How to Explain Individual Classification Decisions
arXiv (Cornell University)Published 6 December 2009Open access
David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, Klaus‐Robert Müller
Citations760
Generate an AI Snapshot to get a quick, structured summary of this paper.
Study Snapshot
ObjectiveStudy objective
MethodsResearch methodology
PopulationPopulation studied
Sample sizeSample sizes
OutcomesStudy outcomes here
ResultsStudy results comes here
LimitationsResearch study limitations comes here
A concise AI-generated summary of the paper will appear here once you click Generate AI Snapshot.
Abstract
After building a classifier with modern tools of machine learning we typically have a black box at hand that is able to predict well for unseen data. Thus, we get an answer to the question what is the most likely label of a given unseen data point. However, most methods will provide no answer why the model predicted the particular label for a single instance and what features were most influential for that particular instance. The only method that is currently able to provide such explanations are decision trees. This paper proposes a procedure which (based on a set of assumptions) allows to explain the decisions of any classification method.
Keywords
Computer ScienceMathematics
The Nature of Statistical Learning Theory
39,279 Citations1995Vladimir Vapnik
Springer series in statisticsThe Elements of Statistical Learning
24,344 Citations2001Trevor Hastie, J. Friedman +1 more
Choice Reviews OnlineNeural networks for pattern recognition
18,687 Citations1994
This is the first comprehensive treatment of feed-forward neural networks from the perspective of statistical pattern recognition, and is designed as a text, with over 100 exercises, to benefit anyone involved in the fields of neural computation and pattern recognition.
Annals of EugenicsTHE USE OF MULTIPLE MEASUREMENTS IN TAXONOMIC PROBLEMS
14,727 Citations1936Ronald Aylmer Fisher
The MIT Press eBooksGaussian Processes for Machine Learning
10,409 Citations2005Carl Edward Rasmussen, Christopher K. I. Williams
The book provides a long-needed, systematic and unified treatment of theoretical and practical aspects of GPs in machine learning, targeted at researchers and students in machine learning and applied statistics.
Probabilistic Outputs for Support vector Machines and Comparisons to Regularized Likelihood Methods
4,863 Citations1999John Platt
The output of a lassi(cid:12)er should be a alibrated posterior probability to enable post-pro essing and a method to train a kernel lassi with a logit link and a regularized maximum likelihood is proposed.
TechnometricsRobust Statistics: The Approach Based on Influence Functions
3,793 Citations1987David Ruppert, Frank R. Hampel +3 more
This paper presents a meta-modelling framework for estimating the values of Covariance Matrices and Multivariate Location using one-Dimensional and Multidimensional Estimators.
Applied Physics Letters10.1162/153244303322753616
3,568 Citations2000
The contributions of this special issue cover a wide range of aspects of variable selection: providing a better definition of the objective function, feature construction, feature ranking, multivariate feature selection, efficient search methods, and feature validity assessment methods.
IEEE Transactions on Neural NetworksAn introduction to kernel-based learning algorithms
3,490 Citations2001K. Müller, S. Mika +3 more
This paper provides an introduction to support vector machines, kernel Fisher discriminant analysis, and kernel principal component analysis, as examples for successful kernel-based learning methods.
Stochastic modelling and applied probabilityA Probabilistic Theory of Pattern Recognition
3,277 Citations1996Luc Devroye, László Györfi +1 more
The Bayes Error and Vapnik-Chervonenkis theory are applied as guide for empirical classifier selection on the basis of explicit specification and explicit enforcement of the maximum likelihood principle.
The MIT Press eBooksDataset Shift in Machine Learning
1,516 Citations2008
This volume offers an overview of current efforts to deal with dataset and covariate shift, and places dataset shift in relationship to transfer learning, transduction, local learning, active learning, and semi-supervised learning.
Covariate Shift Adaptation by Importance Weighted Cross Validation
743 Citations2007Masashi Sugiyama, Matthias Krauledat +1 more
Direct Importance Estimation with Model Selection and Its Application to Covariate Shift Adaptation
667 Citations2007Masashi Sugiyama, Shinichi Nakajima +3 more
This paper proposes a direct importance estimation method that does not involve density estimation and is equipped with a natural cross validation procedure and hence tuning parameters such as the kernel width can be objectively optimized.
Journal of Medicinal ChemistryDerivation and Validation of Toxicophores for Mutagenicity Prediction
656 Citations2004Jeroen Kazius, Ross McGuire +1 more
It was concluded that more specific toxicophores derived from the application of new criteria for toxicophore rule derivation and validation to a considerably sized mutagenicity dataset can be applied to risk assessment processes and can guide the design of chemical libraries for hit and lead optimization.
International Journal of Approximate ReasoningDecision theory in expert systems and artificial intelligence
371 Citations1988Eric Horvitz, John S. Breese +1 more
The potential for addressing problems in representation, inference, knowledge engineering, and explanation within the decision-theoretic framework is surveyed and the belief network and influence diagram representations are presented.
Proceedings of the National Academy of SciencesCarcinogens as Frameshift Mutagens: Metabolites and Derivatives of 2-Acetylaminofluorene and Other Aromatic Amine Carcinogens
369 Citations1972Bruce N. Ames, E G Gurney +2 more
The utility of a set of bacterial strains for detecting carcinogens as mutagens is shown and these compounds may be frameshiftmutagens of the class that intercalates into DNA and then reacts covalently with the DNA; various ultimate carcinogens may be of this type.
Journal of Chemical Information and ModelingBenchmark Data Set for in Silico Prediction of Ames Mutagenicity
367 Citations2009Katja Hansen, Sebastian Mika +6 more
A new unique public Ames mutagenicity data set comprising about 6500 nonconfidential compounds together with their biological activity is described and three commercial tools and an off-the-shelf Bayesian machine learner in Pipeline Pilot are compared.
Physical Review LettersFinding Stationary Subspaces in Multivariate Time Series
266 Citations2009Paul von Bünau, Frank C. Meinecke +2 more
This Letter proposes a novel technique, stationary subspace analysis (SSA), that decomposes a multivariate time series into its stationary and nonstationary part and succeeds in finding stationary components that lead to a significantly improved prediction accuracy and meaningful topographic maps which contribute to a better understanding of the underlyingnonstationary brain processes.
IEEE Transactions on Knowledge and Data EngineeringExplaining Classifications For Individual Instances
265 Citations2008Marko Robnik‐Šikonja, Igor Kononenko
It is demonstrated that the generated explanations closely follow the learned models and a visualization technique is presented that shows the utility of the approach and enables the comparison of different prediction methods.
Rice Digital Scholarship Archive (Rice University)Symmetrizing the Kullback-Leibler Distance
260 Citations2001Don H. Johnson, Sinan Sinanović
A new distance measure the resistor-average distance between two probability distributions that is closely related to the Kullback-Leibler distance is defined and its relation to well-known distance measures is determined.
Journal of Machine Learning ResearchAssessing Approximate Inference for Binary Gaussian Process Classification
247 Citations2005Malte Kuß, Carl Edward Rasmussen
This work reviews and compares Laplace's method and Expectation Propagation for approximate Bayesian inference in the binary Gaussian process classification model, and presents a comprehensive comparison of the approximations, their predictive performance and marginal likelihood estimates to results obtained by MCMC sampling.
Journal of Chemical Information and ModelingGaussian Processes: A Method for Automatic QSAR Modeling of ADME Properties
215 Citations2007Olga Obrezanova, Gábor Cśanyi +2 more
The basic concept of the Gaussian Processes technique is described in the context of regression problems and its application to the modeling of several ADME properties: blood-brain barrier, hERG inhibition, and aqueous solubility at pH 7.4.
Journal of VisionCenter-surround patterns emerge as optimal predictors for human saccade targets
192 Citations2009Wolf Kienzle, Matthias Franz +2 more
It is shown that center-surround patterns emerge as the optimal solution for predicting saccade targets from their local image structure, and bottom-up visual saliency may not be computed cortically as has been thought previously.
Neural NetworksA methodology to explain neural network classification
164 Citations2002Raphaël Féraud, Fabrice Clérot
A methodology to explain the classification obtained by a multilayer perceptron is proposed and a saliency measurement is introduced and defined allowing the selection of relevant variables allowing an interpretation of the neural network classifier to be built.
Explanation in Bayesian belief networks
86 Citations1992Henri J. Suermondt
This dissertation contributes a mathematical methodology that lets us determine the separate influences--on the belief-network inference result-- of individual findings, sets of findings, belief- network arcs, and chains of reasoning, which results in a set of functions that can be used to generate explanations.
Journal of Chemical Information and ModelingAccurate Solubility Prediction with Error Bars for Electrolytes: A Machine Learning Approach
80 Citations2007Anton Schwaighofer, Timon Schroeter +7 more
This work presents a statistical modeling of aqueous solubility based on measured data, using a Gaussian Process nonlinear regression model (GPsol), and shows that the developed model achieves much higher accuracy than available commercial tools for the prediction ofsolubility of electrolytes.
Journal of Chemical Information and ModelingGaussian Processes for Classification: QSAR Modeling of ADMET and Target Activity
72 Citations2010Olga Obrezanova, Matthew Segall
The results indicate that, while no method consistently generates the best model, the Gaussian processes classifier often produces more predictive models than those of the random forest or support vector machines and was rarely significantly outperformed.
Journal of Computer-Aided Molecular DesignEstimating the domain of applicability for machine learning QSAR models: a study on aqueous solubility of drug discovery molecules
67 Citations2007Timon Schroeter, Anton Schwaighofer +6 more
This work investigates the use of different Machine Learning methods to construct models for aqueous solubility, evaluating all approaches in terms of their prediction accuracy and in how far the individual error bars can faithfully represent the actual prediction error.
BioinformaticsPOIMs: positional oligomer importance matrices—understanding support vector machine-based signal detectors
67 Citations2008Sören Sonnenburg, Alexander Zien +2 more
To make SVM-based sequence classifiers more accessible and profitable, the concept of positional oligomer importance matrices (POIMs) is introduced and an efficient algorithm for their computation is proposed.
Journal of Computer-Aided Molecular DesignAutomatic QSAR modeling of ADME properties: blood–brain barrier penetration and aqueous solubility
63 Citations2008Olga Obrezanova, Joelle M. R. Gola +2 more
The results demonstrate the effectiveness of the automatic model generation process for two types of data sets commonly encountered in building ADME QSAR models, a small set of in vivo data and a large set of physico-chemical data.
Mutation Research/Fundamental and Molecular Mechanisms of MutagenesisBacterial mutagenicity investigation of epoxides: drugs, drug metabolites, steroids and pesticides
54 Citations1983Hansruedi Glatt, Reinhard Jung +1 more
35 epoxides were investigated for mutagenicity, using reversion of his- Salmonella typhimurium TA98 and TA100 as biological end-point and the cytostatic agent ethoglucide was the only drug having substantial mutagenic activity.
Journal of Chemical Information and ModelingA Probabilistic Approach to Classifying Metabolic Stability
40 Citations2008Anton Schwaighofer, Timon Schroeter +7 more
This work develops Bayesian classification models to predict the probability of a compound being metabolically stable and implicitly takes the "domain of applicability" into account for four different in vitro assays.
Molecular PharmaceuticsMachine Learning Models for Lipophilicity and Their Domain of Applicability
34 Citations2007Timon Schroeter, Anton Schwaighofer +6 more
This study constructs a log D7 model based on 14,556 drug discovery compounds of Bayer Schering Pharma, and considers error bars for each method, and investigates how well they quantify the domain of applicability of each model.
ChemMedChemPredicting Lipophilicity of Drug‐Discovery Molecules using Gaussian Process Models
32 Citations2007Timon Schroeter, Anton Schwaighofer +6 more
This study constructs different log P and log D7 models that exhibit excellent predictions which compare favorably to state-of-the-art tools on both benchmark and in-house data sets.
Lecture notes in computer scienceTowards a Model Independent Method for Explaining Classification for Individual Instances
10 Citations2008Erik Štrumbelj, Igor Kononenko
This paper proposes a new method, based on attribute interactions, that overcomes the shortcomings of the original method and serves as a theoretical framework for further work.