International Edition
Latest News
Technology

Neural Temporal Opinion Model: Joint Prediction of Stance and Posting Time on Social Media

Opinion prediction on social media often relies on the assumption that connected users share similar views, yet forecasting when a user will post next remains a distinct challenge. Existing models typically require manual segmentation of tweet sequences into…

Opinion prediction on social media often relies on the assumption that connected users share similar views, yet forecasting when a user will post next remains a distinct challenge.

Existing models typically require manual segmentation of tweet sequences into fixed intervals, ignoring the reality that posting frequencies vary widely among individuals.

To address this limitation, researchers proposed the Neural Temporal Opinion Model (NTOM), which treats the time interval between consecutive tweets as a random variable. The approach relies on a temporal point process inspired by multivariate Hawkes processes to jointly forecast a user’s next posting time and associated stance.

Architecture of NTOM

The model processes user input by feeding a Bi-LSTM layer with the user’s tweet, a bag-of-word representation, the time elapsed since the previous post, a user embedding, and a queue of neighbor tweets. Neighborhood context is processed through a stacked Bi-LSTM/LSTM layer before passing to an attention module.

Topics are extracted using a Variational Autoencoder to generate a latent topic vector from bag-of-word representations. The system combines tweet features, topic vectors, user representations, and time intervals into a GRU cell. This cell updates hidden states to compute an intensity function predicting both the next post time and stance label.

Performance and Related Methods

Previous tracking tools often depended on manual time intervals or combined recurrent neural networks with marked temporal point processes for geolocation or event tracking. NTOM builds on the Recurrent Marked Temporal Point Process framework but incorporates social relations and topical attention specifically for stance prediction.

Evaluated on datasets covering Brexit and the United States general election, the model outperformed existing approaches in both stance classification and posting time accuracy. The framework demonstrates that combining neural temporal point processes with neighborhood context yields better predictions of dynamic online discussions.

About the author: Anika Shah - Technology

MSc in Computer Science, senior reporter. Anika focuses on AI ethics, cybersecurity, and emerging hardware—frequently moderating panels at CES and Web Summit. “Anika Shah decodes tech breakthroughs and startup disruption shaping tomorrow’s digital landscape.”