THE WIDESPREAD ADOPTION OF ARTIFICIAL INTELLIGENCE (AI) IN SOFTWARE SYSTEMS INTRODUCES NEW LAYERS OF COMPLEXITY TO SOFTWARE QUALITY ASSURANCE. AS TECHNOLOGIES EVOLVE RAPIDLY AND DEVELOPMENT CYCLES ACCELERATE, ORGANIZATIONS OFTEN PRIORITIZE DELIVERY SPEED OVER QUALITY. THIS SHIFT LEADS DEVELOPERS TO MAKE RUSHED DESIGN DECISIONS, RESULTING IN LONG-TERM MAINTAINABILITY AND RELIABILITY ISSUES. WHEN THESE ISSUES EMERGE IN AI COMPONENTS, THEY MANIFEST AS AI-SPECIFIC TECHNICAL DEBT (AITD)—A PHENOMENON THAT CAN SEVERELY COMPROMISE SOFTWARE QUALITY AND LEAD TO PROJECT FAILURE. ALTHOUGH AITD HAS RECEIVED GROWING ATTENTION, PRIOR STUDIES HAVE LARGELY FOCUSED ON CATEGORIZING DEBT TYPES, OFFERING A LIMITED UNDERSTANDING OF HOW TO DETECT AND MANAGE THEM EFFECTIVELY. AS A RESULT, AITD CONTINUES TO ACCUMULATE THROUGHOUT SYSTEM EVOLUTION. THE RESEARCH PROJECT OF THIS THESIS AIMED TO FILL THESE GAPS, COMBINING MINING SOFTWARE REPOSITORIES WITH MIXED-METHODS RESEARCH. THE WORK (I) CHARACTERIZES THE PROPERTIES AND MANIFESTATIONS OF AITD ISSUES, (II) PROPOSES AND VALIDATES NOVEL AUTOMATED DETECTION APPROACHES, AND (III) ANALYZES THE EVOLUTIONARY ACCUMULATION OF AITD IN AI-ENABLED SYSTEMS OVER TIME. THE WORK CENTERS ON THREE PRINCIPAL CATEGORIES OF AITD—ARCHITECTURAL DEBT, MACHINE-LEARNING-SPECIFIC CODE SMELLS (ML-CSS), AND DATA SMELLS—AND EXTENDS THE INVESTIGATION TO SELF-ADMITTED TECHNICAL DEBT (SATD) TO CAPTURE HOW DEBT IS EXPLICITLY ACKNOWLEDGED BY DEVELOPERS IN SOURCE CODE COMMENTS. THE FINDINGS REVEAL DISTINCT PATTERNS FOR EACH DEBT TYPE. ARCHITECTURAL ISSUES ARE DIFFICULT FOR DEVELOPERS TO PERCEIVE DUE TO THEIR SYSTEM-LEVEL VISIBILITY. IN CONTRAST, ML-CSS AND DATA SMELLS ARE FREQUENTLY OVERLOOKED, INDICATING THEIR HIGH PREVALENCE IN AI-ENABLED SYSTEMS. FOR THESE TWO CATEGORIES, THE THESIS PROPOSES TWO TOOLS TO SUPPORT AITD DETECTION—CLAIRE AND CODESMILE—BOTH PUBLICLY AVAILABLE AND EMPIRICALLY EVALUATED. SPECIFICALLY, CLAIRE EFFECTIVELY GENERATES ACCURATE AND HUMAN-UNDERSTANDABLE REPRESENTATIONS OF ML PIPELINES. CODESMILE ACHIEVES EFFECTIVE DETECTION OF 12 ML-CSS, WITH AN AVERAGE F1-SCORE OF 83%. FINALLY, ML-CSS AND SATDS GRADUALLY EXHIBIT LONG LIFESPANS, AND OFTEN ACCUMULATE UNNOTICED THROUGHOUT SYSTEM EVOLUTION. AS A KEY CONTRIBUTION, THE THESIS PROVIDES A COMPREHENSIVE BODY OF KNOWLEDGE AND ACTIONABLE INSIGHTS FOR UNDERSTANDING, DETECTING, AND MANAGING AITD. IT SUPPORTS PRACTITIONERS BY OFFERING PRACTICAL TOOLS AND EMPIRICAL EVIDENCE TO PREVENT DEBT ACCUMULATION, AND IT GUIDES RESEARCHERS BY OUTLINING CURRENT CHALLENGES AND PRESENTING A ROADMAP TOWARD MORE MATURE, REPRODUCIBLE, AND SUSTAINABLE AITD MANAGEMENT PRACTICES IN AI-ENABLED SYSTEMS.

MIND THE DEBT: CHARACTERIZATION, DETECTION, AND EVOLUTION OF TECHNICAL DEBT IN AI-ENABLED SYSTEMS / Gilberto Recupito , 2025 Dec 05. 38. ciclo, Anno Accademico 2024/25.

MIND THE DEBT: CHARACTERIZATION, DETECTION, AND EVOLUTION OF TECHNICAL DEBT IN AI-ENABLED SYSTEMS

Recupito, Gilberto
2025

Abstract

THE WIDESPREAD ADOPTION OF ARTIFICIAL INTELLIGENCE (AI) IN SOFTWARE SYSTEMS INTRODUCES NEW LAYERS OF COMPLEXITY TO SOFTWARE QUALITY ASSURANCE. AS TECHNOLOGIES EVOLVE RAPIDLY AND DEVELOPMENT CYCLES ACCELERATE, ORGANIZATIONS OFTEN PRIORITIZE DELIVERY SPEED OVER QUALITY. THIS SHIFT LEADS DEVELOPERS TO MAKE RUSHED DESIGN DECISIONS, RESULTING IN LONG-TERM MAINTAINABILITY AND RELIABILITY ISSUES. WHEN THESE ISSUES EMERGE IN AI COMPONENTS, THEY MANIFEST AS AI-SPECIFIC TECHNICAL DEBT (AITD)—A PHENOMENON THAT CAN SEVERELY COMPROMISE SOFTWARE QUALITY AND LEAD TO PROJECT FAILURE. ALTHOUGH AITD HAS RECEIVED GROWING ATTENTION, PRIOR STUDIES HAVE LARGELY FOCUSED ON CATEGORIZING DEBT TYPES, OFFERING A LIMITED UNDERSTANDING OF HOW TO DETECT AND MANAGE THEM EFFECTIVELY. AS A RESULT, AITD CONTINUES TO ACCUMULATE THROUGHOUT SYSTEM EVOLUTION. THE RESEARCH PROJECT OF THIS THESIS AIMED TO FILL THESE GAPS, COMBINING MINING SOFTWARE REPOSITORIES WITH MIXED-METHODS RESEARCH. THE WORK (I) CHARACTERIZES THE PROPERTIES AND MANIFESTATIONS OF AITD ISSUES, (II) PROPOSES AND VALIDATES NOVEL AUTOMATED DETECTION APPROACHES, AND (III) ANALYZES THE EVOLUTIONARY ACCUMULATION OF AITD IN AI-ENABLED SYSTEMS OVER TIME. THE WORK CENTERS ON THREE PRINCIPAL CATEGORIES OF AITD—ARCHITECTURAL DEBT, MACHINE-LEARNING-SPECIFIC CODE SMELLS (ML-CSS), AND DATA SMELLS—AND EXTENDS THE INVESTIGATION TO SELF-ADMITTED TECHNICAL DEBT (SATD) TO CAPTURE HOW DEBT IS EXPLICITLY ACKNOWLEDGED BY DEVELOPERS IN SOURCE CODE COMMENTS. THE FINDINGS REVEAL DISTINCT PATTERNS FOR EACH DEBT TYPE. ARCHITECTURAL ISSUES ARE DIFFICULT FOR DEVELOPERS TO PERCEIVE DUE TO THEIR SYSTEM-LEVEL VISIBILITY. IN CONTRAST, ML-CSS AND DATA SMELLS ARE FREQUENTLY OVERLOOKED, INDICATING THEIR HIGH PREVALENCE IN AI-ENABLED SYSTEMS. FOR THESE TWO CATEGORIES, THE THESIS PROPOSES TWO TOOLS TO SUPPORT AITD DETECTION—CLAIRE AND CODESMILE—BOTH PUBLICLY AVAILABLE AND EMPIRICALLY EVALUATED. SPECIFICALLY, CLAIRE EFFECTIVELY GENERATES ACCURATE AND HUMAN-UNDERSTANDABLE REPRESENTATIONS OF ML PIPELINES. CODESMILE ACHIEVES EFFECTIVE DETECTION OF 12 ML-CSS, WITH AN AVERAGE F1-SCORE OF 83%. FINALLY, ML-CSS AND SATDS GRADUALLY EXHIBIT LONG LIFESPANS, AND OFTEN ACCUMULATE UNNOTICED THROUGHOUT SYSTEM EVOLUTION. AS A KEY CONTRIBUTION, THE THESIS PROVIDES A COMPREHENSIVE BODY OF KNOWLEDGE AND ACTIONABLE INSIGHTS FOR UNDERSTANDING, DETECTING, AND MANAGING AITD. IT SUPPORTS PRACTITIONERS BY OFFERING PRACTICAL TOOLS AND EMPIRICAL EVIDENCE TO PREVENT DEBT ACCUMULATION, AND IT GUIDES RESEARCHERS BY OUTLINING CURRENT CHALLENGES AND PRESENTING A ROADMAP TOWARD MORE MATURE, REPRODUCIBLE, AND SUSTAINABLE AITD MANAGEMENT PRACTICES IN AI-ENABLED SYSTEMS.
5-dic-2025
38
INFORMATICA
DI NUCCI, Dario
PALOMBA, Fabio
File in questo prodotto:
File Dimensione Formato  
Gilberto_Recupito___Ph_D__Abstract_cert.pdf

accesso aperto

Descrizione: Abstract
Tipologia: Tesi di dottorato
Dimensione 158.55 kB
Formato Adobe PDF
158.55 kB Adobe PDF Visualizza/Apri
Ph_D__Thesis___Gilberto_cert.pdf

accesso aperto

Descrizione: Tesi Elettronica
Tipologia: Tesi di dottorato
Dimensione 48.16 MB
Formato Adobe PDF
48.16 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11386/4969110
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact