Research‎ > ‎

Recommender Systems


1. Multi-dimensional recommendations using tensor factorization on sparse user generated data

Mentors: Thomas Phan and Jiayu Zhou (Samsung Research America - Silicon Valley)


Given the abundance of online information available to mobile users, particularly tourists and weekend travelers, recommender systems that effectively filter this information and suggest interesting participatory opportunities will become increasingly important. Previous work has explored recommending interesting locations; however, users would also benefit from recommendations for activities in which to participate at those locations along with suitable times and days. Thus, systems that provide collaborative recommendations involving multiple dimensions such as location, activities and time would enhance the overall experience of users.The relationship among these dimensions can be modeled by higher order matrices called tensors which are then solved by tensor factorization. However, these tensors can be extremely sparse. In this paper, we present a system and an approach for performing multidimensional collaborative recommendations for Who (User), What
(Activity), When (Time) and Where (Location), using tensor factorization on sparse user-generated data. We formulate an objective function which simultaneously factorizes coupled tensors and matrices constructed from heterogeneous data sources. We evaluate our system and approach on large-scale real world data sets consisting of 588,000 Flickr photos collected from three major metro regions in USA. We compare our approach with several state-of-the- art baselines and demonstrate that it outperforms all of them.

Related publications:

a. Preeti BhargavaThomas Phan, Jiayu Zhou, Juhan Lee, Who, What, When, and Where: Multi-Dimensional Collaborative Recommendations using Tensor Factorization on Sparse User Generated Data, Proceedings of the 24th International World Wide Web Conference (WWW 2015) [PDF]