Detecting Viral Social Events through Censored Observation with Deep Survival Analysis

TL;DR

This work addresses detecting viral social events from censored cascade data by framing virality as a time-to-event problem and applying deep survival analysis. It introduces VEDSA, a two-stage framework with a γ model that learns a parametric survival function via an LSTM, and a δ discriminator that uses the inferred survival to classify cascades as viral or not, leveraging discretized bin covariates $B^{(i)}$ and a labeled viral threshold $\zeta$. The method evaluates Exponential, Rayleigh, and Weibull distributions, and demonstrates that Weibull generally provides the strongest predictive performance across Twitter, Digg, and Weibo datasets under censored observations. The results highlight the practical potential for early rumor detection, influence estimation, and targeted information management, with robust performance even when only partial cascade data is available.

Abstract

Users increasing activity across various social networks made it the most widely used platform for exchanging and propagating information among individuals. To spread information within a network, a user initially shared information on a social network, and then other users in direct contact with him might have shared that information. Information expanded throughout the network by repeatedly following this process. A set of information that became popular and was repeatedly shared by different individuals was called viral events. Identifying and analyzing viral social events led to valuable insights into the dynamics of information dissemination within a network. However, more importantly, proactive approaches emerged. In other words, by observing the dissemination pattern of a piece of information in the early stages of expansion, it became possible to determine whether this cascade would become viral in the future. This research aimed to predict and detect viral events in social networks by observing granular information and using a deep survival analysis-based method. This model could play a significant role in identifying rumors, predicting the impact of information, and assisting in optimal decision-making in information management and marketing. Ultimately, the proposed method was tested on various real-world datasets from Twitter, Weibo, and Digg.

Figures (3)

• Figure 1: The cascades have censored data. We Divide the observed part of cascade $c^{(i)}$ into multiple bins $b^{(i)}_{j}$s with length $\mathcal{L}$ for generating the input of the model. Each bin counts the number of events.
• Figure 2: A brief overview of the VEDSA method. A cascade of inputs is observed between $[0, \tau]$, corresponding to steps $0$ to $r$ after prepossessing by Figure \ref{['fig:input']}. From the early stage of uncensored data, $\gamma$ fits the survival function and predicts the probability of virality for censored data. A discriminator $\delta$ will learn to output the virality label by classifying the cascades based on the estimated survival functions.
• Figure 3: VEDSA training phases. The first phase fits the survival function using our recurrent model, and the second phase uses the inferred survival function for viral detection. The red boxes ($\psi_\theta$) are cells of LSTMs.