Research On Playing Prediction Of Short-form Video Based On Data Mining Method

In recent years,with the popularization of smart devices and the development of information technology,short video,as a new form of communication,complies with the fragmented and decentralized communication characteristics of the mobile Internet,and gradually enters the public’s field of vision.It is not only rich in content types,but also has strong social attributes,which can meet the diverse content and social needs of netizens,and become the main tool for people’s daily leisure and entertainment,social interaction,and information sharing.With the continuous enrichment and dissemination of content,the number of short video users continues to increase,the market scale continues to expand,and the number of video resources is also growing explosively.In the face of massive resources,the cost of screening is relatively high.How to select high-quality and user-friendly video resources from the massive video data released by a large number of users has become a challenge and problem.Video platforms with data want to improve the quality of videos distributed to users and recommend the types of videos users like;users also hope to see the content they are interested in.If the click-to-play rate of short videos can be accurately predicted,it can help to solve the above problems.Therefore,the playback prediction of short videos is a very important,challenging and valuable research with real application scenarios.Based on the above background,based on the data mining method and feature engineering technology,based on the video click behavior data of industrial real users,through in-depth analysis of user behavior data,mining user attribute characteristics and video features,etc.,the user’s click-to-play behavior is analyzed predictive research.By building a machine learning model to predict the user’s click-through rate on short videos,it can more accurately recommend the content of interest to the user and improve the user’s click-to-watch behavior on the data set.This article reviews the iterative history of the development of click-through rate prediction,introduces the common methods of data mining technology,and summarizes the relevant theoretical basis,respective advantages and parameter adjustment methods of several latest machine learning models in the industry.Analyze and explore the data set,and visualize the analysis results in the form of charts.Mining and extracting features,including:explicit user-side features,resource-side features,and implicit historical click features,etc.Based on the XGBoost algorithm,LightGBM algorithm,CatBoost algorithm,and DeepFM algorithm,the prediction models were constructed,and the hyperparameters were tuned through Bayesian optimization and parameter tuning in stages to obtain the optimal value of each parameter.Then,the influence of different values of each hyperparameter on the prediction results is analyzed and compared,the main characteristics of each model are analyzed,and finally the prediction results are evaluated and compared.After comparison,the prediction effect F1 score and AUC value of the DeepFM model are the best.Compared with the other three models,the AUC value of DeepFM is increased by 4.38%,3.44%,and 3.93%respectively;the LightGBM model is excellent in terms of prediction efficiency and time consumption.,the time consumption is 1/17 of XGBoost;CatBoost performs best without hyperparameter tuning;and the feature experiments show that the addition of statistical features and second-order cross features has a greater impact on the AUC value.Through research and analysis,constructive suggestions can be given to the short video platform on certain model selection and feature construction and selection in short video prediction and recommendation scenarios,and provide data support and decision support for short video recommendation and short video advertising.