Explaining Decisions of Black-box Models using Association Rules

• Author / Creator

  Motallebi Shabestari, Mohammad Hossein

• Present-day advancements in AI, amongst other things, have often been regarding improving the accuracy of classification models.
  One lagging aspect, however, is justifying the decisions made by those models.
  Recently, AI researchers are paying more attention to fill this gap, leading to the introduction of the new field of eXplainable AI (XAI).
  Model-independent explanations are one class of explanation methods in XAI that aim at addressing the mentioned problem using techniques that have no access to the internals of a learned model.

  In this work, we introduce BARBE, a model-independent method that explains the decisions of any black-box classifier for tabular datasets with high precision.
  Moreover, the black-box classifier is not required to provide any probability score to take advantage of BARBE.
  Furthermore, BARBE presents explanations in two alternative forms: 1) the importance score for salient features, which many methods also benefit from; 2) construction of rules, which distinguishes BARBE from other methods.
  Rules are regarded as a better way to provide explanations as they align better with human intuition.
  Furthermore, BARBE exploits association rules, a special kind of rule that takes into consideration the associations between features, helping users comprehend different underlying causes of a decision.

  We also introduce BARBiE, an extension to BARBE that provides interactive explanations. This framework allows users to alter the features of an instance for which the prediction is being explained, and observe how their modifications affect the explanation and the class label.

• Subjects / Keywords

  • Model-independent Explanation
  • Interactive Explanations
  • Black-box explainer
  • Explainable Artificial Intelligence
  • XAI
  • Association Rules

• Graduation date

  Fall 2020

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-zpbq-g361

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.