Lanzetta, Vincenzo (2025) Automatic/semiautomatic labeling for scenario analysis and predictions of economic systems. [Tesi di dottorato]

Documento PDF Lanzetta_Vincenzo_37.pdf Visibile a [TBR] Amministratori dell'archivio Download (2MB) | Richiedi una copia

Abstract

In recent years, Machine Learning (ML) methods have shown remarkable success in uncovering hidden patterns and making highly accurate predictions, even from large and unstructured datasets. These methods have gained significant traction in fields ranging from computer vision and natural language processing to social and economic systems. Interestingly, economic systems exhibit a number of challenging issues. They are dynamic systems, characterized by non-linear interactions and dependencies among economic actors, and exhibit a high degree of complexity, influenced by various socio-economic, cultural, and political factors; according to literature evidence that machine learning methodologies have better capability – then traditional statistical methodologies - for capturing non linear relationship in the data, this thesis aims to contribute to the growing body of literature on the application of ML methods in economic systems - with respect to the regional innovation capability as particular economic topic to be analyzed; thus, this thesis is focused on some critical issues in the development of novel ML methodologies to enhance predictive accuracy and scenario analysis within the regional innovation capability topic (where "regional" refers to the territorial unit NUTS2 - Nomenclature of territorial units for statistics - of the classification system implemented by the European Union for statistical purposes). In particular, this thesis - by starting with the analysis of some methodological weaknesses - related to the regional grouping and labeling problem of the main regional innovation European tool used by European policymakers for regional innovation policy purposes (i.e.: the European Union Regional Innovation Scoreboard - EURIS) - suggests the adoption of alternative grouping/labeling methodologies (unsupervised and supervised ones) than the EURIS one, to be used in pair with a what-if neural network-based tool, with the final aim to develop a synergystic tool of the EURIS; experimental results suggested that our proposed (unsupervised and supervised) grouping/labeling methodologies are able to develop more cohesive groups, resulting in regions having better similarities, than the ones developed by the EURIS methodology. Experimental simulations within the framework of our proposed what-if tool highlights the potential usefulness of the latter methodological tool (neural network-based) for the understanding of the potential effectiveness of each possible and specific regional innovation policy to be implemented. Furthermore, by analyzing the data of the variables underlying the development of regional innovative capacity, a critical aspect emerged - in the context of machine learning approaches - related to the problem of “dataset shift”; this problem occurs with probability distributions that - within the same variable - can change “overtime”. Literature points out that dataset shift can lead to significant deterioration in the performance of ML systems, if not well addressed with appropriate transfer learning techniques, such as - for example - with “domain adaptation” and “domain generalization” techniques; to this end, the present work has also focused on a thorough analysis of transfer learning techniques, as potentially useful techniques to effectively address the above problem; in particular, a detailed literature review was conducted on the use of transfer learning techniques in economic contexts, with the aim of shedding light on the potential, and limitations of such approaches, and with respect to possible applications for the problem under analysis. However, due to time constraints, this line of research has only been preliminarily investigated, and not yet fully developed/tested empirically. As main conclusion of the thesis, this work claims the need to integrate the current EURIS methodology with an/a unsupervised/supervised tool for grouping/labeling purposes, and with a neural network (NN) tool for performing what-if policy analyses; in other words, this thesis claims the potential usefulness of our comprehensive methodology (unsupervised/supervised labeling method, and NN-based method on labels resulting from unsupervised/supervised labeling methods) as a comprehensive synergistic tool of the European Union Regional Innovation Scoreboard, in order to provide European regional policymakers with an EURIS synergistic tool aimed at developing targeted regional innovation policies. Part I is dedicated to the survey of transfer learning in the economic domain; Part II is focused on regional innovation capability as selected economic topic for the experimenting of novel ML methodologies related to prediction and scenario analysis.

Downloads

Actions (login required)

Modifica documento