Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AIDE: Antithetical, Intent-based, and Diverse Example-Based Explanations

Ikhtiyor Nematov, Dimitris Sacharidis, Tomer Sagi, Katja Hose

TL;DR

AIDE addresses the need for explanations that reveal contrasting reasons behind model predictions by identifying influential training samples. It uses a first-order influence-function approximation to compute sample influence and organizes explanations into four quadrants (S, SC, O, OC) to provide contrastive views aligned with three user intents. AIDE employs proximity- and diversity-aware, greedy sampling with an IQR-based outlier filter and a weighted objective to select a compact, informative set of samples for each intent. Across text and image classification tasks, AIDE demonstrates improved faithfulness and continuity over baselines, and a user study supports the value of intent-aware, contrastive explanations for human-AI collaboration.

Abstract

For many use-cases, it is often important to explain the prediction of a black-box model by identifying the most influential training data samples. Existing approaches lack customization for user intent and often provide a homogeneous set of explanation samples, failing to reveal the model's reasoning from different angles. In this paper, we propose AIDE, an approach for providing antithetical (i.e., contrastive), intent-based, diverse explanations for opaque and complex models. AIDE distinguishes three types of explainability intents: interpreting a correct, investigating a wrong, and clarifying an ambiguous prediction. For each intent, AIDE selects an appropriate set of influential training samples that support or oppose the prediction either directly or by contrast. To provide a succinct summary, AIDE uses diversity-aware sampling to avoid redundancy and increase coverage of the training data. We demonstrate the effectiveness of AIDE on image and text classification tasks, in three ways: quantitatively, assessing correctness and continuity; qualitatively, comparing anecdotal evidence from AIDE and other example-based approaches; and via a user study, evaluating multiple aspects of AIDE. The results show that AIDE addresses the limitations of existing methods and exhibits desirable traits for an explainability method.

AIDE: Antithetical, Intent-based, and Diverse Example-Based Explanations

TL;DR

AIDE addresses the need for explanations that reveal contrasting reasons behind model predictions by identifying influential training samples. It uses a first-order influence-function approximation to compute sample influence and organizes explanations into four quadrants (S, SC, O, OC) to provide contrastive views aligned with three user intents. AIDE employs proximity- and diversity-aware, greedy sampling with an IQR-based outlier filter and a weighted objective to select a compact, informative set of samples for each intent. Across text and image classification tasks, AIDE demonstrates improved faithfulness and continuity over baselines, and a user study supports the value of intent-aware, contrastive explanations for human-AI collaboration.

Abstract

For many use-cases, it is often important to explain the prediction of a black-box model by identifying the most influential training data samples. Existing approaches lack customization for user intent and often provide a homogeneous set of explanation samples, failing to reveal the model's reasoning from different angles. In this paper, we propose AIDE, an approach for providing antithetical (i.e., contrastive), intent-based, diverse explanations for opaque and complex models. AIDE distinguishes three types of explainability intents: interpreting a correct, investigating a wrong, and clarifying an ambiguous prediction. For each intent, AIDE selects an appropriate set of influential training samples that support or oppose the prediction either directly or by contrast. To provide a succinct summary, AIDE uses diversity-aware sampling to avoid redundancy and increase coverage of the training data. We demonstrate the effectiveness of AIDE on image and text classification tasks, in three ways: quantitatively, assessing correctness and continuity; qualitatively, comparing anecdotal evidence from AIDE and other example-based approaches; and via a user study, evaluating multiple aspects of AIDE. The results show that AIDE addresses the limitations of existing methods and exhibits desirable traits for an explainability method.
Paper Structure (14 sections, 1 theorem, 12 equations, 11 figures, 10 tables)

This paper contains 14 sections, 1 theorem, 12 equations, 11 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The AIDE Framework
  4. Preliminaries
  5. AIDE Ingredients
  6. AIDE Quadrants
  7. Explanation Intents
  8. Experiments
  9. Datasets, Models, and Methods
  10. Quantitative Evaluation
  11. Qualitative Evaluation
  12. User Study
  13. Conclusion
  14. Limitations

Key Result

Lemma 1

In binary classification with logistic loss, the influence of a training point ${\bm{z}}$ to the predictions of ${\bm{z}}_t = ({\bm{x}}_t, y_t)$ and ${\bm{z}}_t' = ({\bm{x}}_t, 1-y_t)$ is related as follows:

Figures (11)

  • Figure 1: Explanations for a spam classification task, depicting a correctly classified spam message and its influence-based explanations generated by IF and AIDE.
  • Figure 2: Explanations to clarify an ambiguous prediction.
  • Figure 3: Continuity in terms of explanation similarity vs. instance pair similarity in spam dataset.
  • Figure 4: Continuity in terms of explanation similarity vs. instance pair similarity in image dataset.
  • Figure 5: Explanations to interpret a correct prediction.
  • ...and 6 more figures

Theorems & Definitions (2)

  • Lemma 1
  • proof