Youming Zhang, Ruofei Zhu, Zhengzhou Zhu, Qun Guo, Lei Pang, Peking University, China
The problem of Click-through rate(CTR) prediction is the core issue to many real-world applications such as online advertising and recommendation systems. An effective prediction relies on high-order combinatorial features, which are often hand-crafted by experts. Limited by human experience and high implementation costs, combinatorial features cannot be manually captured thoroughly and comprehensively. There have been efforts in improving hand-crafted features automatically by designing feature-generating models such as FMs, DCN, and so on. Despite the great success of these structures, most of the existing models cannot differentiate the high-quality feature interactions from the huge amount of useless feature interactions, which can easily impair their performance. In this paper, we propose a Higher-Order Attentional Network(HOAN) to select high-quality combinatorial features. HOAN is a hierarchical structure, the multiple crossing layers can learn feature interactions of any order in an end-toend manner. Inside the crossing layer, each interaction item has its unique weight with consideration of global information to eliminate useless features and select high-quality features. Besides, HOAN also maintains the integrity of individual feature embedding and offers interpretive feedback to the calculating process. Furthermore, we combine DNN and HOAN, proposing a Deep & Attentional Crossing Network (DACN) to comprehensively model feature interactions from different perspectives. Experiments on sufficient real-world data show that HOAN and DACN outperform state-of-the-art models.
Click-through rate prediction, Feature interaction networks, Attention mechanism, Hybrid model