Design And Application Of Video Recommendation Algorithm Based On Posterior Index

With the development of society,short video has gradually become an indispensable part of people's life,and even has developed to the stage of national creation.Therefore,the amount of video data is increasing day by day,which makes it difficult for users to choose video.This video platform uses the recommendation algorithm to select specific videos for users.In the recommendation algorithm,the recall phase is mainly responsible for the rough sorting of massive data,obtaining a certain number of video candidate sets,which directly has a decisive impact on the final video recommendation list.In this context,this paper designs a reordering process after the recall phase by using posterior features to improve the sorting effect of recall phase.In this reordering process,the modeling idea in You Tube DNN paper is used for reference,and the average playing time of users is taken as the optimization objective to increase the stickiness of users.The reordering process based on posterior index mainly includes user log analysis,feature engineering construction,model training,online reasoning,self feedback optimization mechanism and final A / B test.In the user log analysis process,the physical meaning of the field corresponding to each user's behavior is considered in detail,and the posterior indicators such as CTR(video click through rate),PPI(average display duration of video)are obtained by calculation formula.In the construction of Feature Engineering,the posterior index is used to calculate the corresponding posterior data with time interval as a parameter,and the posterior features with positive feedback are constructed by aggregating the posterior data.After the feature construction is completed,in order to train the model,the posterior index is used to construct the tag for the video data,and the rules and physical meaning of the tag construction are explained.After the construction of posterior features and video data tags,a training process based on the gradient lifting decision tree model(gbdt)is designed,and the posterior features calculated before are fed to the model for training,and then a score prediction model for inverted data is obtained.After the model predicts the inverted data from recall,the prediction score and similarity score are comprehensively considered by online reasoning to get the total score of video,and the self feedback optimization mechanism of reordering algorithm is established to retrain the model by adjusting the weight.The processed inverted data is submitted to the engine for truncation and sorting.Finally,the A / B test is used to verify the experimental results and start the self feedback optimization mechanism.From the analysis of the research results,the algorithm based on a posteriori index can reorder the inverted data from recall in the context of short video recommendation.More meaningful posterior features are considered in the reordering process,which improves the video sorting effect and improves the average display duration of users.It reduces the process complexity of the subsequent fine scheduling stage and improves the efficiency of the overall video recommendation.