ICTACT Journal on Soft Computing

Bhaskaran Venkatsubramaniam ¹, Pallav Kumar Baruah ¹

Affiliations
1 Master of Math and Computer Science, Sri Sathya Sai Institute of Higher Learning, IN

Abstract

Local Interpretable Model Agnostic Explanation (LIME) is a technique to explain a black box machine learning model using a surrogate model approach. While this technique is very popular, inherent to its approach, explanations are generated from the surrogate model and not directly from the black box model. In sensitive domains like healthcare, this need not be acceptable as trustworthy. These techniques also assume that features are independent and provide feature weights of the surrogate linear model as feature importance. In real life datasets, features may be dependent and a combination of a set of features with their specific values can be the deciding factor rather than individual feature importance. They also generate random instances around the point of interest to fit the surrogate model. These random instances need not be part of the original source or may even turn out to be meaningless. In this work, we compare LIME to explanations from the formal concept lattice. This does not use a surrogate model but a deterministic approach by generating synthetic data that respects implications in the original dataset and not randomly generating it. It obtains crucial feature combinations with their values as decision factors without presuming dependence or independence of features. Its explanations not only cover the point of interest but also global explanation of the model, similar and contrastive examples around the point of interest. The explanations are textual and hence easier to comprehend than comprehending weights of a surrogate linear model to understand the black box model.

Keywords

Explainable AI, XAI, Formal Concept Analysis, Lattice for XAI, Deterministic methods for XAI

Full Text

     

Bhaskaran Venkatsubramaniam ¹, Pallav Kumar Baruah ¹

Affiliations
1 Department of Math and Computer Science, Sri Sathya Sai Institute of Higher Learning, IN

Abstract

With multiple methods to extract explanations from a black box model, it becomes significant to evaluate the correctness of these Explainable AI (XAI) techniques themselves. While there are many XAI evaluation methods that need manual intervention, in order to be objective, we use computable XAI evaluation methods to test the basic nature and sanity of an XAI technique. We pick four basic axioms and three sanity tests from existing literature that the XAI techniques are expected to satisfy. Axioms like Feature Sensitivity, Implementation Invariance, Symmetry preservation and sanity tests like Model parameter randomization, Model-Outcome relationship, Input transformation invariance are used. After reviewing the axioms and sanity tests, we apply it on existing XAI techniques to check if they satisfy them or not. Thereafter, we evaluate our lattice based XAI technique with these axioms and sanity tests using a mathematical approach. This work proves these axioms and sanity tests to establish the correctness of explanations extracted from our Lattice based XAI technique.

Keywords

Explainable AI, XAI, Formal Concept Analysis, Lattice for XAI, XAI Evaluation.

Full Text