A1454
Title: Pairwise personalized learning in recommendation system
Authors: Haowen Zhou - University of Virginia (United States) [presenting]
Xiwei Tang - University of Texas at Dallas (United States)
James Lee - University of Virginia (United States)
Abstract: Recommendation systems are essential across domains such as content platforms and e-commerce, yet standard approaches based solely on explicit ratings often face challenges due to data sparsity and heterogeneity. The aim is to introduce a hybrid pairwise personalized learning (HPPL) model that extends the conventional Bayesian personalized ranking method to handle a mixture of explicit and implicit feedback in recommendation tasks. Unlike traditional models that treat observed data as weakly positive and missing data as weakly negative signals, HPPL leverages explicit ratings in a weighted pairwise loss function with latent factors, achieving robust performance in ranking-based evaluations while maintaining computational efficiency. The model bridges the gap between explicit and implicit feedback systems, providing theoretical guarantees for scalable and adaptive stochastic gradient algorithms and offering a practical solution for real-world recommendation systems.
