[Paper Review] Marked Self-Exciting Point Process Modelling of Information Diffusion on Twitter

Feng Chen, Wai Hong Tan|arXiv (Cornell University)|Feb 26, 2018

Complex Network Analysis Techniques27 references3 citations

TL;DR

This paper proposes MaSEPTiDE, a marked self-exciting point process model that predicts tweet popularity by modeling retweet cascades using parametric, time-dependent intensity functions. It outperforms existing methods like TiDeH and SEISMIC, especially with short observation windows (e.g., under 6 hours), due to its fully parametric structure requiring less data for reliable estimation.

ABSTRACT

Information diffusion occurs on microblogging platforms like Twitter as retweet cascades. When a tweet is posted, it may be retweeted and henceforth further retweeted, and the retweeting process continues iteratively and indefinitely. A natural measure of the popularity of a tweet is the number of retweets it generates. Accurate predictions of tweet popularity can assist Twitter to rank contents more effectively and facilitate the assessment of potential for marketing and campaigning strategies. In this paper, we propose a model called the Marked Self-Exciting Process with Time-Dependent Excitation Function, or MaSEPTiDE for short, to model the retweeting dynamics and to predict the tweet popularity. Our model does not require expensive feature engineering but is capable of leveraging the observed dynamics to accurately predict the future evolution of retweet cascades. We apply our proposed methodology on a large amount of Twitter data and report substantial improvement in prediction performance over existing approaches in the literature.

Motivation & Objective

• To develop a robust, data-efficient model for predicting the final popularity of tweets based on early retweet dynamics.
• To address the limitations of existing nonparametric models like TiDeH, which require long observation times and large datasets for reliable infectivity estimation.
• To improve early prediction accuracy by using a fully parametric formulation of the intensity function with optimized component functions.
• To provide a scalable and practical solution for real-time popularity prediction in social media platforms like Twitter.
• To overcome the bias in existing models that rely on highly retweeted cascades (e.g., >2,000 retweets), which are rare and not representative of average tweets.

Proposed method

• Models retweeting as a marked self-exciting point process where each retweet increases the intensity of future retweets.
• Employs a parametric intensity function composed of a time-varying infectivity function and a memory kernel, both estimated via nonlinear optimization.
• Uses B-spline basis functions to flexibly represent the time-dependent infectivity and memory kernel functions.
• Selects optimal parametric forms (e.g., polynomial decay for infectivity, exponential decay for memory kernel) through goodness-of-fit evaluation on training data.
• Applies nonparametric kernel smoothing as a preliminary step to inform parametric model selection, but avoids reliance on it for final inference.
• Predicts final retweet count by integrating the estimated intensity function over time, enabling early forecasting from limited observation windows.

Experimental results

Research questions

• RQ1 Can a fully parametric marked self-exciting point process model achieve superior early prediction of tweet popularity compared to nonparametric alternatives?
• RQ2 How does the prediction accuracy of MaSEPTiDE compare to TiDeH and SEISMIC when trained on short observation windows (e.g., 2–6 hours)?
• RQ3 To what extent does the parametric structure of MaSEPTiDE reduce data requirements and improve estimation stability compared to nonparametric methods?
• RQ4 How does model performance vary across different cascade lengths, especially for less popular tweets?
• RQ5 Can the model be generalized to predict popularity of average tweets, given that most existing datasets are biased toward highly viral content?

Key findings

• MaSEPTiDE achieves significantly lower mean absolute error (MAE) in conditional median predictions than TiDeH and SEISMIC at censoring times of 2, 4, and 6 hours.
• At 8, 10, and 12 hours, MaSEPTiDE's performance is comparable to TiDeH, but the latter only gains a small advantage after 8 hours—when most retweets have already occurred.
• The model's prediction accuracy improves rapidly with observation time, achieving reliable results within two hours of tweet posting.
• The parametric formulation reduces data dependency, enabling stable estimation even with limited retweet sequences.
• The model outperforms competing approaches on the majority of cascades, particularly those with moderate to high popularity, due to its robustness to short observation periods.
• The study confirms that models trained on highly retweeted cascades (e.g., >2,000 retweets) are not representative of average tweets, highlighting a key limitation of existing datasets.