Abstract: Aiming at the problem that the classic click prediction model cannot capture the users dynamic interest and analyze the characteristics of low-level and high-level interaction, this paper proposes a video click prediction model based on the users dynamic interest. The model first maps the discrete data into low-dimensional continuous vectors that are easy to operate after embedding. In order to capture the users dynamic interest changes, the transformer model is introduced to analyze the video sequence clicked by users and the candidate video to be predicted, and the interaction between the video in the behavior sequence and the video to be recommended is extracted. In order to dig deeper into the implicit feature interaction behind the users click behavior, the DeepFM network is introduced and the network structure is optimized and improved to make the model more suitable for sequence-dependent click data. The experimental results show that the prediction accuracy extension of the model proposed and improved in this paper is better than that of the typical deep decomposition model in click rate prediction, and the release of the transformer mechanism can significantly improve the accuracy of click rate prediction.

Key words: click-through rate prediction, user interest, feature interaction, factorization machine, transformer