Mathematical study shows post-hoc explanations are better than attention weights alone.
problem Understanding the internal behavior of attention-based models.
method Mathematical analysis of a simple attention-based architecture.
result Post-hoc explanations provide more useful insights than attention weights alone.
Counterfactual post-hoc interpretability approaches have been proven to be useful tools to generate explanations for the predictions of a trained blackbox classifier. However, the assumptions they make about the data and the classifier make them unreliable in many contexts. In this paper, we discuss three desirable pro…
Post-hoc explanations improve CNNs by replacing final linear layer with k-means classifier.
problem CNNs lack accurate data representation in their built-in prototypes.
method Introduces k-means-based post-hoc explanations for CNNs, leveraging spatial consistency of convolutional receptive fields.
result Using shallower, less compressed feature activations improves semantic fidelity at the cost of slight predictive performance.
TaylorPODA uses Taylor expansions to improve feature attributions for opaque models.
problem Lack of systematic framework for quantifying feature contributions in opaque models.
method Taylor expansion framework with postulates (precision, federation, zero-discrepancy, adaptation).
result TaylorPODA achieves competitive results and provides principled explanations.
EAGLE improves reproducibility and stability of model explanations.
problem Creating reliable explanations for opaque machine learning models.
method Formulates perturbation selection as an information-theoretic active learning problem.
result EAGLE learns a linear surrogate model with feature importance scores and uncertainty estimates.
Most of the work on interpretable machine learning has focused on designing either inherently interpretable models, which typically trade-off accuracy for interpretability, or post-hoc explanation systems, whose explanation quality can be unpredictable. Our method, ExpO, is a hybridization of these approaches that regu…
As machine learning black boxes are increasingly being deployed in domains such as healthcare and criminal justice, there is growing emphasis on building tools and techniques for explaining these black boxes in an interpretable manner. Such explanations are being leveraged by domain experts to diagnose systematic error…
PACE explains ViTs by modeling patch-level concept distributions, surpassing existing methods.
problem Lack of trustworthy post-hoc explanations for Vision Transformers (ViTs)
method Variational Bayesian explanation framework (PACE)
result PACE surpasses state-of-the-art methods in meeting desiderata for ViT explanations.
Post-hoc interpretability approaches have been proven to be powerful tools to generate explanations for the predictions made by a trained black-box model. However, they create the risk of having explanations that are a result of some artifacts learned by the model instead of actual knowledge from the data. This paper f…
New methods identify concepts in trained embeddings reliably without human labels.
problem Identifying interpretable concepts in trained embedding spaces without human labels.
method Explicitly connecting concept discovery to PCA and ICA, proposing novel approaches for dependent concepts.
result Proven methods outperform competitors on a variety of experiments, achieving up to 29% better alignment with ground truth.
LLMs can help explain credit risk models but not autonomously.
problem Leveraging LLMs for post-hoc explainability in credit risk models.
method Comparison of LLM outputs with SHAP and coefficient-based attributions on three LMs.
result LLMs reliably preserve feature-importance rankings but poorly align with autonomous explanations.
ID-ExpO fine-tunes neural networks for more faithful explanations.
problem Improving the faithfulness of explanations for complex machine learning models.
method Differentiable insertion/deletion metric-aware regularizers for optimization.
result Fine-tuned predictors produce more faithful explanations.
Attack reveals model details from counterfactual explanations.
problem Extracting model details from counterfactual explanations.
method Adversary uses counterfactual explanations to build high-fidelity model.
result High-fidelity and high-accuracy model extraction possible.
S-LIME stabilizes LIME for more reliable model explanations.
problem Instability of post hoc explanation methods like LIME.
method Uses hypothesis testing based on central limit theorem to stabilize explanations.
result Demonstrates effectiveness of S-LIME on simulated and real-world data.
Study proposes explainable analytics for manufacturing process planning.
problem Improving data-driven decision-making in manufacturing.
method Combines process mining, machine learning, and XAI. Uses deep learning for prediction and Shapley values/ICE plots for explanations.
result Enhanced decision-making capabilities through local post-hoc explanations.
Study shows explanation disparities in machine learning models are influenced by data and model properties.
problem Disparities in post-hoc machine learning explanation methods across race and gender.
method Simulations and experiments on a real-world dataset to assess challenges to explanation disparities.
result Increased covariate shift, concept shift, and omission of covariates increase explanation disparities, especially for neural network models.
Modern learning algorithms excel at producing accurate but complex models of the data. However, deploying such models in the real-world requires extra care: we must ensure their reliability, robustness, and absence of undesired biases. This motivates the development of models that are equally accurate but can be also e…
XCM improves MTS classification with explainable deep learning.
problem Lack of explainable deep learning models for MTS classification.
method XCM is a compact CNN that extracts variable and timestamp information directly from input data.
result XCM outperforms state-of-the-art MTS classifiers on large and small datasets.
This review explores methods to explain deep neural networks and their applications.
problem Understanding the decision-making process of deep neural networks.
method Overview of interpretability methods, theoretical foundations, and comparative evaluations.
result Demonstrates the effectiveness of explainable AI in various applications.
Explains agent behavior through intended outcomes in reinforcement learning.
problem Proving impossibility of general post-hoc explanations in reinforcement learning.
method Derives local explanations based on intention for Q-function approximations, proving consistency with learned Q-values.
result Demonstrates the necessity of collecting information during training for accurate explanations.
Post-hoc transforms can reverse model performance trends, especially in noisy settings.
problem Post-hoc transforms can reverse model performance trends, especially in noisy settings.
method Empirical study and analysis of post-hoc transforms like temperature scaling, ensembling, and SWA.
result Post-hoc reversal can prevent double descent and mitigate mismatches between test loss and test error.
Study robustness of global feature effect explanations in machine learning models.
problem Vulnerability of global feature effect explanations to data and model perturbations.
method Theoretical bounds and experimental evaluation of partial dependence plots and accumulated local effects.
result Quantifies the gap between best and worst-case scenarios of misinterpreting machine learning predictions globally.
New method quantifies intrinsic causal contributions in neural networks.
problem Measuring the causal influence of input features in deep neural networks.
method Proposes an identifiable generative post-hoc framework to quantify intrinsic causal contributions (ICC) as structural causal models.
result ICC generates more intuitive and reliable explanations compared to existing global explanation techniques.
Post-hoc calibration improves uncertainty under domain shift.
problem Improving uncertainty calibration under domain shift.
method Apply perturbations to validation set before post-hoc calibration.
result Perturbation step results in better calibration under domain shift.
Framework evaluates post-hoc interpretability methods in time-series classification.
problem Lack of suitable post-hoc interpretability methods for time-series classification.
method Proposes a framework with quantitative metrics to assess interpretability methods.
result Addresses several drawbacks of existing methods, including dependence on human judgement and data distribution shift.
Proposes a method for explaining tabular data using copulas.
problem Lack of ground truth for explainability in complex datasets.
method Uses copulas to specify statistical properties and build intuition.
result Demonstrates improved explainability on logistic regression and correlation use cases.
Most of the work on interpretable machine learning has focused on designing either inherently interpretable models, which typically trade-off accuracy for interpretability, or post-hoc explanation systems, which lack guarantees about their explanation quality. We propose an alternative to these approaches by directly r…
Study shows more data improves model explanations, aiding reliable knowledge extraction.
problem Challenges in deriving reliable knowledge from machine learning models due to the Rashōmon effect.
method Examined the influence of sample size on explanations from models in a Rashōmon set using SHAP.
result Explanations from <128 samples are highly variable, but agreement improves with more data.
CalArena benchmarks post-hoc calibration methods across various tasks.
problem Inconsistent evaluations of post-hoc calibration methods.
method Large-scale benchmark with 2000 experiments, covering diverse models and settings.
result Smooth calibration functions outperform binning-based approaches.
Develops causal explanations for black-box classifiers.
problem Creating understandable explanations for complex models.
method Generative model with information-theoretic measures of causal influence.
result Learned latent factors influence classifier outputs.
Meta-Cal improves post-hoc calibration of neural networks.
problem Improving the accuracy of uncalibrated neural network predictions.
method Meta-Cal uses a base calibrator and a ranking model with constraints to provide high-probability bounds.
result Meta-Cal significantly outperforms existing methods in post-hoc multi-class classification calibration.
PiNets provide faithful explanations for neural networks.
problem Lack of true explanations for neural network predictions.
method Pointwise-interpretable Networks (PiNets) that form linear models instance-wise.
result PiNets offer explanations that are meaningful, aligned, robust, and sufficient.
CUBE explains models by balanced experiments and contrasts.
problem Post-hoc explanation of trained predictive models.
method Design-based framework using balanced low-high probes.
result Reveals dominant learned effect structure and clarifies query efficiency.
Post-hoc explanations of machine learning models are crucial for people to understand and act on algorithmic predictions. An intriguing class of explanations is through counterfactuals, hypothetical examples that show people how to obtain a different prediction. We posit that effective counterfactual explanations shoul…
In this paper, we study two challenging problems in explainable AI (XAI) and data clustering. The first is how to directly design a neural network with inherent interpretability, rather than giving post-hoc explanations of a black-box model. The second is implementing discrete k-means with a differentiable neural net…
Proposes MOC method for better counterfactual explanations in ML models.
problem Difficulties in balancing multiple objectives for counterfactual explanations.
method Translates counterfactual search into a multi-objective optimization problem.
result Returns diverse counterfactuals with different trade-offs and maintains feature diversity.
Study post-hoc Learning to Defer using density-ratio losses.
problem Optimizing decision-making between models and experts.
method Density-ratio losses for post-hoc L2D scorers, derived from class-probability estimation.
result The approach recovers known results and introduces new connections to expert comparison and anomaly detection.
While statistics and machine learning offers numerous methods for ensuring generalization, these methods often fail in the presence of adaptivity---the common practice in which the choice of analysis depends on previous interactions with the same dataset. A recent line of work has introduced powerful, general purpose a…
Post-hoc calibration of neural networks using g-Layers proves theoretical justification.
problem Ensuring the confidence of neural network decisions in real-world applications.
method Proves theoretical justification for post-hoc calibration methods by adding g-Layers and minimizing NLL.
result Proves that adding g-Layers and minimizing NLL can lead to a calibrated network.
End-to-end method improves neural network calibration during training.
problem Improving neural network calibration for regression problems.
method Quantile Recalibration Training integrates post-hoc calibration into model training.
result Improved predictive accuracy and calibration in a large-scale experiment.
A post-hoc framework improves model performance by calibrating different feature spaces.
problem Improving AUC performance on binary classification tasks for overconfident models.
method Identifies heterogeneous partitions of the feature space and applies post-hoc calibration techniques to each partition.
result Theoretical optimality of the framework for any model, demonstrated on deep neural networks.
CTE improves explanation estimation with less data and faster computation.
problem Inefficient and inaccurate explanation estimation in machine learning models.
method Distribution compression through kernel thinning to reduce sample size.
result CTE significantly improves accuracy and stability of explanation estimation.
Researchers show how to manipulate Partial Dependence plots to deceive explanations of predictive models.
problem The robustness and trustworthiness of Partial Dependence (PD) explanations are compromised.
method Data poisoning using genetic and gradient algorithms to manipulate PD plots.
result PD explanations can be fooled and manipulated to mislead understanding of predictive models.
Post-detection analysis identifies responsible coordinates for multivariate change-points.
problem Identifying which coordinates in multivariate time series change after a detected change-point.
method Two-sample testing procedures with nonparametric tests for Type I error control.
result Strong performance of proposed post hoc statistical procedures.
The paper detects and identifies bias in data using a counterfactual approach.
problem Detecting and identifying bias in data, especially in medical image classification.
method A global explanation framework using the counterfactual approach to identify bias causing artifacts.
result Black frames significantly influence Convolutional Neural Network's prediction, changing benign to malignant.
Paper introduces GAMs for interpretable learning-to-rank models.
problem Need for transparent ranking models in legal or policy scenarios.
method Developed generalized additive models (GAMs) for ranking tasks using neural networks.
result Neural ranking GAMs achieve better performance than traditional GAMs while maintaining interpretability.
RAW-Explainer generates interpretable subgraph explanations for link predictions in knowledge graphs.
problem Interpreting GNN predictions for link prediction in heterogeneous settings is challenging.
method RAW-Explainer uses random walk objective and neural network to generate connected, concise subgraph explanations.
result RAW-Explainer strikes a balance between explanation quality and computational efficiency.
Most existing interpretable methods explain a black-box model in a post-hoc manner, which uses simpler models or data analysis techniques to interpret the predictions after the model is learned. However, they (a) may derive contradictory explanations on the same predictions given different methods and data samples, and…