Predicting Retweet Scale Using Log-Normal Distribution

Abstract

In social network analysis, retweet scale prediction is one important studying focus. Generally speaking, there are two different approaches to predict the retweet scale: timeseries approach and non-time-series approach. In this paper, we conduct a research on the distribution of the reaction time in retweeting activity and introduce a time-series prediction model. We show that in retweeting activity, the reaction time has the feature of heavy-tailed distribution and the log-normal distribution fits the real reaction time data well. Within the framework of time-series prediction, for the direct retweets, we make the prediction by solving the parameter estimation problem of truncated log-normal distribution. For retweets at deeper depths, we make a prediction based on the general information diffusion theory. Experiments are carried out on real data downloaded from SINA weibo. We test the full model on retweet graphs and compare our model with the auto regression model and a perceptron model using tweet text. Our method outperforms the other two models and in experiment, on average, there is a 2% advantage over the auto-regression model when one-hour data are given.

Bayesian Nonparametric Poisson-Process Allocation for Time-Sequence Modeling

“…Through time-sequence modeling, we can try to understand the retweet patterns. For example, one reason could be that the tweets posted at an inactive hour (late at night) will regain the attention from the followers several hours later next morning (Ding and Wu, 2015;Gao et al, 2015). BaNPPA could help us understand such reasons as illustrated in Figure 2.…”

Prediction of online topics’ popularity patterns

“…Time series prediction methods were also imported to solve the problem of popularity volume prediction [20,21]. Ding and Ji [22] studied on the distribution of the reaction time in retweeting activity and presented a time series method to predict the retweet scale using a partial retweet graph. Hu et al [23] constructed a time series feature space to capture the behaviours of popularity of viral topics and utilised a method for predicting the short-term popularity of a given viral topic by using only data of historical popularity of the topic.…”