Causality-Inspired Taxonomy for Explainable Artificial Intelligence

TL;DR

The paper proposes a causality-inspired taxonomy for explainable AI and validates it through a biometrics case study, analyzing 81 papers to map current xAI approaches onto a Ladder of xAI. By reframing explanations through causality, it identifies where methods provide observational versus interventional or counterfactual insights, highlighting gaps toward true causal understanding. The biometrics review reveals heavy reliance on first-rung visualization techniques, with substantial opportunities to adopt higher-order, semantically rich explanations and domain-informed priors. The work argues that a causality-grounded framework can improve evaluation, feedback, and privacy while enabling bias detection and safer deployment across biometric modalities, and it outlines concrete future directions including multi-modal data, human-aligned explanations, and integrated in-model and post-hoc strategies.

Abstract

As two sides of the same coin, causality and explainable artificial intelligence (xAI) were initially proposed and developed with different goals. However, the latter can only be complete when seen through the lens of the causality framework. As such, we propose a novel causality-inspired framework for xAI that creates an environment for the development of xAI approaches. To show its applicability, biometrics was used as case study. For this, we have analysed 81 research papers on a myriad of biometric modalities and different tasks. We have categorised each of these methods according to our novel xAI Ladder and discussed the future directions of the field.

Figures (11)

• Figure 1: Different tasks of biometric systems. In (a) users add their information to the database. In (b) users claim to be someone and and the system attempts an 1:1 comparison. In (c) the system tries to match the user with all the identities in an 1:N problem.
• Figure 2: Vulnerability of face recognition systems to natural-looking adversarial input images that drive the model to incorrect output predictions.
• Figure 3: The Ladder of Causality (adapted from pearl2018book).
• Figure 4: How we can explain black box models: a summary of current post hoc explainability approaches (adapted from BarredoArrieta2020).
• Figure 5: A figure illustrating the World's Causal Model which represents all the causal relationships that we observe and rule the functioning of the world. Below we see a representation of the Causal Model of a Deep Neural Network. While the latter represents the causal relationships within the "rules" learnt by the DNN, the former represents the true causal relationships of the world. These two models are very likely to disagree, and the latter does not aim to explain the causal relationships of the world, just the specific deep learning model it represents.