Abstract

A personalized product recommendation can significantly help customers find their preferred items and assist business owners in obtaining more income. However, due to the complexity of the users’ decision-making process, which is influenced by various aspects in different item domains such as functional complementarity and visual compatibility, item recommendations usually suffer from many challenging problems. Although several visually-aware recommendation methods have been proposed, most do not simultaneously incorporate both the influence of functional complementarity and the effect of visual compatibility into latent factor models. To address this issue, we first build an item-complementarity network using “frequently-bought-together” item information. Then, a Visual and Relational Probabilistic Matrix Factorization (VRPMF) model is proposed, which models a user’s preference for a given item as the combination of visual contents and item-complementarity relationships. Due to the dynamic nature of online data, such a comprehensive model poses efficiency challenge in learning parameters of the model. To overcome this scalability issue, we present a novel and Fast Alternating Least Squares (FALS) algorithm, to efficiently optimize the proposed model. Finally, to evaluate the VRPMF method, we conduct comprehensive experiments with several state-of-the-art competitors and evaluation metrics on multiple real-world datasets. The empirical results show that our method achieves significant improvements in terms of both rating prediction accuracy and running time. Our implementation of VRPMF is publicly available at: https://github.com/wubin7019088/VRPMF

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call