Deep Interest Network for Taobao advertising data Click-Through Rate Prediction

Click-through rate prediction has been widely used in recommendation system, search engine optimization, advertising and other fields. Traditional machine learning methods such as LR, FM, and GBDT require a large amount of artificial feature construction processes and only learn the linear relationship between features. With the rapid development of deep learning in natural language processing, computer vision, many studies such as DeepFM, PNN, and Wide&Deep, apply DNN models to Click-through rate prediction estimation. In this work, I use the modified deep interest network (DIN) for the click-through rate prediction task. The mixed loss is proposed as the training loss to fit the DIN model. Attention mechanism is improved according to different input features. I evaluate proposed deep interest network model on Taobao ad click data. Results show the model has superior performance than other deep learning methods. Experiments show that the modified DIN model is effective for the Taobao advertising data click-through rate prediction task.