A novel evaluation framework for recommender systems in big data environments

Recommender systems were first introduced to solve information overload problems in enterprises. Over the last few decades, recommender systems have found applications in several major websites related to e-commerce, music and video streaming, travel and movie sites, social media, and mobile app stores. Several methods have been proposed over the years to build recommender systems. However, very little work has been done in recommender system evaluation metrics. The most common approach to measuring recommender system's performance in offline settings is to employ micro or macro averaged versions of standard machine-learning measures. Profit or other business-oriented metrics have been proposed for other predictive analytics problems, such as churn prediction. However, no such metrics have emerged for the recommender system context. In this work, we propose a novel evaluation metric that incorporates information from the online-platform userbase's behavior. This metric's rationale is that the recommender system ought to improve customers' repeated use of an online platform beyond the baseline level (i.e. in the absence of a recommender system). An empirical application of this novel metric is also presented in a real-world mobile app store, which integrates the dynamics of large-scale big data environments, which are common deployment scenarios for these types of recommender systems. The resulting profit metric is shown to correlate with the existing metrics while also being capable of integrating cost information, thereby providing an additional business benefit context, which allows us to differentiate between two similarly performing models.