Exploring Multi-Task Learning for Explainability

Foivos Charalampakos, Iordanis Koutsopoulos

Abstract: Machine Learning (ML) model understanding and interpretation is an essential component of several applications in different domains. Several explanation techniques have been developed in order to provide insights about decisions of complex ML models. One of the most common explainability methods, Feature Attribution, assigns an importance score to each input feature that denotes its contribution (relative significance) to the complex (black-box) ML model’s decision. Such scores can be obtained through another model that acts as a surrogate, e.g., a linear one, which is trained after the black-box model so as to approximate its predictions. In this paper, we propose a training procedure based on Multi-Task Learning (MTL), where we concurrently train a black-box neural network and a surrogate linear model whose coefficients can then be used as feature significance scores. The two models exchange information through their predictions via the optimization objective which is a convex combination of a predictive loss function for the black-box model and of an explainability metric which aims to keep the predictions of the two models close together. Our method manages to make the surrogate model achieve a more accurate approximation of the black-box one, compared to the baseline of separately training the black-box and surrogate models, and therefore improves the quality of produced explanations, both global and local ones. We also achieve a good trade-off between predictive performance and explainability with minimal to negligible accuracy decrease. This enables black-box models acquired from the MTL training procedure to be used instead of normally trained models whilst being more interpretable.

External IDs:dblp:conf/ecai/CharalampakosK23