Large Scale Pathways Recommender Systems (PaRS) at Verizon

Millions of customers visited the Verizon Website and stores monthly to shop, from plans to products and accessories. Numerous task-specific recommender systems have been built at Verizon to enhance customer experience. Motivated by the latest development in deep learning, multi-task learning architectures can better understand customer behavior therefore greatly improve model performance, and also save time to learn new tasks. However, it's challenging to build such a system to: 1) digest heterogeneous data from purchases, online clicks, store visits to customer services and many more; and 2) accomplish various tasks where some of them might contradict with each other; and 3) efficiently train and tune large DL models at scale . At Verizon, we greatly benefited from the Ray ecosystem and built a novel multi-task learning recommender system, namely Pathways Recommender system (PaRS), which directly handles multiple abstract forms of data and is able to generalize across multiple recommendation tasks at Verizon. Last but not least, we explored built-in bias mitigation in PaRS to promote diverse, relevant and fair recommendations.