微博个性化转发行为预测新算法

目前,对微博转发行为预测主要是对所有微博用户的历史数据进行学习,从而得到转发模型.但是这类模型需要对所有用户的转发行为进行全局预测,存在同质性且无法对具体用户进行个性化预测的缺陷.针对这些问题,提出了基于多任务学习的个性化微博转发行为预测算法.对新浪微博进行了数据抓取、分析和特征选择,根据社会学中影响力的理论,针对微博用户之间进行社交信息交流而导致相互影响的特点,引入了多任务学习方法,以逻辑回归预测模型作为基准算法,将预测模型分为全局模型与个性模型进行学习.预测模型把对每个用户转发行为的预测对应为多个任务,根据微博用户间的社交交互对这些任务进行关联.实验结果表明,所提出的模型能够有效地对单个用户的微博转发行为进行预测,并且提高了转发行为预测的准确率.

Novel algorithm for predicting personalized retweet behavior

Recently, models for predicting the user retweet behavior are based mainly on the historical retweet data of all users. However, these models are of homogeneity and could not predict a particular user's behavior. To overcome these problems, we propose an algorithm for predicting personalized retweet behavior. Based on crawled Weibo data, we have conducted an analysis and a selection of retweet features. According to the influential theory, we introduce the multi-task learning framework to divide the tasks into common global tasks and many individual tasks. Our massive experiments show that our algorithm is effective in predicting personalized retweet behavior.