| High-speed development of the mobile internet service is benefited from 5G technology.As a new type of social media,short videos have the characteristi cs of short duration,rich content,strong entertainment,easy to share,and quickly integrated into people’s daily life.While enriching people’s daily life,it also changes people’s life patterns.However,with the increasing number of customers,the scope of short videos and markets,it has caused a very serious problem of information overload,which makes people unable to quickly obtain their suitable videos from massive resources.Therefore,there is an urgent need for a system that can make personalized video recommendations.CTR prediction is an extremely important issue in the recommendation system,and it is determined by the size of the CTR to determine which videos need to be recommended.Deep learning technology is advancing rapidly,and some scholars have applied it to short video click prediction models in recent years.But it ignores the fact that users’ interests may change over time,fails to consider the characteristics of multi-dimensional characteristics of short videos,and ignores the correlation between multiple click behaviors,so that the accuracy of click rate prediction is not high.Therefore,aiming at these problems,this paper has carried out a series of studies.The main works are embodied in the following three aspects:1.Sequence feature model based on user behavior.In view of the fact that the traditional click prediction model negletcs the fact that the user’s interests may change with time,a sequence feature model based on user behavior is proposed,which divides user behavior according to time,uses negative sampling method to screen high-quality behavior data,and uses embedding modules(word embedding model Word2 Vec and graph embedding model)to process the user’s historical behavior sequence and extract a series of sequence feature information,thereby improving the accuracy of click prediction.2.Multi-behavior click prediction model for short videos based on user behavior sequence.Since the existing click prediction models do not consider the influence of multi-dimensional features of short video on the accuracy of behavioral click prediction,a short video multi-behavioral click prediction model based on user behavior sequence is proposed,and an embedded module and Light GBM module are introduced to construct multi-dimensional features of short videos on the basis of the Deep FM model to improve the accuracy of click prediction.3.Short video multi-behavior click prediction model based on multi-task learning.Aiming at the fact that the multi-task learning model ignores the intrinsic relationship and difference between different behaviors and will affect the prediction accuracy of the behavior when predicting multi-behavior clicks,a short video multi-behavior click prediction model based on multi-task learning is proposed,and the embedded module is introduced to extract sequence features in the user’s historical behavior,and the FM model is introduced on the basis of the MMOE model,which can not only learn the connections and differences of different behaviors,but also learn the interaction of low-order features,so that the model has the memory performance of low-order features and the generalization performance of high-order features.Improve click prediction accuracy. |
