Exploiting Intra and Inter-field Feature Interaction with Self-Attentive Network for CTR Prediction

Click-Through Rate (CTR) prediction models have achieved huge success mainly due to the ability to model arbitrary-order feature interactions. Recently, Self-Attention Network (SAN) has achieved significant success in CTR prediction. However, most of the existing SAN-based methods directly perform feature interaction operations on raw features. We argue that such operations, which ignore the intra-field information and inter-field affinity, are not designed to model richer feature interactions. In this paper, we propose an Intra and Inter-field Self-Attentive Network (IISAN) model for CTR prediction. Specifically, we first design an effective embedding block named Gated Fusion Layer (GFL) to refine raw features. Then, we utilize self-attention to model the feature interactions for each item to form meaningful high-order features via a multi-head attention mechanism. Next, we use the attention mechanism to aggregate all interactive embeddings. Finally, we assign DNNs in the prediction layer to generate the final output. Extensive experiments on three real public datasets show that IISAN achieves better performance than existing state-of-the-art approaches for CTR prediction.