Research On Feature Selection Of Weibo Users' Forwarding Behavior Prediction Features

Weibo users' forwarding prediction research has important academic value for the research of social network's information dissemination and recommendation.It has important application value in public opinion management,personalized recommendation and marketing,etc.This paper mainly studies the important influencing factors of ordinary users' forwarding behavior in Weibo.The purpose is to better understand the forwarding behavior of users and improve the effect of forwarding prediction.This paper analyzes the influencing factors of Weibo users' forwarding behavior,synthesizes related literatures,and summarizes many features that affect users' forwarding behavior.Then on the large-scale real Weibo datasets,feature extraction and feature selection are performed,a Factorization Machines prediction model is established,and the forwarding behavior of users in the test set is predicted.Based on the prediction results,this paper compares the effectiveness of various features and feature combinations in Weibo user's forwarding behavior prediction.The work of this paper has four main aspects: 1.Based on the comprehensive literatures,it analyzes and summarizes a large number of features affecting users' forwarding behavior,and implements feature extraction on the real Sina Weibo dataset.2.Through using a set of type features training to establish a Factorization Machines prediction model,it studies the impact on users' forwarding behavior of five different types of features which include user features,author features,Weibo features,interest features and social features.Experiments show that interest features and Weibo features have the greatest impact on model's prediction performance.3.Filter feature selection and Wrapper feature selection are implemented on the complete set of Weibo features.The effects of various features and feature subsets on the prediction performance of the model are studied.Experiments show that the forwarding similarity feature has the highest correlation with the classification prediction.The optimal feature subset selected by the Wrapper method greatly reduces the feature dimension and enhances the operational efficiency while almost guaranteeing the prediction effect.4.Using the Factorization Machines prediction model established by the optimal feature subset with the best prediction performance to predict the user's forwarding behavior,the prediction accuracy reaches 89.0%,the F1 metric reaches 66.8%,and the AUC area reaches 95.0%.