Performance Trade-offs in Explainability-as-a-Service Using Surrogate Linear Models

Abstract

Surrogate linear models are often employed as an explainability tool, allowing us to extract the most important features to explain a black-box model’s decisions. In the Explainability-as-a-Service paradigm (XaaS), explainability (and linear surrogate models as a by-product) can be considered as a separate service from the machine-learning (ML) model provisioning service from which they were derived. In this paper, we wish to investigate whether such surrogate models can be used in place of the original black-box models on economic grounds related to the different costs of obtaining ML models (including natively linear models) instead of surrogate models. By applying the models to three datasets (pertaining to healthcare, manufacturing, and motion sensing applications), we show that the accuracy of surrogate models is better than that of natively linear models and close to that of deep-learning models in most cases. On the other hand, surrogate models can be obtained at a lower procurement price than black-box ones. On that basis, we formulate an economics-based framework where those models can be compared on purely economic grounds, allowing us to determine the conditions under which surrogate models are economically convenient over black-box ones.

Ähnliche Arbeiten