GFEN: Graph Feature Extract Network for Click-Through Rate Prediction

Click-Through Rate (CTR) is a fundamental task in personalized advertising and recommender systems. It is vital to model different orders feature interactions in click-through rate. There are many proposed methods in this field such as FM and its variants. However, many current methods can not adequately and precisely extract feature interactions. In this paper, we improve an effective light weight method called the Graph Feature Extract Network (GFEN) to further explicitly or implicitly model low-order and high-order feature interactions information via Graph Convolutional Network (GCN) and Global Recombination Network (GRN). GCN explicitly models local high-order feature interactions. GRN automatically captures global high-order feature interactions via the diversity pooling layer and recombines global and local feature interactions via fully connection layer. We conduct extensive experiments on there real-world datasets and show that our model achieves the best performance compared to the existing state-of-the-art methods such as CKE, DKN, Wide&Deep, RippleNet etc. Our proposed GFEN is a very light weight model, which can be applied to other complicate model such as Ripplenet based knowledge graph and deep learning models based CTR. The whole model can be efficiently fit on large-scale raw input feature.