New Performance Metrics for Offline Content-Based TV Recommender System.

The past decade has seen a fast rise in popularity of recommendation systems provided by many entertainment and social media services. However, despite the recognised advances in different recommendation approaches and technologies, there remain many challenges, particularly in TV content recommendation systems. More precisely, machine learning based TV content recommendation systems suffer from a class imbalance problem; hence, it is difficult to evaluate the system using traditional metrics. Moreover, specific challenges arise during the development phase, when the system operates in ‘offline’ mode. This means the recommendations are not actually presented to users - making it even more difficult to measure the quality of those recommendations. This paper presents a proof-of-concept demonstrator of a television recommendation system, based on Content-based Filtering, as a contribution towards building a full-scale intelligent recommendation system. New evaluation metrics are proposed for ‘offline’ testing mode, while also tackling the class imbalance problem. The experimental results, based on real usage data, are promising and help in defining the future path as presented along with the conclusion.