Adaptive Noise Injection in Variational Autoencoders for Enhancing Fairness in Group Recommendations
Ahmad, Emaz Uddin (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-10-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025102310082
https://urn.fi/URN:NBN:fi:tuni-2025102310082
Tiivistelmä
Recommender systems play a vital role in digital platforms by personalizing user experiences across different domains such as e-commerce, media, and social networks. However, the major challenge faced by recommender systems is to ensure the fairness and diversity especially in group settings due to diverse user preferences and biases inherent in interaction data.
This thesis proposes an enhanced Variational Autoencoder (VAE) based framework that introduces adaptive noise injection into the latent space to promote fairness and satisfaction in group recommendations. In contrast to conventional VAEs that depend on static Gaussian noise the proposed model dynamically learns data-dependent, adaptive noise from user representations, enhancing its ability to predict uncertainty and reducing bias towards dominant user preferences. The framework is further enhanced through Bayesian optimization, which is employed to fine-tune the hyperparameters of the Variational Autoencoder.
Comprehensive experiments on the MovieLens 10M dataset across homogeneous, heterogeneous, and mixed groups demonstrate that the adaptive noise VAE model consistently outperforms the static noise baseline, achieving higher satisfaction, improved ranking quality (NDCG and Recall), and reduced unfairness (DFH). The proposed approach contributes a novel viewpoint on fairness regulation in deep generative recommender systems and establishes a foundation for more equal, diversified, and reliable group recommendations.
This thesis proposes an enhanced Variational Autoencoder (VAE) based framework that introduces adaptive noise injection into the latent space to promote fairness and satisfaction in group recommendations. In contrast to conventional VAEs that depend on static Gaussian noise the proposed model dynamically learns data-dependent, adaptive noise from user representations, enhancing its ability to predict uncertainty and reducing bias towards dominant user preferences. The framework is further enhanced through Bayesian optimization, which is employed to fine-tune the hyperparameters of the Variational Autoencoder.
Comprehensive experiments on the MovieLens 10M dataset across homogeneous, heterogeneous, and mixed groups demonstrate that the adaptive noise VAE model consistently outperforms the static noise baseline, achieving higher satisfaction, improved ranking quality (NDCG and Recall), and reduced unfairness (DFH). The proposed approach contributes a novel viewpoint on fairness regulation in deep generative recommender systems and establishes a foundation for more equal, diversified, and reliable group recommendations.