Web-Scale Personalized Real-Time Recommender System on Suumo

In this paper we investigate the performance of machine learning based recommender system with real-time log streaming on a large real-estate site, in the views of system robustness, business productivity and algorithm performance. Our proposed recommender system, providing personalized contents as opposed to item/query based recommendation, consists of a real-time log processor, auto-scaling recommender API and machine learning modules. System is carefully designed to let data scientists focus on improving core algorithms and features (instead of taking care of distributing systems) and achieves weekly release cycle in production environment. On Suumo, the largest real-estate portal site in Japan, the system returns more than 99.9% of the API calls successfully in real-time and shows finally a 250% improvement of conversion rate compared to the existing recommendation. With its flexible nature, we would also expect the system to be applied in various kinds of real-time recommendation in the near future.