minimal present/absent features that justify a classifier’s decision.
• Define and optimize perturbations to generate pertinent positives (PP) and pertinent negatives (PN) for a given input.
• Ensure explanations are close to the data manifold when possible via autoencoder-based reconstruction.
• Validate the approach across diverse domains (MNIST, procurement fraud, brain imaging) and compare with existing methods like LRP and LIME.

Proposed method

• Formulate PP and PN as optimization problems over perturbations of the input.
• Use hinge-like losses to push the perturbed input toward a different class (PN) or maintain the same top class (PP).
• Incorporate elastic-net regularization (L1 and L2 terms) for sparsity and stability.
• Optionally constrain perturbations to lie near the data manifold via an autoencoder reconstruction loss.
• Solve using projected FISTA to handle the L1 regularization efficiently.

Experimental results

Research questions

• RQ1 Can we generate instance-specific, contrastive explanations that specify minimal present and absent features for a given prediction?
• RQ2 Do Pertinent Positives and Pertinent Negatives provide more intuitive and human-aligned explanations than existing methods?
• RQ3 How does CEM perform across image and non-image domains (MNIST, procurement fraud, fMRI) compared to LRP and LIME?
• RQ4 Does enforcing proximity to the data manifold via an autoencoder improve explanation quality?

Key findings

• CEM produces interpretable explanations by highlighting which features must be present and which must be absent for a classification.
• In MNIST, CEM explanations (with and without autoencoder) are more interpretable than LRP/LIME, with autoencoder aiding clarity.
• In procurement fraud data, PP/PN explanations align closely with expert judgments, outperforming proxies for competing methods.
• In ABIDE fMRI data, CEM identifies network- and region-level patterns that align with known autism connectivity alterations, outperforming LRP in identifying pertinent positives.
• Quantitative evaluation on all three datasets shows PP/PN perturbations are 100% effective at preserving or flipping the original class when tested as separate inputs.