Research On Vectorized Content Recall Algorithm Of Video Recommendation System

With the rapid development of the mobile Internet era,our country’s mainstream video network companies continue to grow,and more companies use the Internet,social platforms,e-commerce and other channels to enhance brand value,optimize corporate profit models,and meet consumer needs.People are gradually entering the era of information overload from the era of lack of information.In order to provide accurate information flow to the vast information consumers,the recommendation system has gradually become the core key technology in the era of mobile Internet.With the rapid increase in the number of video streams and the number of video stream users in the network,the video recommendation system often encounters a cold start problem during the accumulation of the number.This article will alleviate the cold-start problem in recommendation recall based on the user’s historical behavior on the video,provide users with personalized recommendation services,and then optimize the recall algorithm in the video recommendation system.The recommendation system usually consists of two stages: recall and sorting.The recall stage refers to the process of selecting some of the resources that users are interested in from the massive resource pool as the candidate set.The sorting stage is to sort the candidate set according to specific optimization goals.The content of the recall determines the upper limit of the accuracy of the sorting stage.In order to improve the ability of the recall model to extract user behavioral interest,this paper has carried out the following optimization and research work on the algorithm of the recall phase.First of all,a user model is proposed for the user’s historical behavior on the video.The user model will collect the user’s historical click records and search records,and combine these users’ historical behavior with the rest of the user’s own characteristics through a deep neural network.It is encoded into a user vector that can match the video vector;secondly,a content model is proposed based on the pre-training model of Bidirectional Encoder Representations from Transformers,which encodes the video text information into a content vector with semantic information.Finally,based on the idea of a deep semantic matching model,a recall model with a double-tower structure for vector matching is proposed to train user vectors and content vectors.By analyzing the advantages and disadvantages of the two negative sample construction strategies of exposure unclicked and random sampling,a suitable negative sample construction strategy is selected.Finally,a series of model comparison experiments are designed on the two video data sets to verify the effectiveness of the model proposed in this paper in vectorized content recall based on user behavior,and can effectively alleviate the cold start problem faced in the early stage of recommendation recall.