R-RNN: Extracting User Recent Behavior Sequence for Click-Through Rate Prediction

Abstract

Click-through rate (CTR) prediction is of great importance in such web applications as recommender systems, web search, and online advertising. Exploring the interest of a user through feature interactions behind massive behaviors of users is essential for CTR prediction. Recently, deep learning-based models have been proposed, which follow a similar embedding and multi-layer perception (MLP) paradigm and achieve great success in both business and research fields. However, existing studies ignore the sequential characteristics of click-through behaviors of users. To overcome this limitation, we propose a novel model named recent recurrent neural network (R-RNN) to adaptively learn the representation of the interest of a user from his or her overall historical click-through behaviors. R-RNN puts an emphasis on recent click-through behaviors via a novel neural network architecture. Compared to the state-of-the-art methods including the deep interest network (DIN) model recently proposed by Alibaba, R-RNN not only applies the attention mechanism to help capture the representation of the main interest of a user but also incorporates a long short-term memory (LSTM) unit for exploring the trend of the change of the interest of a user behind his or her recent click-through behaviors. Comprehensive experiments are conducted to demonstrate the effectiveness of R-RNN for CTR prediction on a benchmark data set. The results show that R-RNN significantly outperforms existing deep learning models. The results also demonstrate that the length of the recent click-through behavior sequence has an important effect on the prediction performance of the model.