On The Reasons Behind Decisions

Recent work has shown that some common machine learning classifiers can be compiled into Boolean circuits that have the same input-output behavior. We present a theory for unveiling the reasons behind the decisions made by Boolean classifiers and study some of its theoretical and practical implications. We define notions such as sufficient, necessary and complete reasons behind decisions, in addition to classifier and decision bias. We show how these notions can be used to evaluate counterfactual statements such as "a decision will stick even if ... because ... ." We present efficient algorithms for computing these notions, which are based on new advances on tractable Boolean circuits, and illustrate them using a case study.

[1]  Carlos Guestrin,et al.  Anchors: High-Precision Model-Agnostic Explanations , 2018, AAAI.

[2]  W. Quine.  On Cores and Prime Implicants of Truth Functions , 1959 .

[3]  Sofia Cassel,et al.  Graph-Based Algorithms for Boolean Function Manipulation , 2012 .

[4]  Jean-Marie Lagniez,et al.  An Improved Decision-DNNF Compiler , 2017, IJCAI.

[5]  Peter L. Hammer,et al.  Boolean Functions - Theory, Algorithms, and Applications , 2011, Encyclopedia of mathematics and its applications.

[6]  Adnan Darwiche,et al.  DPLL with a Trace: From SAT to Knowledge Compilation , 2005, IJCAI.

[7]  Olivier Coudert,et al.  Fault Tree Analysis: 1020 Prime Implicants and Beyond , 1993 .

[8]  Carlos Guestrin,et al.  "Why Should I Trust You?": Explaining the Predictions of Any Classifier , 2016, ArXiv.

[9]  Adnan Darwiche,et al.  Compiling Neural Networks into Tractable Boolean Circuits , 2019 .

[10]  Nripendra N. Biswas,et al.  Minimization of Boolean Functions , 1971, IEEE Transactions on Computers.

[11]  A. Darwiche,et al.  The Language of Search , 2007, J. Artif. Intell. Res..

[12]  Olivier Coudert,et al.  Implicit Prime Cover Computation: An Overview , 2011 .

[13]  Adnan Darwiche,et al.  A Symbolic Approach to Explaining Bayesian Network Classifiers , 2018, IJCAI.

[14]  Willard Van Orman Quine,et al.  The Problem of Simplifying Truth Functions , 1952 .

[15]  Umut Oztok,et al.  On Compiling CNF into Decision-DNNF , 2014, CP.

[16]  Raymond Reiter,et al.  Characterizing Diagnoses and Systems , 1992, Artif. Intell..

[17]  Adnan Darwiche,et al.  Formal Verification of Bayesian Network Classifiers , 2018, PGM.

[18]  Adnan Darwiche,et al.  Compiling Bayesian Network Classifiers into Decision Graphs , 2019, AAAI.

[19]  Adnan Darwiche,et al.  Decomposable negation normal form , 2001, JACM.

[20]  Joao Marques-Silva,et al.  Abduction-Based Explanations for Machine Learning Models , 2018, AAAI.

[21]  Adnan Darwiche,et al.  Reasoning about Bayesian Network Classifiers , 2002, UAI.

[22]  Adnan Darwiche,et al.  Verifying Binarized Neural Networks by Angluin-Style Learning , 2019, SAT.

[23]  Nina Narodytska,et al.  On Validating, Repairing and Refining Heuristic ML Explanations , 2019, ArXiv.

[24]  Jürg Kohlas,et al.  Handbook of Defeasible Reasoning and Uncertainty Management Systems , 2000 .

[25]  Shin-ichi Minato,et al.  Fast Generation of Prime-Irredundant Covers from Binary Decision Diagrams , 1993 .

[26]  Adnan Darwiche,et al.  New Advances in Compiling CNF into Decomposable Negation Normal Form , 2004, ECAI.

[27]  Pierre Marquis,et al.  A Knowledge Compilation Map , 2002, J. Artif. Intell. Res..

[28]  Joao Marques-Silva,et al.  On Relating Explanations and Adversarial Examples , 2019, NeurIPS.

[29]  Umut Oztok,et al.  An Exhaustive DPLL Algorithm for Model Counting , 2018, J. Artif. Intell. Res..

[30]  Andreas Herzig,et al.  Propositional Update Operators Based on Formula/Literal Dependence , 2013, TOCL.

[31]  Felix Lindner,et al.  Extracting Reasons for Moral Judgments Under Various Ethical Principles , 2019, KI.