F2DeepRS: A Deep Recommendation Framework Applied to ICRC Platforms

The applications of recommendation system (RS) are ubiquitous in our daily lives. Many RSs have exhibited excellent performances, especially those based on neural networks in recent years. However, many existing methods gained superiority in performance at the expense of hardware or computational costs, which are not applicable to applications that run on information and computation resource constrained (ICRC) platforms. The ICRC applications, such as in-vehicle infotainment system where there is no powerful computer resources, pose new challenges for developing a RS, including uncertainty in the availability of user profile or item content, less powerful computation resources but huge size of user and item pools, etc. With a focus on applications in such context, we developed a Fast and Flexible Deep Recommendation System (F2DeepRS) framework. In this framework, existing user/item knowledge learned by other even non-neural models can be leveraged to make the learning process computationally efficient. And the framework can be flexibly configured to incorporate user/item content information depending on availability. Experiments were conducted using R2-Yahoo! Music dataset, which is, to our best knowledge, the largest public dataset publicly accessible to our knowledge. The performance is compared with both traditional and state-of-the-art RS. The results have indicated that the proposed F2DeepRS framework is competitive in preference prediction tasks, and particularly, is very computationally efficient.