The paper "Deep Interest Network for Click-Through Rate Prediction" addresses the challenge of click-through rate (CTR) prediction in online advertising, particularly in e-commerce settings. Traditional deep learning models, which follow an Embedding&MLP paradigm, often struggle with capturing diverse user interests due to their fixed-length representation vectors. To overcome this, the authors propose the Deep Interest Network (DIN), which introduces a local activation unit to adaptively learn the representation of user interests from historical behaviors for a given ad. This allows the representation vector to vary across different ads, enhancing the model's expressive power. The paper also develops two novel techniques: mini-batch aware regularization and a data adaptive activation function (Dice). Mini-batch aware regularization simplifies the computation of regularization by only updating parameters of non-zero features in each mini-batch, making it feasible for large-scale industrial networks. Dice is an extension of the PReLU activation function, which adaptively adjusts the rectified point based on the distribution of input data, improving the model's performance. Experiments on public datasets and an Alibaba production dataset demonstrate the effectiveness of DIN and its training techniques. DIN outperforms state-of-the-art methods, achieving superior CTR prediction accuracy. The model has been successfully deployed in Alibaba's online display advertising system, contributing to significant improvements in business performance.The paper "Deep Interest Network for Click-Through Rate Prediction" addresses the challenge of click-through rate (CTR) prediction in online advertising, particularly in e-commerce settings. Traditional deep learning models, which follow an Embedding&MLP paradigm, often struggle with capturing diverse user interests due to their fixed-length representation vectors. To overcome this, the authors propose the Deep Interest Network (DIN), which introduces a local activation unit to adaptively learn the representation of user interests from historical behaviors for a given ad. This allows the representation vector to vary across different ads, enhancing the model's expressive power. The paper also develops two novel techniques: mini-batch aware regularization and a data adaptive activation function (Dice). Mini-batch aware regularization simplifies the computation of regularization by only updating parameters of non-zero features in each mini-batch, making it feasible for large-scale industrial networks. Dice is an extension of the PReLU activation function, which adaptively adjusts the rectified point based on the distribution of input data, improving the model's performance. Experiments on public datasets and an Alibaba production dataset demonstrate the effectiveness of DIN and its training techniques. DIN outperforms state-of-the-art methods, achieving superior CTR prediction accuracy. The model has been successfully deployed in Alibaba's online display advertising system, contributing to significant improvements in business performance.