Details

Autor(en) / Beteiligte

• Venugopal, Shreya

Titel

User Segmentation in Recommender Systems

Ort / Verlag

ProQuest Dissertations & Theses

Erscheinungsjahr

2024

Link zum Volltext

Quelle

ProQuest Dissertations & Theses A&I

Beschreibungen/Notizen

• Personalized recommendations are key for recommendation systems, which make profits by tailoring marketing strategies and targeting product offerings to distinct user groups. However, current personalized recommendation algorithms do not consider structures in user groups thereby producing non-robust recommendations. Our key insight is to integrate the previously ignored local structure of users into the recommendation algorithm, where we perform a structured user segmentation that considers the hierarchy in user structure. This improves the quality of the recommendations since they enhance their capability to decipher the essence of the user preferences over the noise. Capitalizing on the inherent hierarchical structure of user segments, our method reduces the model size and results in improved accuracy. We conduct experiments using four different approaches for computing user segments and evaluate their performance across various hyperparameter configurations. As far as we know, this is the first comprehensive evaluation on understanding the right user structure to employ for recommendations. Our results demonstrate that our method yields significant improvements in performance metrics across three diverse datasets. The improvement ranges from 9% to 13% across 5 well-known metrics for a large-scale dataset and around 1-2% on two other small datasets. It also gives a significant improvement of around 20% for Relevance to Other Users (ROU) metric that captures the proportion of similar users who have liked an item being recommended to the user.