Towards the Formalization of a Trustworthy AI for Mining Interpretable Models explOiting Sophisticated Algorithms

TL;DR

The paper addresses the need for trustworthy AI by formalizing a framework (MIMOSA) that produces interpretable-by-design models while embedding causal, fairness, and privacy properties. It introduces a unified supervised-learning formulation across diverse data types, categorizes three interpretable model families (feature-importance, rule-based, instance-based), and details interpretability dimensions, reasoning, and complexity for each. It formalizes ethical properties with corresponding verification procedures and discusses inherent trade-offs, emphasizing how these properties can be embedded within the model generation process. The Rashomon effect is highlighted to motivate exploring multiple equally good models to identify solutions that jointly satisfy performance, interpretability, and ethical constraints, paving the way for practical, high-stakes AI deployment. The work lays a theoretical foundation for designing AI systems that are accurate, transparent, fair, privacy-preserving, and causally aware by design, with future directions including scalability, broader model families, empirical validation, and integration of quantum-inspired approaches.

Abstract

Interpretable-by-design models are crucial for fostering trust, accountability, and safe adoption of automated decision-making models in real-world applications. In this paper we formalize the ground for the MIMOSA (Mining Interpretable Models explOiting Sophisticated Algorithms) framework, a comprehensive methodology for generating predictive models that balance interpretability with performance while embedding key ethical properties. We formally define here the supervised learning setting across diverse decision-making tasks and data types, including tabular data, time series, images, text, transactions, and trajectories. We characterize three major families of interpretable models: feature importance, rule, and instance based models. For each family, we analyze their interpretability dimensions, reasoning mechanisms, and complexity. Beyond interpretability, we formalize three critical ethical properties, namely causality, fairness, and privacy, providing formal definitions, evaluation metrics, and verification procedures for each. We then examine the inherent trade-offs between these properties and discuss how privacy requirements, fairness constraints, and causal reasoning can be embedded within interpretable pipelines. By evaluating ethical measures during model generation, this framework establishes the theoretical foundations for developing AI systems that are not only accurate and interpretable but also fair, privacy-preserving, and causally aware, i.e., trustworthy.

Figures (8)

• Figure 1: Univariate time series on January 1st, 2011 from the Bike Sharing dataset, showing hourly variations in rental demand.
• Figure 2: Examples of medical images with different numbers of color channels: (a) grayscale and (b) RGB.
• Figure 3: Trajectory extracted from the Taxi-Drive dataset and visualized over a map of San Francisco. Black dots represent individual GPS observations; the green marker is the starting point, while the red marker is the ending point.
• Figure 4: Comparison of decision boundaries from a Logistic Regression model (Left) and a GAM (Right) on a synthetic dataset for Covid-19 diagnosis. Input features include lung capacity and exhaled carbon monoxide (CO) level. The Logistic Regression defines a linear boundary, while the GAM captures non-linear effects, enhancing flexibility and interpretability.
• Figure 5: Comparison between local and global interpretations for Logistic Regression and GAM. Local views highlight instance-specific feature contributions, i.e., weights can be multiplied by their corresponding features, yielding local feature contributions. Global views describe the general behavior of the models across the data space, i.e., weights are instead invariant.