基于混合特征学习的微博转发预测方法

摘要

微博转发预测是研究信息传播的关键问题之一，对于舆情监控、广告投放、商业决策具有重要意义。用户兴趣、微博作者影响力及微博内容等信息均影响信息传播过程。转发行为预测的挑战性问题在于如何捕获更多有意义的影响因素以提高预测性能。提出基于混合特征学习的转发预测方法，该方法首先引入并分析了局部社会影响力特征、用户特征、微博内容特征的计算方法；接着，基于分类器建立预测模型；最后，比较了不同类型微博的转发预测效果。在新浪微博平台数据的实验结果表明，局部社会影响力特征、用户特征、微博内容特征都对转发预测有较大影响，其中微博内容特征的影响最大。随机森林预测效果最好，准确率达到83．1％；与朴素贝叶斯、逻辑回归、支持向量机模型相比，准确率平均提高约7．4％，最高提高约10．8％。另外，该方法对自然灾害、环境、审判、维权等类型的微博进行转发预测时，效果更加明显，说明这类事件转发的规律性更强。%Microblogging retweet prediction is one of the key problems in information dissemination,which plays important roles in public opinion monitoring,advertising,and business decision making.The process of information dissemination is influenced by many factors such as user interest,microblogging author’s influence,and content of post,etc.The challenge of improving prediction performance is how to capture the important features for retweet prediction.In this paper,we propose a retweet prediction method based on hybrid features learning. Firstly,the method introduces and analyses the impacts of hybrid features including social influence locality,user features,and microblogging content features.Then,it builds the retweet prediction model based on classification algorithms.Finally,it compares the results of different types of microblog.Experimental results on Sina Weibo datasets show that local social influence features,user features and microblogging content features affect the retweet prediction,and the greatest impact is the micro-blog content features.Random forest method has the best performance,and the accuracy rate can reach 83.1%.Compared to Naive Bayes,logistic regression and SVM,the accuracy rate increased by an average of about 7.4%,the highest increase of about 10.8%.In addition,the method has an advantage on topics about natural disasters,environment,trial,rights,which shows that these kinds of events contain stronger retweet patterns.