A Novel Ensemble Approach for Click-Through Rate Prediction Based on Factorization Machines and Gradient Boosting Decision Trees

Abstract

Click-Through Rate (CTR) prediction is a significant technique in the field of computational advertising, its accuracy directly affects companies profits and user experience. Achieving great ability of generalization by learning complicated feature interactions behind user behaviors is critical in improving CTR for recommender systems. Factorization Machines (FM) is a hot recommender method for efficiently modeling features’ second-order interactions. Nevertheless, FM cannot capture the nonlinear and complex modes implied in the real-world data while it models feature in a linear way and just uses the second-order feature interactions. In this paper, we propose a model named GFM, which is an ensemble learning of FM and Gradient Boosting Decision Trees (GBDT) for recommendations. We use FM to model linear features and second-order feature interactions and use GBDT to model the side information for transforming the raw features to cross-combined features. In addition, we import the attention mechanism to calculate users’ latent attention on different features. To illustrate the performance of GFM, we conduct experiments on two real-world datasets, including a movie dataset and a music dataset, the results show that our model is effective in providing accurate recommendations.