Öneri sistemlerine yapılan profil enjeksiyon ataklarının popülerlik yanlılığı sorunu üzerindeki etkilerinin incelenmesi

Digital platforms, in their quest to engage users with personalized recommendations, face an unforeseen challenge: popularity bias. Recommendation algorithms often favour popular content, limiting users' access to new and niche items, thereby reducing content diversity on these platforms. This thesis proposes an innovative approach to reduce popularity bias in recommendation systems, enhancing users' access to a broader range of content. The unique aspect of this study is its strategic use of profile injection attack models, traditionally seen as security risks, to mitigate popularity bias. Attack models such as Average, Bandwagon, and Random Attack, typically used to manipulate systems, are here repurposed to restore diversity and balance in recommendations. In particular, the Random Attack model emerges as a strong alternative, promoting less popular content to achieve a fairer and more diverse recommendation system. Additionally, traditional debiasing techniques, such as XQUAD, ERPAug, ERPMul, and VAR, are analysed and compared against attack models for their effectiveness in reducing popularity bias. This study is conducted using the widely adopted MovieLens dataset and employs the VAECF (Variational Autoencoder for Collaborative Filtering) algorithm. The experiments assess each method's impact on popularity bias and recommendation quality through core performance metrics such as APLT, F1-Score, nDCG, and Novelty. Results indicate that XQUAD is the most effective method for reducing popularity bias, while Random Attack (Attack Size 10, Filler Size 10) provides a balanced performance, preserving both recommendation accuracy and diversity. On the other hand, ERPMul and VAR offer high diversity but may compromise some recommendation accuracy. These findings serve as practical insights for digital platforms aiming to provide fairer, more diverse, and user-centric recommendations. In conclusion, this thesis presents an innovative perspective on mitigating popularity bias in recommendation systems, contributing to strategies that enhance fairness and diversity. The findings can serve as a foundation for future studies, which may broaden this approach by exploring various datasets and algorithms, ultimately advancing solutions for fair and diverse content delivery in recommendation systems.