Click-Through Rate Prediction Models based on Interest Modeling

Deep learning frameworks have achieved extraordinary results in popular technological frontiers, which also greatly inspires considerable researchers in recommender system. As one of the mainstream research frontiers, click-through rate (CTR) predictions have received increasing attention from both industrial and academic circles. Recently, Alibaba group has come up with novel idea of interest modeling that can effectively overcome the limitations brought by previous deep learning CTR prediction models, which is also the key to improve performance. For the past few years, more and more relevant works begin to explore approaches taking advantage of the learning, representation, and interpretation ability from neural networks. However, there are few comprehensive surveys targeted at this topic. In this paper, we systematically review and summarize recent CTR prediction models based on interest modeling. First, we elaborate their different design ideas and pay additional attention to their intrinsic relationship including improvements and extensions, similarities and differences, advantages and disadvantages and so on. Next, we deploy them separately according to official open-source code and compare their performance assessment in terms of accuracy and speed, looking for best optimizers and other hyperparameters. Furthermore, we briefly discuss some possible refinement directions and future research trends.