An Attention-based Deep Network for CTR Prediction

Click-through rate (CTR) prediction is a crucial topic in online advertising system. Early researchers proposed numerous shallow models to analyze this issue, such as logistic regression, factorization machines and Gradient boosting decision tree. In order to improve the model performance furthermore, researchers propose some deep models, such as the factorization-machine supported neural networks, Wide&Deep, DeepFM. Normally, user's historical behavior data contains abundant feature information, but above models lack of modeling and analysis of historical data. To address this problem, the paper proposes a deep CTR prediction model based on attention mechanism and GRU model, which can make use of the users' historical behaviors. The experimental results demonstrate that compared with other popular models, our proposed model can improve the prediction performance by extracting the implied interest features from user historical behaviors.