Hybrid Models of Factorization Machines with Neural Networks and Their Ensembles for Click-through Rate Prediction

Prediction of Click-through rate (CTR) is getting considerable attention of online campaign and advertising management systems as well as search and recommender engines. The reason behind these great interest in CTR prediction is: It facilitates exploring the interest of web/mobile users based on their interactions in online environments. With the rapid proliferation of deep learning in recent years, CTR prediction has also expanded its effects in terms of both prediction accuracy and widespread usage in Real-Time Bidding (RTB) services. This study examines the latest status of CTR prediction works particularly in the field of neural network extended models, and also aims to devise hybrid models of Factorization Machines with Neural Networks. Also, ensembles of these hybrid models are studied and how much difference ensemble models might make is presented. In the end, our best ensemble model is implemented on a real digital campaign to show both prediction performance and the financial outcome of the CTR prediction comparatively. Among the proposed implementations in this study, the best performing one is based on the serial ensem-bling of hybrid models via Gradient Boosting Algorithm. It achieves 32.4% reduction in mean-squared error compared to the baseline Matrix Factorization Model, and 76.1% reduction of cost (eCPC) on the campaign on the seventh day of it.