BANDO PRIN 2022D. D. N. 104 DEL 2 FEBBRAIO 2022

TITOLO DEL PROGETTO: TRex-SE: Trustworthy Recommenders x Software Engineers 

CODICE CUP:  D53D23008730006

Budget:  € 87.060,00

P.I. o Responsabile U.R.: Fedelucio Narducci

Altre Unità di Ricerca o eventuali Sub Unità: 
Università degli Studi del SANNIO di BENEVENTO (Unisannio) 
Università degli Studi dell'AQUILA (Univaq) 

Breve descrizione del progetto
Recommender Systems for Software Engineers (RSSEs) support development tasks by suggesting APIs, code snippets, or experts for change management. The availability of data from platforms such as GitHub and Stack Overflow, together with advances in machine learning, has significantly improved their capabilities. However, this abundance of training data also introduces risks. Recommended code may be vulnerable or malicious, especially when popularity-based metrics are manipulated through fake projects or misleading discussions. Moreover, developers need transparency regarding the provenance and trustworthiness of recommendations, as security, privacy, and legal concerns are increasingly relevant—also in light of emerging AI regulations. 
In addition, RSSEs rarely incorporate user feedback effectively. Developers should be able to accept, reject, or modify recommendations, enabling human-in-the-loop scenarios that improve system accuracy and reliability. 

Finalità
The TRex-SE (Trustworthy Recommenders x Software Engineers) project aims to address these challenges by developing methods to enhance the trustworthiness of RSSEs. 

Risultati attesi
Identification of vulnerabilities and bias in RSSE recommendations: 
After assessing the presence of vulnerable (or in any case, risky) recommendations provided by RSSEs, as well as of recommendations affected by bias, this objective aims at developing vulnerability and bias-aware mining tools for RSSEs. 
Design and Evaluation of Trustworthy RSSEs: After analyzing the techniques and methodologies for designing fair, unbiased, and secure RSs, this objective aims at designing and evaluating recommendation approaches for trustworthy RSSEs. Human-in-the-loop in RSSEs: After analyzing existing RSs that provide human-in-the-loop capabilities, this objective aims to develop techniques and tools for managing user feedback in the context of RSSEs.

Risultati raggiunti (questa sezione non sarà compilata per i progetti PRIN 2022 oggetto di scorrimento, in quanto tuttora in corso) 
Identification of vulnerabilities and bias in RSSEs 

• Developing LLM-based mining and detection tools (e.g., multi-agent malware detection, semantic noise filtering) to identify vulnerable, poisoned, or inconsistent data.
• Analyzing bias and reliability issues in LLM outputs and benchmarks (e.g., prompt sensitivity, multilingual bias, data leakage).
• Building taxonomies and empirical studies to understand risks in AI-assisted development and RSSE inputs. 
   

Design and evaluation of trustworthy RSSEs 

• Designing trustworthy recommendation approaches addressing privacy (differential privacy), fairness (bias mitigation), and reproducibility (DataRec).
• Developing LLM-based tools and prototypes (e.g., SATD classification, CI/CD recommendation, test generation).
• Performing extensive empirical validation, highlighting both the effectiveness and limitations of LLM-based RSSEs. 
   

Human-in-the-loop in RSSEs 

• Creating feedback-aware tools that leverage developer input (e.g., SATD detection, improved knowledge retrieval).
• Designing mechanisms to integrate and validate user feedback in RSSE workflows.
• Enhancing transparency and trust, e.g., through AI-generated code detection (GPTSniffer) and improved interaction with LLMs.