Abduction-Based Explanations for Machine Learning Models.

摘要: The growing range of applications Machine Learning (ML) in a multitude settings motivates the ability computing small explanations for predictions made. Small are generally accepted as easier human decision makers to understand. Most earlier work on is based heuristic approaches, providing no guarantees quality, terms how close such solutions from cardinality- or subset-minimal explanations. This paper develops constraint-agnostic solution any ML model. proposed exploits abductive reasoning, and imposes requirement that model can be represented sets constraints using some target constraint reasoning system which problem answered with oracle. experimental results, obtained well-known datasets, validate scalability approach well quality computed solutions.