Broad Learning for Recommender System

Recommender system (RecSys) has been deployed in most online platforms to assist users in finding their interested items. Though achieving great success and creating tremendous business value, it still suffers from a series of problems such as data sparsity and filter bubble issue. This thesis demonstrates how we can alleviate the two problems and learn a more powerful RecSys with broad learning from different kinds of information. Specifically, five research topics are covered. Three published papers are presented to illustrate how we can apply broad learning by directly modeling the user’s social graph and item’s context information within the recommendation model. Then one published paper is illustrated to improve recommendation diversity with external item category information. And finally a multi-task training algorithm to embed external information into user/item representation is further demonstrated. All experiments are conducted on public datasets (Ciao, Epinion, Steam, Amazon). The experiments do not involve human subjects, animals, or recombinant DNA. Therefore, approvals from the Institutional Review Board, the Animal Care Committee, or the Institutional Biosafety Committee are not needed.