This paper surveys Machine Learning approaches to build predictive models that know what they don't know. The consequential action of this knowledge can consist of abstaining from providing an output (rejection), deferring to another model (dynamic model selection), deferring to a human expert (learning to defer), or informing the user (uncertainty estimation). We formally state the problems each approach solves and point to key references. We discuss open issues that deserve investigation from the scientific community.

Deferring systems extend supervised Machine Learning (ML) models with the possibility to defer predictions to human experts. However, evaluating the impact of a deferring strategy on system accuracy is still an overlooked area. This paper fills this gap by evaluating deferring systems through a causal lens. We link the potential outcomes framework for causal inference with deferring systems, which allows to identify the causal impact of the deferring strategy on predictive accuracy. We distinguish two scenarios. In the first one, we have access to both the human and ML model predictions for the deferred instances. Here, we can identify the individual causal effects for deferred instances and the aggregates of them. In the second one, only human predictions are available for the deferred instances. Here, we can resort to regression discontinuity design to estimate a local causal effect. We evaluate our approach on synthetic and real datasets for seven deferring systems from the literature.

Concept Bottleneck Models (CBMs) are interpretable machine learning models that ground their predictions on human-understandable concepts, allowing for targeted interventions in their decision-making process. However, when intervened on, CBMs assume the availability of humans that can identify the need to intervene and always provide correct interventions. Both assumptions are unrealistic and impractical, considering labor costs and human error-proneness. In contrast, Learning to Defer (L2D) extends supervised learning by allowing machine learning models to identify cases where a human is more likely to be correct than the model, thus leading to deferring systems with improved performance. In this work, we gain inspiration from L2D and propose Deferring CBMs (DCBMs), a novel framework that allows CBMs to learn when an intervention is needed. To this end, we model DCBMs as a composition of deferring systems and derive a consistent L2D loss to train them. Moreover, by relying on a CBM architecture, DCBMs can explain the reasons for deferring on the final task. Our results show that DCBMs can achieve high predictive performance and interpretability by deferring only when needed.

Abstract A critical problem for several real world applications is class imbalance. Indeed, in contexts like fraud detection or medical diagnostics, standard machine learning models fail because they are designed to handle balanced class distributions. Existing solutions typically increase the rare class instances by generating synthetic records to achieve a balanced class distribution. However, these procedures generate not plausible data and tend to create unnecessary noise. We propose a change of perspective where instead of relying on resampling techniques, we depend on unsupervised features engineering approaches to represent records with a combination of features that will help the classifier capturing the differences among classes, even in presence of imbalanced data. Thus, we combine a large array of outlier detection, features projection, and features selection approaches to augment the expressiveness of the dataset population. We show the effectiveness of our proposal in a deep and wide set of benchmarking experiments as well as in real case studies.

In recent decades, advancements in information technology allowed Artificial Intelligence (AI) systems to predict future outcomes with unprecedented success. This brought the widespread deployment of these methods in many fields, intending to support decision-making. A pressing question is how to make AI systems robust to common challenges in real-life scenarios and trustworthy. In my work, I plan to explore ways to enhance the trustworthiness of AI through the selective classification framework. In this setting, the AI system can refrain from predicting whenever it is not confident enough, allowing it to trade off coverage, i.e. the percentage of instances that receive a prediction, for performance.

Selective classification (also known as classification with reject option) conservatively extends a classifier with a selection function to determine whether or not a prediction should be accepted (i.e., trusted, used, deployed). This is a highly relevant issue in socially sensitive tasks, such as credit scoring. State-of-the-art approaches rely on Deep Neural Networks (DNNs) that train at the same time both the classifier and the selection function. These approaches are model-specific and computationally expensive. We propose a model-agnostic approach, as it can work with any base probabilistic binary classification algorithm, and it can be scalable to large tabular datasets if the base classifier is so. The proposed algorithm, called SCROSS, exploits a cross-fitting strategy and theoretical results for quantile estimation to build the selection function. Experiments on real-world data show that SCROSS improves over existing methods.

There is a fast-growing literature in addressing the fairness of AI models (fair-AI), with a continuous stream of new conceptual frameworks, methods, and tools. How much can we trust them? How much do they actually impact society? We take a critical focus on fair-AI and survey issues, simplifications, and mistakes that researchers and practitioners often underestimate, which in turn can undermine the trust on fair-AI and limit its contribution to society. In particular, we discuss the hyper-focus on fairness metrics and on optimizing their average performances. We instantiate this observation by discussing the Yule's effect of fair-AI tools: being fair on average does not imply being fair in contexts that matter. We conclude that the use of fair-AI methods should be complemented with the design, development, and verification practices that are commonly summarized under the umbrella of trustworthy AI.

Selective classification (or classification with a reject option) pairs a classifier with a selection function to determine whether or not a prediction should be accepted. This framework trades off coverage (probability of accepting a prediction) with predictive performance, typically measured by distributive loss functions. In many application scenarios, such as credit scoring, performance is instead measured by ranking metrics, such as the Area Under the ROC Curve (AUC). We propose a model-agnostic approach to associate a selection function to a given probabilistic binary classifier. The approach is specifically targeted at optimizing the AUC. We provide both theoretical justifications and a novel algorithm, called AUCROSS, to achieve such a goal. Experiments show that our method succeeds in trading-off coverage for AUC, improving over existing selective classification methods targeted at optimizing accuracy.

We describe the design, the architecture, and the evaluation of the Leasing Score Prediction (LSP) system - a credit scoring and credit rating system for the leasing sector deployed at the Italian association of leasing companies. Due to its challenging objectives, the design and complexity of the LSP system represent a unique contribution to the best practices in the field. We cover requirements by managers, users, and regulations about rigorous backtesting, statistical validation, calibration, explainability, robustness and uncertainty self-assessment. LSP relies on a machine learning model trained on a mixture of data distributions contributed by many associated leasing companies. We describe the technical solutions adopted and report on their performance evaluation, including the management of the data shifts due to the COVID-19 pandemic.