A Novel Ensemble Strategy Combining Gradient Boosted Decision Trees and Factorization Machine Based Neural Network for Clicks Prediction

Abstract

Click-through rate (CTR) prediction is generally formulated as a supervised classification problem. One challenge in CTR prediction, especially the features with high-sparsity, is to exploit the potential generalization ability under the given samples. In this paper, we first present a novel Factorization Machine (FM) based Neural Network (FNN), which helps capture the nonlinear interactions between sparse inputs. And then, the gradient boosted decision trees (GBDT) model is combined with FNN via cascading and boosting (i.e., GBDT2FNN, GBDT+FNN) respectively to improve the CTR predictive accuracy. To illustrate the performance, we employ them on the open dataset, JData. The experiment results show that the proposed ensembles significantly increase AUC and RIG compared with the baseline GBDT2LR.