The paper introduces the Deep Interest Evolution Network (DIEN), a novel model for click-through rate (CTR) prediction, which aims to capture and model the dynamic nature of user interests. DIEN addresses the limitations of existing CTR models by focusing on the representation of latent user interests and their evolution over time. The model consists of two main components: the Interest Extractor Layer and the Interest Evolving Layer. The Interest Extractor Layer uses a GRU to capture temporal interests from historical behavior sequences, incorporating an auxiliary loss to supervise the learning of interest states. The Interest Evolving Layer employs a GRU with an attentional update gate (AUGRU) to model the evolving process of interests relative to the target item, enhancing the representation of relative interests and reducing the impact of interest drifting. Experimental results on both public and industrial datasets demonstrate that DIEN significantly outperforms state-of-the-art methods, achieving a 20.7% improvement in CTR on the Taobao display advertisement system. The paper also discusses the online serving techniques used to handle high traffic and latency requirements in commercial systems.The paper introduces the Deep Interest Evolution Network (DIEN), a novel model for click-through rate (CTR) prediction, which aims to capture and model the dynamic nature of user interests. DIEN addresses the limitations of existing CTR models by focusing on the representation of latent user interests and their evolution over time. The model consists of two main components: the Interest Extractor Layer and the Interest Evolving Layer. The Interest Extractor Layer uses a GRU to capture temporal interests from historical behavior sequences, incorporating an auxiliary loss to supervise the learning of interest states. The Interest Evolving Layer employs a GRU with an attentional update gate (AUGRU) to model the evolving process of interests relative to the target item, enhancing the representation of relative interests and reducing the impact of interest drifting. Experimental results on both public and industrial datasets demonstrate that DIEN significantly outperforms state-of-the-art methods, achieving a 20.7% improvement in CTR on the Taobao display advertisement system. The paper also discusses the online serving techniques used to handle high traffic and latency requirements in commercial systems.