Similitude Attentive Relation Network for Click-Through Rate Prediction

In online advertising systems, having a good knowledge of user behavior is crucial for click-through rate (CTR) prediction. In recent years, many researchers turn to seek a better way of user representation by modeling the behavior sequences with recurrent neural network (RNN). However, recurrent layers implicitly adopt the assumption that elements with different orders are fundamentally different, which is inefficient in many practical scenarios with much uncertainty and complicated hidden states. In this paper, we follow the paradigm of Relation Network (RN), and propose a new model called Similitude Attentive Relation Network (SARN). The user behavior is modeled as a graph, where nodes correspond to the visited items and edges correspond to the relations. To capture the latent user interest better, the model concentrates on the relations between items, rather than the translation on the time series. More specifically, the model tries to learn the similarity between items in a semantic space through a learnable dot-product operation and blend both of the item representations and relational information together as the final relations. We define our user representation on an attentive pooling of the relations directly. To verify the effectiveness of our method, extensive experiments on two public datasets and one real-world online advertising dataset are conducted. Experimental results show that our methods achieve usually better performance than others. Besides, we explore the properties of our model by controlled experiments and show the learned relational knowledge by visualizing the inner states of SARN.