Fairness-Aware Recommender Systems : Extending the SoCRATe Framework with Neural and Generative Models
Imam, Shaheer Ghani (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-12-03
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025112510865
https://urn.fi/URN:NBN:fi:tuni-2025112510865
Tiivistelmä
Recommender systems play a vital role in helping users discover relevant items in domains such as e-commerce, media, and crowdsourcing. However, most existing approaches assume that items are always available in unlimited quantities. In reality, recommendations are often made under limited availability, where assigning an item to one user may restrict its access for others. This challenge raises questions of fairness, as users may experience unequal satisfaction depending on the timing and allocation of recommendations.
To address this, the SoCRATe framework was developed as a fairness-aware recommendation system that compensates users for such disparities across multiple iterations. This thesis extends the original SoCRATe framework by integrating two modern recommender models Graph Neural Networks (GNNs) and Variational Autoencoders (VAEs) alongside the existing KNN baseline. The goal is to explore how more expressive learning models influence the fairness–utility trade-off in recommendation settings with limited availability.
The study builds on SoCRATe’s iterative workflow, where users receive recommendations, consume items, and accumulate fairness losses that guide compensation in later rounds. The extended framework introduces a fairness-aware GNN, which learns user and item embeddings through graph propagation, and a VAE-based recommender, which captures latent preference distributions for improved generalization under sparse data. Both models are evaluated within SoCRATe’s orchestration logic using preference-driven and round-robin strategies.
Experiments were conducted on synthetic and real-world datasets to compare the performance of KNN, GNN, and VAE models in terms of fairness and accuracy. Results show that while GNNs and VAEs improve overall recommendation quality, their impact on fairness varies depending on data sparsity, model sensitivity, and the chosen allocation strategy. The findings highlight the importance of balancing representation learning with fairness-aware orchestration to achieve equitable outcomes in recommendation systems.
This work contributes to a better understanding of how modern deep learning recommenders interact with fairness-aware frameworks like SoCRATe. It provides both theoretical and practical insights for designing recommendation systems that remain efficient while ensuring fair treatment of users under real-world availability constraints.
To address this, the SoCRATe framework was developed as a fairness-aware recommendation system that compensates users for such disparities across multiple iterations. This thesis extends the original SoCRATe framework by integrating two modern recommender models Graph Neural Networks (GNNs) and Variational Autoencoders (VAEs) alongside the existing KNN baseline. The goal is to explore how more expressive learning models influence the fairness–utility trade-off in recommendation settings with limited availability.
The study builds on SoCRATe’s iterative workflow, where users receive recommendations, consume items, and accumulate fairness losses that guide compensation in later rounds. The extended framework introduces a fairness-aware GNN, which learns user and item embeddings through graph propagation, and a VAE-based recommender, which captures latent preference distributions for improved generalization under sparse data. Both models are evaluated within SoCRATe’s orchestration logic using preference-driven and round-robin strategies.
Experiments were conducted on synthetic and real-world datasets to compare the performance of KNN, GNN, and VAE models in terms of fairness and accuracy. Results show that while GNNs and VAEs improve overall recommendation quality, their impact on fairness varies depending on data sparsity, model sensitivity, and the chosen allocation strategy. The findings highlight the importance of balancing representation learning with fairness-aware orchestration to achieve equitable outcomes in recommendation systems.
This work contributes to a better understanding of how modern deep learning recommenders interact with fairness-aware frameworks like SoCRATe. It provides both theoretical and practical insights for designing recommendation systems that remain efficient while ensuring fair treatment of users under real-world availability constraints.