Research areas

This line of research investigates computer science education, with a particular focus on programming. Contributions include an analysis of computer science teaching in Italian schools [Bellettini et al., 2014a] and a critique of the view that reduces programming to encoding algorithms in a programming language [Lonati et al., 2015]. Topics investigated include introductory programming through specific projects and tools [Bulgheroni and Malchiodi, 2009; Paterson et al., 2015], interdisciplinary learning activities connecting programming and music [Ludovico et al., 2017; Baratè et al., 2017a; Baratè et al., 2017b], and the teaching of advanced programming topics [Lonati et al., 2016; Lonati et al., 2017d]. A constructionist approach to programming education was analysed in [Monga et al., 2018; Lodi et al., 2019], while [Condorelli and Malchiodi, 2022] describes the design of a course on big data architectures in collaboration with an industry partner.

In collaboration with laboratorio ALaDDIn (Università degli Studi di Milano)

These activities, carried out with the ALaDDIn laboratory, aim to promote understanding of the foundations of computer science in schools and among a wider audience. They include computing competitions for students, such as Kangourou dell'Informatica [Lissoni et al., 2012; Lissoni et al., 2013; Lissoni et al., 2014; Lissoni et al., 2015] and the Italian edition of Bebras. The latter's tasks have been studied as teaching resources and analysed in terms of difficulty [Bellettini et al., 2015b; Lonati et al., 2017c; Calcagni et al., 2017; Lonati et al., 2017a; Morpurgo et al., 2018]. Constructionist activities have also been proposed to support the learning of programming and computational thinking, including activities based on music and LEGO® bricks [Baratè et al., 2017b; Baratè et al., 2017a; Monga et al., 2018; Lodi et al., 2019]. These initiatives are complemented by teacher training [Bellettini et al., 2015a; Bellettini et al., 2018a; Bellettini et al., 2018b] and contributions aimed at teachers and the general public [Monga et al., 2017; Bellettini et al., 2020].

In collaboration with laboratorio ALaDDIn (Università degli Studi di Milano)

[Bellettini et al., 2014c] introduces the algomotricity approach, which places automated information processing at the centre of computer science education. The aim is to move beyond teaching focused almost exclusively on how specific technological tools work [Lonati et al., 2015; Bellettini et al., 2014a; Bellettini et al., 2020]. The approach was evaluated in teacher qualification programmes [Bellettini et al., 2015a], with particular attention to the constructivist perspective [Bellettini et al., 2018b; Bellettini et al., 2018a]. The educational use of computing contests in non-competitive settings was also investigated [Lonati et al., 2017c], analysing how task presentation affects their educational effectiveness [Lonati et al., 2017a].

In collaboration with laboratorio ALaDDIn (Università degli Studi di Milano)

This activity investigates the use of computing contests as learning resources in primary and secondary schools, emphasising their educational purpose over competition. It includes organising nationwide contests [Lissoni et al., 2012; Lissoni et al., 2013; Lissoni et al., 2014; Lissoni et al., 2015], critically analysing their outcomes [Bellettini et al., 2015b; Lonati et al., 2017a], and studying their use in teaching [Lonati et al., 2017c; Calcagni et al., 2017; Morpurgo et al., 2018].

In collaboration with laboratorio ALaDDIn (Università degli Studi di Milano)

The algomotricity approach introduced in [Bellettini et al., 2014c; Bellettini et al., 2014b] has been applied to teaching core computer science concepts, including information representation [Bellettini et al., 2012; Bellettini et al., 2013; Baratè et al., 2017a], programming fundamentals [Baratè et al., 2017b], and recursive and greedy algorithmic strategies [Lonati et al., 2016; Lonati et al., 2017d; Lonati et al., 2017b].

In collaboration with laboratorio ALaDDIn (Università degli Studi di Milano)

This line of research investigates algorithms for learning from partially labelled data or uncertain labels, alongside methods for fuzzy set induction. In particular, [Malchiodi and Legnani, 2014] proposes an extension of support vector classification algorithms that handles both partially labelled data and uncertain labels, while [Malchiodi and Pedrycz, 2013] introduces an algorithm for learning fuzzy set membership functions, subsequently adapted to shadowed set induction in [Malchiodi and Zanaboni, 2019].

Bootstrap techniques use data resampling to estimate the sampling distribution of a statistic and quantify the uncertainty of estimates. This line of research investigates a variant first proposed in [Apolloni et al., 2006a] and further developed in [Apolloni et al., 2009; Apolloni et al., 2007a], which enables the construction of confidence regions for regression curves without relying on the usual assumptions about the distribution of measurement errors. Applications to linear and nonlinear regression are described in [Apolloni et al., 2008c], while [Apolloni et al., 2007c] presents applications in medicine.

This line of research develops Algorithmic Inference, a theoretical framework that unifies inferential problems in statistics, such as point and interval estimation of distribution parameters, and computer science, such as approximation error estimation in computational learning [Apolloni et al., 2006a; Apolloni et al., 2005e; Apolloni et al., 2002e; Apolloni et al., 2002d; Apolloni and Malchiodi, 2001b; Malchiodi, 2000]. The framework builds on earlier theoretical results on the sample complexity of learning problems [Apolloni and Malchiodi, 2001a]. It has been applied to estimating classification risk for support vector machines [Apolloni et al., 2007e; Apolloni et al., 2005b; Apolloni and Malchiodi, 2002; Apolloni and Malchiodi, 2001b] and constructing confidence regions for regression lines without assuming a Gaussian error distribution [Apolloni et al., 2005f; Apolloni et al., 2002f]. In medicine, it has been used to construct confidence regions for the hazard function associated with time to recurrence in specific cancers [Apolloni et al., 2007c; Apolloni et al., 2005g; Apolloni et al., 2002i].

Dedicated hardware circuits for computing estimates from data can reduce processing times, but require algorithms compatible with architectural constraints. This line of research applied the inference among gossips model, developed in [Malchiodi, 1996], to the construction of a family of estimators for Bernoulli populations that can be implemented directly on pRAM boards [Apolloni et al., 1997]. The same model was subsequently applied to the study of information exchange in social networks [Apolloni et al., 2013].

In computational learning theory, the structural risk minimisation principle balances model complexity against training error, with the aim of promoting generalisation. This line of research applies this principle to classifiers based on logical expressions in disjunctive or conjunctive normal form. In [Apolloni et al., 2006b; Apolloni et al., 2005h; Apolloni et al., 2003a; Apolloni et al., 2002h; Apolloni et al., 2002b], an algorithm was developed to simplify these expressions through stochastic optimisation of the parameters of the fuzzy sets that describe them.

Hybrid learning systems combine subsymbolic modules, often based on neural networks, with symbolic modules based on logical constructs. This line of research investigates how to integrate these representations and other learning models. [Apolloni et al., 2005a; Apolloni et al., 2004a] describe a system that extracts independent Boolean components (bICA) from input data features. These components are interpreted as truth values and used to construct logical formulae that symbolically represent relationships among the data [Apolloni et al., 2006b; Apolloni et al., 2003a; Apolloni et al., 2002a; Apolloni et al., 2000]. The system was applied to emotion recognition from speech signals [Apolloni et al., 2004a]. Systems of this kind were also used to monitor driver attention through biosignals [Apolloni et al., 2004c; Apolloni et al., 2004b; Apolloni et al., 2003d; Apolloni et al., 2003c; Apolloni et al., 2003b], within the ORESTEIA project (mOdular hybRid artEfactS wiTh adaptivE functIonAlity, IST-2000-26091), funded by the European Union under the IST-FET area of the Fifth Framework Programme between 2001 and 2003. A second strand integrates a fuzzy system for measuring the relevance of training examples with classification and regression models. Specifically, [Apolloni and Malchiodi, 2006] describes integration with a linear support vector classifier, while [Apolloni et al., 2005c] presents integration with a linear regression model.

This work applied the Algorithmic Inference framework described in [Apolloni et al., 2006a; Malchiodi, 2000] to automatic emotion classification from audiovisual signals [Apolloni et al., 2004a; Apolloni et al., 2002a]. The research was conducted within the PHYSTA project (Principled Hybrid Systems: Theory and Applications), funded by the European Union under the TMR programme of the Fourth Framework Programme between 1998 and 2000. The results were presented at an international school on computational learning organised within the same project.

The ORESTEIA project (mOdular hybRid artEfactS wiTh adaptivE functIonAlity) investigated the design, implementation and analysis of intelligent systems for pervasive and ubiquitous computing, which embed computing capabilities in the environment and everyday objects. The project was funded by the European Union under the IST-FET area of the Fifth Framework Programme between 2001 and 2003. As part of research into attention state detection [Kasderidis et al., 2003], a prototype was developed to recognise a car driver's attention state from physiological signals [Apolloni et al., 2004c; Apolloni et al., 2004b; Apolloni et al., 2003d; Apolloni et al., 2003c; Apolloni et al., 2003b].

Scientific computing systems support simulations and the interactive, incremental exploration of mathematical problems, providing tools for designing educational activities [Bulgheroni and Malchiodi, 2009; Malchiodi, 2008b]. A commercial system was described in detail in [Malchiodi, 2007] and extended with functionality for information encoding [Malchiodi, 2006d], remote procedure calls [Malchiodi, 2006c; Malchiodi, 2006a], and the production of scientific documentation [Malchiodi, 2011]. Further extensions address optimisation [Malchiodi, 2006b], support vector learning [Malchiodi et al., 2009b; Malchiodi et al., 2009a], and software validation [Malchiodi, 2013a]. The resulting code underpins the simulations described in [Apolloni et al., 2007d; Apolloni and Malchiodi, 2006]. Finally, [Malchiodi, 2010] presents a machine learning library built on an open-source scientific computing system.

This line of research investigates the mathematical modelling of conflict situations through an alternative to classical game theory. In [Apolloni et al., 2006d; Apolloni et al., 2003e; Apolloni et al., 2002g; Apolloni et al., 2002c], two-player games are considered, modelling conflict through the approximation of the solution to an NP-hard problem. The Algorithmic Inference framework is used to determine how limited computational resources should be allocated to the players. [Apolloni et al., 2006c] extends the approach to team games. Biological applications are described in [Apolloni et al., 2007b; Apolloni et al., 2005d], while [Apolloni et al., 2010b] uses the model to determine the computation time to allocate to learning algorithms based on local error minimisation.