Meaningful insights: explainability techniques for black-box models on tabular data

Artificial Intelligence (AI) and Machine Learning (ML) are novel data analysis techniques providing very accurate prediction results. They are widely adopted in a variety of industries to improve efficiency and decision-making, but they are also being used to develop intelligent systems. Their success grounds upon complex mathematical models, whose decisions and rationale are usually difficult to comprehend for human users to the point of being dubbed as black-boxes. This is particularly relevant in sensitive and highly regulated domains. To mitigate and possibly solve this issue, the Explainable AI (XAI) field became prominent in recent years. XAI consists of models and techniques to enable understanding of the intricated patterns discovered by black-box models. In this thesis, we consider model-agnostic XAI techniques, which can be applied to Tabular data, with a particular focus on the Credit Scoring domain. Special attention is dedicated to the LIME framework, for which we propose several modifications to the vanilla algorithm, in particular: a pair of complementary Stability Indices that accurately measure LIME stability, and the OptiLIME policy which helps the practitioner finding the proper balance among explanations' stability and reliability. We subsequently put forward GLEAMS a model-agnostic surrogate interpretable model which requires to be trained only once, while providing both Local and Global explanations of the black-box model. GLEAMS produces feature attributions and what-if scenarios, from both dataset and model perspective. Eventually, we argue that synthetic data are an emerging trend in AI, being more and more used to train complex models instead of original data. To be able to explain the outcomes of such models, we must guarantee that synthetic data are reliable enough to be able to translate their explanations to real-world individuals. To this end we propose DAISYnt, a suite of tests to measure synthetic tabular data quality and privacy.

Decomposition and reformulation of integer linear programming problems

This thesis deals with an investigation of Decomposition and Reformulation to solve Integer Linear Programming Problems. This method is often a very successful approach computationally, producing high-quality solutions for well-structured combinatorial optimization problems like vehicle routing, cutting stock, p-median and generalized assignment . However, until now the method has always been tailored to the specific problem under investigation. The principal innovation of this thesis is to develop a new framework able to apply this concept to a generic MIP problem. The new approach is thus capable of auto-decomposition and autoreformulation of the input problem applicable as a resolving black box algorithm and works as a complement and alternative to the normal resolving techniques. The idea of Decomposing and Reformulating (usually called in literature Dantzig and Wolfe Decomposition DWD) is, given a MIP, to convexify one (or more) subset(s) of constraints (slaves) and working on the partially convexified polyhedron(s) obtained. For a given MIP several decompositions can be defined depending from what sets of constraints we want to convexify. In this thesis we mainly reformulate MIPs using two sets of variables: the original variables and the extended variables (representing the exponential extreme points). The master constraints consist of the original constraints not included in any slaves plus the convexity constraint(s) and the linking constraints(ensuring that each original variable can be viewed as linear combination of extreme points of the slaves). The solution procedure consists of iteratively solving the reformulated MIP (master) and checking (pricing) if a variable of reduced costs exists, and in which case adding it to the master and solving it again (columns generation), or otherwise stopping the procedure. The advantage of using DWD is that the reformulated relaxation gives bounds stronger than the original LP relaxation, in addition it can be incorporated in a Branch and bound scheme (Branch and Price) in order to solve the problem to optimality. If the computational time for the pricing problem is reasonable this leads in practice to a stronger speed up in the solution time, specially when the convex hull of the slaves is easy to compute, usually because of its special structure.

Interactive theorem provers: issues faced as a user and tackled as a developer

Interactive theorem provers (ITP for short) are tools whose final aim is to certify proofs written by human beings. To reach that objective they have to fill the gap between the high level language used by humans for communicating and reasoning about mathematics and the lower level language that a machine is able to “understand” and process. The user perceives this gap in terms of missing features or inefficiencies. The developer tries to accommodate the user requests without increasing the already high complexity of these applications. We believe that satisfactory solutions can only come from a strong synergy between users and developers. We devoted most part of our PHD designing and developing the Matita interactive theorem prover. The software was born in the computer science department of the University of Bologna as the result of composing together all the technologies developed by the HELM team (to which we belong) for the MoWGLI project. The MoWGLI project aimed at giving accessibility through the web to the libraries of formalised mathematics of various interactive theorem provers, taking Coq as the main test case. The motivations for giving life to a new ITP are: • study the architecture of these tools, with the aim of understanding the source of their complexity • exploit such a knowledge to experiment new solutions that, for backward compatibility reasons, would be hard (if not impossible) to test on a widely used system like Coq. Matita is based on the Curry-Howard isomorphism, adopting the Calculus of Inductive Constructions (CIC) as its logical foundation. Proof objects are thus, at some extent, compatible with the ones produced with the Coq ITP, that is itself able to import and process the ones generated using Matita. Although the systems have a lot in common, they share no code at all, and even most of the algorithmic solutions are different. The thesis is composed of two parts where we respectively describe our experience as a user and a developer of interactive provers. In particular, the first part is based on two different formalisation experiences: • our internship in the Mathematical Components team (INRIA), that is formalising the finite group theory required to attack the Feit Thompson Theorem. To tackle this result, giving an effective classification of finite groups of odd order, the team adopts the SSReflect Coq extension, developed by Georges Gonthier for the proof of the four colours theorem. • our collaboration at the D.A.M.A. Project, whose goal is the formalisation of abstract measure theory in Matita leading to a constructive proof of Lebesgue’s Dominated Convergence Theorem. The most notable issues we faced, analysed in this part of the thesis, are the following: the difficulties arising when using “black box” automation in large formalisations; the impossibility for a user (especially a newcomer) to master the context of a library of already formalised results; the uncomfortable big step execution of proof commands historically adopted in ITPs; the difficult encoding of mathematical structures with a notion of inheritance in a type theory without subtyping like CIC. In the second part of the manuscript many of these issues will be analysed with the looking glasses of an ITP developer, describing the solutions we adopted in the implementation of Matita to solve these problems: integrated searching facilities to assist the user in handling large libraries of formalised results; a small step execution semantic for proof commands; a flexible implementation of coercive subtyping allowing multiple inheritance with shared substructures; automatic tactics, integrated with the searching facilities, that generates proof commands (and not only proof objects, usually kept hidden to the user) one of which specifically designed to be user driven.

I beni relazionali nella ricerca sociologica: che cosa sono e quali effetti producono

This paper studies relational goods as immaterial assets creating real effects in society. The work starts answering to this question: what kind of effects do relational goods produce? After an accurate literature examination we suppose relational goods are social relations of second order. In the hypotesis they come from the emergence of two distinct social relations: interpersonal and reflexive relations. We describe empirical evidences of these emergent assets in social life and we test the effects they produce with a model. In the work we focus on four targets. First of all we describe the emergence of relational goods through a mathematical model. Then we individualize social realities where relational goods show evident effects and we outline our scientific hypotesis. The following step consists in the formulation of empirical tests. At last we explain final results. Our aim is to set apart the constitutive structure of relational goods into a checkable model coherently with the empirical evidences shown in the research. In the study we use multi-variate analysis techniques to see relational goods in a new way and we use qualitative and quantitative strategies. Relational goods are analysed both as dependent and independent variable in order to consider causative factors acting in a black-box model. Moreover we analyse effects of relational goods inside social spheres, especially in third sector and capitalistic economy. Finally we attain to effective indexes of relational goods in order to compare them with some performance indexes.

Theoretical and Experimental Analysis of micro-CHP Energy Systems

In the framework of the micro-CHP (Combined Heat and Power) energy systems and the Distributed Generation (GD) concept, an Integrated Energy System (IES) able to meet the energy and thermal requirements of specific users, using different types of fuel to feed several micro-CHP energy sources, with the integration of electric generators of renewable energy sources (RES), electrical and thermal storage systems and the control system was conceived and built. A 5 kWel Polymer Electrolyte Membrane Fuel Cell (PEMFC) has been studied. Using experimental data obtained from various measurement campaign, the electrical and CHP PEMFC system performance have been determinate. The analysis of the effect of the water management of the anodic exhaust at variable FC loads has been carried out, and the purge process programming logic was optimized, leading also to the determination of the optimal flooding times by varying the AC FC power delivered by the cell. Furthermore, the degradation mechanisms of the PEMFC system, in particular due to the flooding of the anodic side, have been assessed using an algorithm that considers the FC like a black box, and it is able to determine the amount of not-reacted H2 and, therefore, the causes which produce that. Using experimental data that cover a two-year time span, the ageing suffered by the FC system has been tested and analyzed.

From trust who to trust what. A fresh perspective on trust in the era of blockchain

In the last few years, a remarkable technological revolution has been taking place around the world. Such a technological revolution is not just a means to the end of improving efficiency, but an important end of social order. In this work, I show that blockchain is one important example. I discuss that the success of a disintermediated transaction is conditioned on the level of trust in the rules-of-code. In what follows, the dissertation is organized in three studies. Study 1 is a theoretical paper where I present and discuss a model of how blockchain-like technologies can automate and algorithmically dictate the nurturing of trust. Based on study 1, study 2 is aimed to develop and validate a quantitative scale to measure each component of the proposed trust model. Lastly, study 3 is intended to investigate how telematic equipment can affect individual trusting behavior under condition of information asymmetry. The introduction of black box in the Italian automobile insurance market is the set of this study. Overall, the dissertation offers some major contributions for theory and practice about the dynamics underlying the development of trust in the rising era of blockchain transactions.

Intervista camminata e altri metodi qualitativi per monitorare l'efficacia dei progetti di Educazione Avventura con adolescenti a rischio

INTRODUZIONE - La presente ricerca è incentrata sul monitoraggio dell’efficacia dei progetti di Educazione Avventura con adolescenti difficili, in particolare del progetto “Lunghi cammini educativi”. A partire da un’analisi della letteratura sull’educazione esperienziale nature-based e in particolare sull’Adventure Education con adolescenti difficili, è stata progettata una rilevazione empirica attraverso cui sperimentare un metodo di monitoraggio finalizzato a cogliere la dimensione processuale (che nella ricerca nell’ambito resta spesso inesplorata, poiché sono maggiormente diffusi i metodi di monitoraggio cosiddetti “black-box”), utilizzando un sistema integrato di diverse tecniche di rilevazione. Le due principali domande che hanno guidato la ricerca sono state: 1.Quali processi educativi significativi si innescano e possono essere osservati durante l’esperienza? 2.Il metodo dell’intervista camminata, integrato ad altri metodi, è utile per individuare e monitorare questi processi? METODO - Collocandosi all’interno di un framework metodologico qualitativo (influenzato da riflessioni post-qualitative, paradigma delle mobilità e sguardo fenomenologico), la ricerca prende la forma di uno studio di caso singolo con due unità di analisi, e prevede la triangolazione di diversi metodi di raccolta dei dati: analisi documentale; osservazione partecipante nei cammini e nelle riunioni di équipe; interviste (prima, durante, dopo il cammino) con differenti tecniche: camminata, “image-elicited”, tradizionale, online. RISULTATI - L’analisi tematica abduttiva delle interviste e delle osservazioni conferma quanto già evidenziato dalla letteratura circa la centralità della dilatazione del campo d’esperienza e del lavoro su alcune life skills (in particolare, competenze personali e growth mindset). Emergono anche alcuni key findings inattesi: il notevole “peso” dello stile educativo dell’accompagnatore; la “scoperta” del ruolo della quotidianità all’interno dell’esperienza straordinaria; la necessità di consapevolezza riguardo al potenziale educativo dell’ambiente (naturale e/o antropizzato), per una maggiore intenzionalità nelle scelte strategiche di cammino. L’intervista camminata, nonostante alcuni limiti, si conferma come metodo effettivamente utile a cogliere la dimensione processuale, e coerente con il contesto indagato.

An experimental and theoretical approach to correct for the scattered radiation in an X-ray computer tomography system for industrial applications

The main problem connected to cone beam computed tomography (CT) systems for industrial applications employing 450 kV X-ray tubes is the high amount of scattered radiation which is added to the primary radiation (signal). This stray radiation leads to a significant degradation of the image quality. A better understanding of the scattering and methods to reduce its effects are therefore necessary to improve the image quality. Several studies have been carried out in the medical field at lower energies, whereas studies in industrial CT, especially for energies up to 450 kV, are lacking. Moreover, the studies reported in literature do not consider the scattered radiation generated by the CT system structure and the walls of the X-ray room (environmental scatter). In order to investigate the scattering on CT projections a GEANT4-based Monte Carlo (MC) model was developed. The model, which has been validated against experimental data, has enabled the calculation of the scattering including the environmental scatter, the optimization of an anti-scatter grid suitable for the CT system, and the optimization of the hardware components of the CT system. The investigation of multiple scattering in the CT projections showed that its contribution is 2.3 times the one of primary radiation for certain objects. The results of the environmental scatter showed that it is the major component of the scattering for aluminum box objects of front size 70 x 70 mm2 and that it strongly depends on the thickness of the object and therefore on the projection. For that reason, its correction is one of the key factors for achieving high quality images. The anti-scatter grid optimized by means of the developed MC model was found to reduce the scatter-toprimary ratio in the reconstructed images by 20 %. The object and environmental scatter calculated by means of the simulation were used to improve the scatter correction algorithm which could be patented by Empa. The results showed that the cupping effect in the corrected image is strongly reduced. The developed CT simulation is a powerful tool to optimize the design of the CT system and to evaluate the contribution of the scattered radiation to the image. Besides, it has offered a basis for a new scatter correction approach by which it has been possible to achieve images with the same spatial resolution as state-of-the-art well collimated fan-beam CT with a gain in the reconstruction time of a factor 10. This result has a high economic impact in non-destructive testing and evaluation, and reverse engineering.

Multiple testing in spatial epidemiology: a Bayesian approach

In this work we aim to propose a new approach for preliminary epidemiological studies on Standardized Mortality Ratios (SMR) collected in many spatial regions. A preliminary study on SMRs aims to formulate hypotheses to be investigated via individual epidemiological studies that avoid bias carried on by aggregated analyses. Starting from collecting disease counts and calculating expected disease counts by means of reference population disease rates, in each area an SMR is derived as the MLE under the Poisson assumption on each observation. Such estimators have high standard errors in small areas, i.e. where the expected count is low either because of the low population underlying the area or the rarity of the disease under study. Disease mapping models and other techniques for screening disease rates among the map aiming to detect anomalies and possible high-risk areas have been proposed in literature according to the classic and the Bayesian paradigm. Our proposal is approaching this issue by a decision-oriented method, which focus on multiple testing control, without however leaving the preliminary study perspective that an analysis on SMR indicators is asked to. We implement the control of the FDR, a quantity largely used to address multiple comparisons problems in the eld of microarray data analysis but which is not usually employed in disease mapping. Controlling the FDR means providing an estimate of the FDR for a set of rejected null hypotheses. The small areas issue arises diculties in applying traditional methods for FDR estimation, that are usually based only on the p-values knowledge (Benjamini and Hochberg, 1995; Storey, 2003). Tests evaluated by a traditional p-value provide weak power in small areas, where the expected number of disease cases is small. Moreover tests cannot be assumed as independent when spatial correlation between SMRs is expected, neither they are identical distributed when population underlying the map is heterogeneous. The Bayesian paradigm oers a way to overcome the inappropriateness of p-values based methods. Another peculiarity of the present work is to propose a hierarchical full Bayesian model for FDR estimation in testing many null hypothesis of absence of risk.We will use concepts of Bayesian models for disease mapping, referring in particular to the Besag York and Mollié model (1991) often used in practice for its exible prior assumption on the risks distribution across regions. The borrowing of strength between prior and likelihood typical of a hierarchical Bayesian model takes the advantage of evaluating a singular test (i.e. a test in a singular area) by means of all observations in the map under study, rather than just by means of the singular observation. This allows to improve the power test in small areas and addressing more appropriately the spatial correlation issue that suggests that relative risks are closer in spatially contiguous regions. The proposed model aims to estimate the FDR by means of the MCMC estimated posterior probabilities b i's of the null hypothesis (absence of risk) for each area. An estimate of the expected FDR conditional on data (\FDR) can be calculated in any set of b i's relative to areas declared at high-risk (where thenull hypothesis is rejected) by averaging the b i's themselves. The\FDR can be used to provide an easy decision rule for selecting high-risk areas, i.e. selecting as many as possible areas such that the\FDR is non-lower than a prexed value; we call them\FDR based decision (or selection) rules. The sensitivity and specicity of such rule depend on the accuracy of the FDR estimate, the over-estimation of FDR causing a loss of power and the under-estimation of FDR producing a loss of specicity. Moreover, our model has the interesting feature of still being able to provide an estimate of relative risk values as in the Besag York and Mollié model (1991). A simulation study to evaluate the model performance in FDR estimation accuracy, sensitivity and specificity of the decision rule, and goodness of estimation of relative risks, was set up. We chose a real map from which we generated several spatial scenarios whose counts of disease vary according to the spatial correlation degree, the size areas, the number of areas where the null hypothesis is true and the risk level in the latter areas. In summarizing simulation results we will always consider the FDR estimation in sets constituted by all b i's selected lower than a threshold t. We will show graphs of the\FDR and the true FDR (known by simulation) plotted against a threshold t to assess the FDR estimation. Varying the threshold we can learn which FDR values can be accurately estimated by the practitioner willing to apply the model (by the closeness between\FDR and true FDR). By plotting the calculated sensitivity and specicity (both known by simulation) vs the\FDR we can check the sensitivity and specicity of the corresponding\FDR based decision rules. For investigating the over-smoothing level of relative risk estimates we will compare box-plots of such estimates in high-risk areas (known by simulation), obtained by both our model and the classic Besag York Mollié model. All the summary tools are worked out for all simulated scenarios (in total 54 scenarios). Results show that FDR is well estimated (in the worst case we get an overestimation, hence a conservative FDR control) in small areas, low risk levels and spatially correlated risks scenarios, that are our primary aims. In such scenarios we have good estimates of the FDR for all values less or equal than 0.10. The sensitivity of\FDR based decision rules is generally low but specicity is high. In such scenario the use of\FDR = 0:05 or\FDR = 0:10 based selection rule can be suggested. In cases where the number of true alternative hypotheses (number of true high-risk areas) is small, also FDR = 0:15 values are well estimated, and \FDR = 0:15 based decision rules gains power maintaining an high specicity. On the other hand, in non-small areas and non-small risk level scenarios the FDR is under-estimated unless for very small values of it (much lower than 0.05); this resulting in a loss of specicity of a\FDR = 0:05 based decision rule. In such scenario\FDR = 0:05 or, even worse,\FDR = 0:1 based decision rules cannot be suggested because the true FDR is actually much higher. As regards the relative risk estimation, our model achieves almost the same results of the classic Besag York Molliè model. For this reason, our model is interesting for its ability to perform both the estimation of relative risk values and the FDR control, except for non-small areas and large risk level scenarios. A case of study is nally presented to show how the method can be used in epidemiology.

A panchromatic view of the evolution of supermassive black holes

This PhD Thesis is devoted to the accurate analysis of the physical properties of Active Galactic Nuclei (AGN) and the AGN/host-galaxy interplay. Due to the broad-band AGN emission (from radio to hard X-rays), a multi-wavelength approach is mandatory. Our research is carried out over the COSMOS field, within the context of the XMM-Newton wide-field survey. To date, the COSMOS field is a unique area for comprehensive multi-wavelength studies, allowing us to define a large and homogeneous sample of QSOs with a well-sampled spectral coverage and to keep selection effects under control. Moreover, the broad-band information contained in the COSMOS database is well-suited for a detailed analysis of AGN SEDs, bolometric luminosities and bolometric corrections. In order to investigate the nature of both obscured (Type-2) and unobscured (Type-1) AGN, the observational approach is complemented with a theoretical modelling of the AGN/galaxy co-evolution. The X-ray to optical properties of an X-ray selected Type-1 AGN sample are discussed in the first part. The relationship between X-ray and optical/UV luminosities, parametrized by the spectral index αox, provides a first indication about the nature of the central engine powering the AGN. Since a Type-1 AGN outshines the surrounding environment, it is extremely difficult to constrain the properties of its host-galaxy. Conversely, in Type-2 AGN the host-galaxy light is the dominant component of the optical/near-IR SEDs, severely affecting the recovery of the intrinsic AGN emission. Hence a multi-component SED-fitting code is developed to disentangle the emission of the stellar populationof the galaxy from that associated with mass accretion. Bolometric corrections, luminosities, stellar masses and star-formation rates, correlated with the morphology of Type-2 AGN hosts, are presented in the second part, while the final part concerns a physically-motivated model for the evolution of spheroidal galaxies with a central SMBH. The model is able to reproduce two important stages of galaxy evolution, namely the obscured cold-phase and the subsequent quiescent hot-phase.

Quantum integrability as a new method for exact results on N=2 supersymmetric gauge theories and black holes

In this thesis I show a triple new connection we found between quantum integrability, N=2 supersymmetric gauge theories and black holes perturbation theory. I use the approach of the ODE/IM correspondence between Ordinary Differential Equations (ODE) and Integrable Models (IM), first to connect basic integrability functions - the Baxter’s Q, T and Y functions - to the gauge theory periods. This fundamental identification allows several new results for both theories, for example: an exact non linear integral equation (Thermodynamic Bethe Ansatz, TBA) for the gauge periods; an interpretation of the integrability functional relations as new exact R-symmetry relations for the periods; new formulas for the local integrals of motion in terms of gauge periods. This I develop in all details at least for the SU(2) gauge theory with Nf=0,1,2 matter flavours. Still through to the ODE/IM correspondence, I connect the mathematically precise definition of quasinormal modes of black holes (having an important role in gravitational waves’ obervations) with quantization conditions on the Q, Y functions. In this way I also give a mathematical explanation of the recently found connection between quasinormal modes and N=2 supersymmetric gauge theories. Moreover, it follows a new simple and effective method to numerically compute the quasinormal modes - the TBA - which I compare with other standard methods. The spacetimes for which I show these in all details are in the simplest Nf=0 case the D3 brane in the Nf=1,2 case a generalization of extremal Reissner-Nordström (charged) black holes. Then I begin treating also the Nf=3,4 theories and argue on how our integrability-gauge-gravity correspondence can generalize to other types of black holes in either asymptotically flat (Nf=3) or Anti-de-Sitter (Nf=4) spacetime. Finally I begin to show the extension to a 4-fold correspondence with also Conformal Field Theory (CFT), through the renowned AdS/CFT correspondence.