A new method for measuring conditional feature importance using generative models.
problem Challenges in evaluating feature importance given other feature values.
method Adversarial Random Forest (ARF) for generating on-manifold data points.
result cARFi method yields robust importance scores adaptable for various feature importance notions.
Proposes a new method for interpreting feature importance and effects in dependent feature models.
problem Challenges in interpreting feature importance when features are dependent and interactions are present.
method Conditional Subgroup Approach
result Conditional PFI and PDP estimates based on this approach often outperform existing methods.
In recent years, a large amount of model-agnostic methods to improve the transparency, trustability and interpretability of machine learning models have been developed. We introduce local feature importance as a local version of a recent model-agnostic global feature importance method. Based on local feature importance…
New methods reduce extrapolation errors in feature importance.
problem Flawed feature importance methods using unrestricted permutations lead to extrapolation errors.
method Three new approaches: conditional model reliance, Knockoffs with Gaussian transformation, and restricted ALE plot designs.
result Theoretical and numerical results show our strategies reduce/eliminate extrapolation.
Introduces RFI for assessing feature importance relative to any subset of features.
problem Lack of nuanced feature importance computation.
method Generalizes PFI and CFI to assess relative feature importance.
result Derives general interpretation rules for RFI.
DFI maps covariates to latent representations for feature importance.
problem Feature importance when predictors are statistically dependent.
method Disentangled Feature Importance (DFI) using entropic optimal transport.
result DFI yields stable, interpretable, uncertainty-quantified attributions of shared predictive signal.
NGMs create mirrored features to assess neural network feature importance.
problem Lack of feature relevance information in DNNs limits their applicability.
method Structured perturbation and kernel-based conditional dependence measure for feature importance evaluation.
result Controls feature selection error rate and maintains high selection power with correlated features.
RAMPART ranks top-k features more accurately than existing methods.
problem Accurate ranking of important features in machine learning.
method Adaptive sequential halving strategy combined with ensembling techniques.
result RAMPART achieves the correct top-k ranking with high probability.
Study examines challenges in variable importance ranking due to feature correlation.
problem Challenges in variable importance ranking under correlation.
method Simulation study and theoretical analysis of feature knockoffs and conditional predictive impact (CPI).
result Highly correlated features increase the correlation of knockoff variables, posing a limitation for CPI.
Efficiently estimates SAGE values using causal structure learning.
problem Computational infeasibility of exact SAGE calculations.
method Uses causal structure learning to identify conditional independencies and accelerate SAGE approximation.
result Empirically demonstrates efficient and accurate estimation of SAGE values.
New method for mixed data FI controls type I error and achieves high power.
problem Statistical inadequacy of feature importance measures for mixed data.
method Combining CPI framework with sequential knockoffs for mixed data.
result Our method controls type I error and achieves high power for mixed data.
A new variable importance measure for DRFs detects broader impacts on output distributions.
problem Estimating full conditional distributions of multivariate outputs given inputs.
method Based on the drop and relearn principle and MMD distance.
result Consistent and high-performing variable importance measure for DRFs.
Unified feature importance for machine learning models tackles sufficiency and necessity limitations.
problem Insufficient and incomplete explanations of machine learning models.
method Formalized sufficiency and necessity notions, proposing a unified importance measure.
result Unified importance measure detects features missed by sufficiency and necessity alone.
New method R-LOCO improves local feature importance analysis.
problem Local attribution methods fail to accurately identify important features.
method R-LOCO segments input space into regions and applies global methods within.
result R-LOCO delivers more accurate local attributions.
Proposes a new method to find features affecting treatment effect distribution.
problem Existing methods fail to detect differences in treatment effect distribution parameters other than the mean.
method Formulates and estimates a feature importance measure that quantifies feature influence on potential outcome distribution discrepancies. Develops a feature selection algorithm to control type I error rate.
result Successfully discovers important features and outperforms existing mean-based methods.
MACQ method explains deep learning models by analyzing feature contributions across prediction levels.
problem Explaining deep learning model predictions.
method Global gradient-based, model-agnostic approach focusing on marginal attribution.
result MACQ separates feature contributions from interaction effects and visualizes 3-way relationships.
New method quantifies feature interactions in machine learning models.
problem Capturing high-order interactions and feature contributions in predictive models.
method Information-theoretic approach using Conditional Mutual Information (CMI) via k-NN.
result Accurately recovers feature interactions in synthetic and real-world datasets.
Enhances functional classifier performance with new tree-based methods and unbiased feature importance assessment.
problem Challenges of high-dimensional functional data and biased feature importance assessment.
method Augmented functional classification trees and random forests with ad-hoc conditional permutations for unbiased feature importance.
result Significant enhancement in predictive power of functional classifiers through new feature importance assessment.
Note that a newer expanded version of this paper is now available at: arXiv:1802.03888 It is critical in many applications to understand what features are important for a model, and why individual predictions were made. For tree ensemble methods these questions are usually answered by attributing importance values to i…
Determining an appropriate number of features for each layer in a neural network is an important and difficult task. This task is especially important in applications on systems with limited memory or processing power. Many current approaches to reduce network size either utilize iterative procedures, which can extend …
An important problem in machine learning and statistics is to identify features that causally affect the outcome. This is often impossible to do from purely observational data, and a natural relaxation is to identify features that are correlated with the outcome even conditioned on all other observed features. For exam…
The goal of feature selection is to identify important features that are relevant to explain an outcome variable. Most of the work in this domain has focused on identifying globally relevant features, which are features that are related to the outcome using evidence across the entire dataset. We study a more fine-grain…
New local MDI variable importances derived from global scores match Shapley values.
problem Local feature relevance in tree-based models.
method Deriving local MDI importance measure from global scores and linking it to Shapley values.
result Local MDI importances have a natural connection with Shapley values.
Paper presents efficient IS for tail risk estimation with machine learning features.
problem Estimating Value at Risk and Conditional Value at Risk with black-box access.
method Efficient Importance Sampling algorithm with self-structuring transformation.
result Asymptotically optimal variance reduction in logarithmic scale.
This paper provides a guide to feature importance methods for better scientific inference.
problem Limited understanding of data-generating process due to opaque ML model mechanisms.
method Comprehensive review and new proofs of global feature importance methods.
result Facilitates a thorough understanding and concrete recommendations for FI methods.
New method provides calibrated feature importance explanations for regression models.
problem Lack of uncertainty quantification in existing local explanation methods.
method Extension of Calibrated Explanations method to support regression and probabilistic regression.
result Calibrated Explanations for regression provides quantified uncertainty and robust explanations.
XAI methods struggle with identifying true predictors from suppressors in linear datasets.
problem XAI methods misidentify suppressor variables as important features.
method Carefully crafted linear ground-truth dataset to study suppressor variables; evaluated various XAI methods.
result Most XAI methods fail to distinguish true predictors from suppressors in linear settings.
PredDiff measures prediction changes while marginalizing features, offering new insights into interaction effects.
problem Understanding interaction effects in black-box models.
method Model-agnostic, local attribution method based on probability theory.
result Introduced a new measure for interaction effects between arbitrary feature subsets.
This research simplifies computation of feature attribution methods under certain conditions.
problem Computational complexity of feature attribution methods, especially power indices.
method Identifying conditions for polynomial computation and introducing new indices.
result Conditions for efficient computation of feature attribution methods are identified.
Nonparametric estimation of the conditional distribution of a response given high-dimensional features is a challenging problem. It is important to allow not only the mean but also the variance and shape of the response density to change flexibly with features, which are massive-dimensional. We propose a multiscale dic…
Optimal Bayesian feature selection (OBFS) is a multivariate supervised screening method designed from the ground up for biomarker discovery. In this work, we prove that Gaussian OBFS is strongly consistent under mild conditions, and provide rates of convergence for key posteriors in the framework. These results are of …
A framework for quantifying uncertainty in feature importance values.
problem Stable interpretation of feature importance values in machine learning models.
method A novel method based on pairwise comparisons of feature importance values to produce confidence intervals for feature ranks.
result The method produces simultaneous confidence intervals for feature ranks, enabling selection of top-k important features.
Combining feature importance estimates improves reliability of machine learning predictions.
problem Lack of consensus on feature importance quantification makes explanations unreliable.
method Proposes a feature importance fusion framework combining multiple quantifiers.
result Feature importance ensembles reduce prediction error by 15%.
New method disentangles feature importance scores in machine learning.
problem Misinterpretation of feature importance scores due to interactions and dependencies.
method Derive DIP (Disentangled Importance) decomposition of feature importance scores.
result DIP decomposition uniquely separates standalone contributions from interactions and dependencies.
This paper shows feature importance remains valid even in low-performing models.
problem Feature importance validity in low-performing machine learning models for biomedical data.
method Experiments with synthetic and real biomedical datasets to compare feature rank stability under different data reductions.
result Feature importance can be maintained even at low performance levels if data size is adequate.
Pruning method removes less important features in linear models.
problem Removing less important features in linear models trained by gradient flow.
method Iterative Magnitude Pruning (IMP) applied to linear models trained by gradient flow.
result IMP prunes features with smallest projection onto the data.
Convolutional neural networks have had a great success in numerous tasks, including image classification, object detection, sequence modelling, and many more. It is generally assumed that such neural networks are translation invariant, meaning that they can detect a given feature independent of its location in the inpu…
Paper defines feature impact and importance from data, not models.
problem Misinterpretation of feature importance as impact leads to flawed insights.
method Mathematical definitions of feature impact and importance derived from partial dependence curves.
result Feature rankings by these definitions are competitive with existing techniques.
New features from early battery cycles predict lifetime with high accuracy.
problem Accurately predicting battery lifetime under varying conditions is challenging due to manufacturing variability and usage-dependent degradation.
method Extracted features from regularly scheduled reference performance tests and used them to predict battery lifetime using a hierarchical Bayesian regression model.
result Demonstrated a lifetime prediction of in-distribution cells with 15.1% mean absolute percentage error using only the first 15% of data.
A new method explains mixed features for predictive models using conditional inference trees.
problem Explaining complex machine learning models with mixed features.
method Proposes a method to explain mixed features (continuous, discrete, ordinal, categorical) using conditional inference trees.
result Our method often outperforms current industry standards in various simulation studies and real-world financial data.
BIF assesses feature importance using Dirichlet distribution and Bayesian inference.
problem Quantitative feature importance assessment in statistical models.
method Utilizes Dirichlet distribution for probabilistic feature importance assessment via approximate Bayesian inference.
result Learned importance provides relative significance and confidence quantification of features.
A framework infers feature importance with uncertainties for high-dimensional data.
problem Estimating feature importance in high-dimensional data with uncertainty.
method Shapley value based framework, sub-SAGE, bootstrapping.
result Uncertainties in feature importance can be estimated from bootstrapping.
We examine counterfactual explanations for explaining the decisions made by model-based AI systems. The counterfactual approach we consider defines an explanation as a set of the system's data inputs that causally drives the decision (i.e., changing the inputs in the set changes the decision) and is irreducible (i.e., …
New methods compare local and global feature importance scores for bioinformatics models.
problem Improving feature importance estimation in tree-based models.
method Comparison of SHAP values and Conditional Feature Contributions (CFCs) for 164 bioinformatics problems.
result SHAP values and CFCs yield similar rankings and interpretations for random forests.
This study ranks feature-block importance in multiblock neural networks.
problem Understanding feature contributions in multiblock neural networks.
method Three methods: composite, knock-in, and knock-out strategies.
result Each strategy has its merits for specific application scenarios.
A new imputation method considers feature importance for missing value completion.
problem Missing values in datasets reduce accuracy and increase processing difficulty.
method Iterative matrix completion with feature importance learning.
result The method outperforms existing imputation algorithms on various datasets.
New approach interprets machine learning models through feature space transformations.
problem Interpreting models with strongly dependent features in high-dimensional spaces.
method Feature space transformations, including PCA and partial orthogonalization.
result Enhances model interpretation tools for domain experts.
DEDACT breaks down feature importance into direct and associative components.
problem Lack of clear distinction between direct and associative feature importance.
method DEDACT framework to decompose direct and associative importance measures.
result Provides insight into sources of prediction-relevant information and feature pathways.