The Application of Recommender Systems in a Multi Site, Multi Domain Environment

Abstract

Recommender systems have cemented themselves in the daily experiences of most online users. In this work we will elaborate on the different challenges faced when creating recommendations in the following domains - Online marketplaces: Two sided marketplaces where buyers and sellers can interact and sell items with each other. - Online News: Online news sites where users consume the latest news articles related to current affairs. - Generic Recommendations: Sites which create generic recommendations based on generalised algorithms. We will review how we address these different challenges in Schibsted. Schibsted is an international media company with over 200 million unique users a month, split across 39 countries across the world. Concretely we will review, and compare the primary challenges between the different domains mentioned as well as the commonalities and general lessons we have learnt. For example in a two sided marketplace, it is important that both actors in the interaction are considered when creating recommendations. Constraints such as price sensitivity and geographical location become important when identifying good quality recommendations for our users. Alternatively, in online news we need to consider issues such as freshness and topical relevance when creating recommendations for users, while also striving to ensure we have editorial satisfaction. Finally we can look to generic recommendation solutions where we provide simple recommendation API end points. In this case it is important to ensure good quality recommendations while ensuring a generic enough solution that it can be used in many different scenarios. What makes these challenges particularly interesting is that we approach these different challenges with a holistic view of for improving the overall user experience for our users in Schibsted.