Fairness and Explanations in Group Recommender Systems. Towards Trustworthiness and transparency

Projektinformationen

FIDELITY

ID Finanzhilfevereinbarung: 101106164

DOI

10.3030/101106164

Projekt abgeschlossen

EK-Unterschriftsdatum 17 Juli 2023

Startdatum 1 Mai 2024

Enddatum 30 April 2026

Finanziert unter

• Marie Skłodowska-Curie Actions (MSCA)

Gesamtkosten

Keine Daten

EU-Beitrag

€ 181 152,96

Koordiniert durch

UNIVERSIDAD DE JAEN
Spain

Periodic Reporting for period 1 - FIDELITY (Fairness and Explanations in Group Recommender Systems. Towards Trustworthiness and transparency)

Berichtszeitraum: 2024-05-01 bis 2026-04-30

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

The Ethics Guidelines for Trustworthy AI and its subsequent updates, published by the European Commission in 2019, state that the development, deployment, and use of artificial intelligence (AI) systems should adhere to the ethical principles of respect for human autonomy, prevention of harm, fairness, and explainability. Two key elements for achieving this goal are compliance with fairness and explainability requirements. Fairness implies a commitment to ensuring that individuals and groups are free from unfair bias and discrimination in AI systems, thereby guaranteeing an equitable distribution of benefits and costs. Explainability refers to the ability to make both the technical processes of an AI system and the associated human decisions understandable. A key challenge is to improve explainability while maintaining accuracy.

Recommender systems (RSs) are essential AI-based systems in which the concepts of fairness and explainability are currently receiving increasing attention. RSs generate personalised recommendations or guide users in selecting appealing and relevant items within an overloaded information space. They have rapidly expanded across multiple domains (e.g. e-commerce, e-learning, and e-health), and this trend is expected to continue given their growing importance for businesses: recommendations account for 85% of Netflix video viewing and 30% of Amazon purchases, and the global RS market is forecast to reach USD 16.13 billion by 2026. Traditionally, most RS research has focused on recommending items to individual users. However, the need to recommend items consumed jointly by multiple users (e.g. in e-marketplaces, tourism packages, or restaurant selection) has shifted attention towards group recommender systems (GRSs).

To overcome these limitations, FIDELITY aims to develop novel algorithms and computational tools to enhance explainability, fairness, and their interplay, with a particular focus on GRSs, thereby delivering a stronger impact for end users in real-world scenarios.

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

FIDELITY develops novel post-hoc explanation models focused on group recommendations. Specifically, it introduces two families of models, respectively centred on item-based and feature-based explanations. In the former, several handcrafted templates are explored, accounting for different levels of involvement of group members in the explanation process. In the latter, a local rule-based discovery approach is employed to learn surrogate models that support the generation of feature-based explanations. In addition, SHAP is leveraged to further facilitate the understanding of explanations in group recommender systems (GRSs).

These explanation approaches have been evaluated through offline experiments on well-known public datasets, as well as on a newly collected e-service dataset gathered within the project. Furthermore, several user studies have been conducted to assess satisfaction, perceived consensus, and perceived fairness of group recommendation explanations. Additional relevant and emerging dependent variables, such as objective and subjective understanding, have also been measured.

Regarding fairness, it has been audited from both provider and consumer perspectives using public datasets and a collected e-service dataset. Moreover, a novel optimisation-based approach has been proposed to improve fairness in group recommendations, demonstrating generalisation capabilities that enable the provision of fair top-N recommendations in previously unseen scenarios.

Finally, FIDELITY develops a methodology to explain the fairness strategies used to generate top-N recommendations, particularly fairness-aware re-ranking. An extensive offline evaluation of this methodology has been carried out, showing that the proposed explanations can cover most recommendation scenarios while maintaining an appropriate level of accuracy. In addition, a user study has been conducted demonstrating the superiority of fairness-oriented explanations in terms of user satisfaction, confidence, transparency, and persuasiveness. A github repository was also deployed for incorporating these methods.

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

FIDELITY has developed novel approaches to explainability, fairness, and fairness-oriented explanations, providing an innovative framework applicable to both individual and group recommender systems. These contributions advance beyond the state of the art by explicitly addressing the interaction between fairness and explainability, a dimension still insufficiently explored in current research.

The proposed methods have been validated through offline experiments and user studies, showing improvements in user satisfaction, perceived fairness, and understanding of recommendation processes. Initial evaluations in domain-specific scenarios also indicate the potential applicability of the approaches in real-world environments.

From an impact perspective, FIDELITY contributes to the development of more transparent and trustworthy AI-based recommender systems, with particular relevance for data-intensive domains such as e-commerce and digital services. The results support the adoption of responsible AI practices aligned with European Trustworthy AI principles.

Further exploitation will benefit from large-scale validation, integration into operational platforms, and collaboration with industry stakeholders. Future work may also focus on scalability, domain adaptation, and standardisation of evaluation protocols to facilitate broader uptake and commercial applications. The FIDELITY outputs will also be used as a basis for an MSCA Staff Exchange application, within the framework of trustworthy AI systems.