Fairness-Aware Reranking and LLM-Assisted Explanations for Recommender Systems
Peteti, Mukesh (2026)
2026
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-07-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202607148346
https://urn.fi/URN:NBN:fi:tuni-202607148346
Tiivistelmä
Recommender systems are widely used to support users in finding suitable items from large digital collections. Although collaborative filtering methods are effective for personalization, they often give repeated visibility to items that already have many interactions. As a result, less popular movies may receive limited exposure even when they are relevant to user preferences. This creates problems related to popularity bias, catalog use, and transparency, especially when users are not given clear reasons for the recommendations they receive.
This thesis implements and evaluates a hybrid recommender framework on the MovieLens dataset. The framework uses the Smart Adaptive Recommendations (SAR) algorithm to produce candidate movie recommendations, followed by a fairness-aware reranking step based on item popularity and novelty. A lightweight Large Language Model-assisted explanation component is added after reranking to describe the final recommendations in natural language. The LLM component is therefore used for explanation support rather than as a replacement for the SAR recommendation model.
The results show an accuracy–fairness trade-off. The debiased model reduces MAP@10 from 0.2447 to 0.1865, but improves fairness by increasing catalog coverage from 0.1338 to 0.1914 and reducing average popularity from 235.17 to 165.89. The long-tail ratio also increases from 0.0030 to 0.0259. Diversity results are mixed, while the explanation layer improves recommendation transparency.
The findings indicate that post-processing reranking can make SAR-based recommendations less dependent on highly popular movies, although this improvement requires a measurable reduction in ranking accuracy. The study therefore demonstrates a practical hybrid design in which collaborative filtering remains responsible for recommendation generation, while reranking improves exposure fairness and the LLM-assisted component supports interpretability.
This thesis implements and evaluates a hybrid recommender framework on the MovieLens dataset. The framework uses the Smart Adaptive Recommendations (SAR) algorithm to produce candidate movie recommendations, followed by a fairness-aware reranking step based on item popularity and novelty. A lightweight Large Language Model-assisted explanation component is added after reranking to describe the final recommendations in natural language. The LLM component is therefore used for explanation support rather than as a replacement for the SAR recommendation model.
The results show an accuracy–fairness trade-off. The debiased model reduces MAP@10 from 0.2447 to 0.1865, but improves fairness by increasing catalog coverage from 0.1338 to 0.1914 and reducing average popularity from 235.17 to 165.89. The long-tail ratio also increases from 0.0030 to 0.0259. Diversity results are mixed, while the explanation layer improves recommendation transparency.
The findings indicate that post-processing reranking can make SAR-based recommendations less dependent on highly popular movies, although this improvement requires a measurable reduction in ranking accuracy. The study therefore demonstrates a practical hybrid design in which collaborative filtering remains responsible for recommendation generation, while reranking improves exposure fairness and the LLM-assisted component supports interpretability.