Alberto Carlevaro

The supervised classification recognizes patterns in the data to separate classes of behaviors. Canonical solutions contain misclassification errors that are intrinsic to the numerical approximating nature of machine learning (ML). The data analyst may minimize the classification error on a class at the expense of increasing the error of the other classes. The error control of such a design phase is often done in a heuristic manner. In this article, it is key to develop theoretical foundations capable of providing probabilistic certifications to the obtained classifiers. In this perspective, we introduce the concept of probabilistic safety region to describe a subset of the input space in which the number of misclassified instances is probabilistically controlled. The notion of adjustable classifiers, a special class of classifiers that share the property of being controllable by a scalar parameter, is then exploited to link the tuning of ML with error control. Several tests and examples corroborate the approach. They are provided through the synthetic data in order to highlight all the steps involved, as well as notable benchmark datasets and a smart mobility application.

Recent advances in Artificial Intelligence (AI) in healthcare are driving research into solutions that can provide personalized guidance. For these solutions to be used as clinical decision support tools, the results provided must be interpretable and consistent with medical knowledge. To this end, this study explores the use of explainable AI to characterize the risk of developing cardiovascular disease in patients diagnosed with chronic obstructive pulmonary disease. A dataset of 9613 records from patients diagnosed with chronic obstructive pulmonary disease was classified into three categories of cardiovascular risk (low, moderate, and high), as estimated by the Framingham Risk Score. Counterfactual explanations were generated with two different methods, MUlti Counterfactuals via Halton sampling (MUCH) and Diverse Counterfactual Explanation (DiCE). An error control mechanism is introduced in the preliminary classification phase to reduce classification errors and obtain meaningful and representative explanations. Furthermore, the concept of counterfactual conformity is introduced as a new way to validate single counterfactual explanations in terms of their conformity, based on proximity with respect to the factual observation and plausibility. The results indicate that explanations generated with MUCH are generally more plausible (lower implausibility) and more distinguishable (higher discriminative power) from the original class than those generated with DiCE, whereas DiCE shows better availability, proximity and sparsity. Furthermore, filtering the counterfactual explanations by eliminating the non-conformal ones results in an additional improvement in quality. The results of this study suggest that combining counterfactual explanations generation with conformity evaluation is worth further validation and expert assessment to enable future development of support tools that provide personalized recommendations for reducing individual risk by targeting specific subsets of biomarkers.

Explainability has become crucial in Artificial Intelligence studies and, as the complexity of the model increases, so does the complexity of its explanation. However, the higher the complexity of the problem, the higher the amount of information it may provide, and this information can be exploited to generate a more precise explanation of how the model works. One of the most valuable ways to recover such input-output relation is to extract counterfactual explanations that allow us to find minimal changes from an observation to another one belonging to a different class. In this article, we propose a novel methodology to extract multiple counterfactual explanations (MUCH, MUlti Counterfactual via Halton sampling) from an original Multi-Class Support Vector Data Description algorithm. To evaluate the performance of the proposed method, we extracted a set of counterfactual explanations from three state-of-the-art datasets achieving satisfactory results that pave the way to a range of real-world applications.

Today, the cities we live in are far from being truly smart: overcrowding, pollution, and poor transporta- tion management are still in the headlines. With wide-scale deployment of advanced Artificial Intelligence (AI) solutions, however, it is possible to reverse this course and apply appro- priate countermeasures to take a step forward on the road to sustainability. In this research, explainable AI techniques are applied to provide public transportation experts with suggestions on how to control crowding on subway platforms by leveraging interpretable, rule-based models enhanced with counterfactual explanations. The experimental scenario relies on agent-based simulations of the De Ferrari Hitachi subway station of Genoa, Italy. Numerical results for both prediction of crowding and counterfactual (i.e., countermeasures) properties are encouraging. Moreover, an assessment of the quality of the proposed explainable methodology was submitted to a team of experts in the field to certify and validate the model.

Despite the growing availability of artificial intelligence models for predicting type 2 diabetes, there is still a lack of personalized approaches to quantify minimum viable changes in biomarkers that may help reduce the individual risk of developing disease. The aim of this article is to develop a new method, based on counterfactual explanations, to generate personalized recommendations to reduce the one-year risk of type 2 diabetes. Ten routinely collected biomarkers extracted from Electronic Medical Records of 2791 patients at low risk and 2791 patients at high risk of type 2 diabetes were analyzed. Two regions characterizing the two classes of patients were estimated using a Support Vector Data Description classifier. Counterfactual explanations (i.e., minimal changes in input features able to change the risk class) were generated for patients at high risk and evaluated using performance metrics (availability, validity, actionability, similarity, and discriminative power) and a qualitative survey administered to seven expert clinicians. Results showed that, on average, the requested minimum viable changes implied a significant reduction of fasting blood sugar, systolic blood pressure, and triglycerides and a significant increase of high-density lipoprotein in patients at risk of diabetes. A significant reduction in body mass index was also recommended in most of the patients at risk, except in females without hypertension. In general, greater changes were recommended in hypertensive patients compared to non-hypertensive ones. The experts were overall satisfied with the proposed approach although in some cases the proposed recommendations were deemed insufficient to reduce the risk in a clinically meaningful way. Future research will focus on a larger set of biomarkers and different comorbidities, also incorporating clinical guidelines whenever possible. Development of additional mathematical and clinical validation approaches will also be of paramount importance.

Machine learning (ML) algorithms are nowadays widely adopted in different contexts to perform autonomous decisions and predictions. Due to the high volume of data shared in the recent years, ML algorithms are more accurate and reliable since training and testing phases are more precise. An important concept to analyze when defining ML algorithms concerns adversarial machine learning attacks. These attacks aim to create manipulated datasets to mislead ML algorithm decisions. In this work, we propose new approaches able to detect and mitigate malicious adversarial machine learning attacks against a ML system. In particular, we investigate the Carlini-Wagner (CW), the fast gradient sign method (FGSM) and the Jacobian based saliency map (JSMA) attacks. The aim of this work is to exploit detection algorithms as countermeasures to these attacks. Initially, we performed some tests by using canonical ML algorithms with a hyperparameters optimization to improve metrics. Then, we adopt original reliable AI algorithms, either based on eXplainable AI (Logic Learning Machine) or Support Vector Data Description (SVDD). The obtained results show how the classical algorithms may fail to identify an adversarial attack, while the reliable AI methodologies are more prone to correctly detect a possible adversarial machine learning attack. The evaluation of the proposed methodology was carried out in terms of good balance between FPR and FNR on real world application datasets: Domain Name System (DNS) tunneling, Vehicle Platooning and Remaining Useful Life (RUL). In addition, a statistical analysis was performed to improve the robustness of the trained models, including evaluating their performance in terms of runtime and memory consumption.