Author:
Ji Gang,Zhu Yuxuan,Niu Yukai,Hu Kai
Abstract
Abstract
In recent years, with the rapid development of the Internet, especially the mobile Internet, social networks have entered the stage of vigorous development and become one of the main sources of information. User-generated contents (UGC) on social platforms can spread information along social networks at an astonishing speed. Existing literature has proposed many prediction methods for the popularity prediction on social networks. This paper presents a classification and establishes a unified evaluation framework of popularity prediction methods for microblogs. More specifically, we divide these mainstream prediction methods into four types: feature based methods, time series methods, collaborative filtering methods and deep learning methods and conduct experiments on the real-world weibo data using these methods to predict. Finally, according to four indicators, including accuracy, efficiency, robustness and bias, we evaluate and compare the methods. Based on the prediction and evaluation results, this paper summarizes and draws the following research conclusions:(1) The deep learning method has the characteristics of high accuracy, high robustness and low bias. The DeepFM method, one of the deep learning methods, performs better than the other three prediction methods when using temporal data as its input. (2) The feature based methods only using temporal features are basically consistent with those using all available features, indicating that the temporal feature has strong prediction power. Therefore, the ‘peeking’ strategy that monitors the early response of users in the initial period after the items are posted is effective. Additionally, the predictive power of temporary features can be further amplified in time series methods and deep learning methods. (3) Due to the sparse user-item interaction in social networks, the accuracy and efficiency of collaborative filtering methods are low, which makes it impossible to predict the popularity of items in social networks well.
Subject
General Physics and Astronomy
Reference15 articles.
1. Probabilistic matrix factorization;Mnih;Advances in neural information processing systems,2007
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献