LLM-Driven Fairness-Aware Recommender System
Masum, Md Abdulla Al (2026)
2026
Matematiikan ja tilastollisen data-analyysin maisteriohjelma - Master's Programme in Mathematics and Statistical Data Analytics
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-06-05
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606057010
https://urn.fi/URN:NBN:fi:tuni-202606057010
Tiivistelmä
Recommender systems are expected to balance the interests of multiple stakeholders — consumers, items, and providers — rather than optimising for accuracy alone. Existing multistakeholder fairness frameworks address this challenge through mathematical optimization with carefully tuned penalty terms. While effective, such approaches are rigid, require parameter calibration, and struggle to balance competing fairness objectives flexibly. This thesis investigates how Large Language Models (LLMs) can perform as a ranking policy that is capable of reasoning about multiple fairness objectives.
This thesis proposes LLMFARS (LLM-Driven Fairness-Aware Recommender System), a fairness-aware recommendation framework in which a base recommender (VAECF) generates candidate items and an LLM-driven ranking policy selects the final top-K items by reasoning over user history, candidate metadata, structured fairness signals, and structured prompt engineering. The framework defines eight operational modes covering all combinations of consumer, item, and provider fairness. LLMFARS is evaluated against the CIPFRS baseline on the Amazon Software dataset using two LLMs of comparable scale, Qwen 3.5 9B and Gemma 4e 4B, across six metrics directly comparable to CIPFRS and two supplementary accuracy metrics.
The results suggest that LLM-based reasoning is substantially better than mathematical optimization for provider fairness and catalog coverage, comparable for consumer fairness and accuracy, and worse for item fairness and novelty. Combined fairness modes provide graceful trade-offs and occasional synergies between objectives, including a notable repair of single-mode failures in consumer fairness. The two LLMs agree on the direction of the findings, but systematically differ on magnitudes. This suggests that the choice of LLM is itself a fairness parameter, not an implementation detail. In our thesis, we propose a new LLM-based fairness architecture, provide empirical insights into the trade-offs between optimization-based and LLM-based fairness, and offer practical guidance for deploying LLMs as ranking policies in multi-stakeholder recommender systems.
This thesis proposes LLMFARS (LLM-Driven Fairness-Aware Recommender System), a fairness-aware recommendation framework in which a base recommender (VAECF) generates candidate items and an LLM-driven ranking policy selects the final top-K items by reasoning over user history, candidate metadata, structured fairness signals, and structured prompt engineering. The framework defines eight operational modes covering all combinations of consumer, item, and provider fairness. LLMFARS is evaluated against the CIPFRS baseline on the Amazon Software dataset using two LLMs of comparable scale, Qwen 3.5 9B and Gemma 4e 4B, across six metrics directly comparable to CIPFRS and two supplementary accuracy metrics.
The results suggest that LLM-based reasoning is substantially better than mathematical optimization for provider fairness and catalog coverage, comparable for consumer fairness and accuracy, and worse for item fairness and novelty. Combined fairness modes provide graceful trade-offs and occasional synergies between objectives, including a notable repair of single-mode failures in consumer fairness. The two LLMs agree on the direction of the findings, but systematically differ on magnitudes. This suggests that the choice of LLM is itself a fairness parameter, not an implementation detail. In our thesis, we propose a new LLM-based fairness architecture, provide empirical insights into the trade-offs between optimization-based and LLM-based fairness, and offer practical guidance for deploying LLMs as ranking policies in multi-stakeholder recommender systems.