A new preprint reports that a machine-learning model forecast the future popularity of social-media information cascades more accurately than the other systems tested on two datasets. The reported gains ranged from 2.01% to 12.76% on one error measure and from 2.88% to 18.12% on another, across the datasets, observation settings and metrics examined.
The result is about prediction, not control. The model estimated how much a cascade would grow after an initial observation period; it did not show that the system causes information to spread more widely or narrowly.
A forecast built from several views
The system, called TFM4POP, combines several kinds of information about a post and its diffusion. Its central tabular branch uses a pretrained tabular foundation model, meaning a model designed to learn useful patterns from structured data before being adapted to a specific task. A second branch uses a Neural-ODE to represent cascade activity as changing continuously over time, and cross-attention joins the two representations.
The prediction target was incremental popularity: the number of future diffusion events minus the number already observed. The researchers measured performance with Mean Squared Logarithmic Error and Mean Absolute Percentage Error, two measures of prediction error.
To reduce the risk that a training example would indirectly reveal its own answer, the training process built out-of-fold contexts. In practice, a query was kept from appearing alongside its own label when those contexts were constructed.
The reported gains across two datasets
The benchmark used 7,205 Twitter cascades involving 225,998 users, and 33,455 cascades or posts from EventCas involving 1,109,542 users. EventCas was collected from Sina Weibo around 148 hot events and retained posts with more than five retweets.
The tests used different forecasting setups on the two sources. Twitter was observed for either one or three hours before a three-day prediction horizon. EventCas was observed for either half an hour or one hour before a 24-hour horizon.
The reported TFM4POP mean MSLE and MAPE values were 1.2382 and 0.2145 for the one-hour Twitter setting, and 1.2917 and 0.2459 for the three-hour setting. On EventCas, the corresponding values were 2.7544 and 0.2326 after half an hour of observation, and 2.2276 and 0.2218 after one hour.
The authors report that the model led all listed baselines in both datasets and observation settings, with gains across five random seeds. The main benchmark results were summarized as means with standard deviations.
Removing pieces weakened the forecast
Tests that removed one information view at a time all worsened the reported ablation performance, with the MSLE increase reaching 27.70%. The comparison suggests that the combined views were useful within the tested configurations, although the ablation results came from a single fixed random seed.
The largest reported drop came when the pretrained tabular backbone was replaced with a parameter-matched multilayer perceptron, or MLP. MSLE rose by 14.53% on Twitter and 70.76% on EventCas in that comparison.
The adaptation method also mattered in the reported ablation. IA3 performed better than full fine-tuning, LoRA and a frozen backbone while changing only about one ten-thousandth of the backbone parameters. Replacing the continuous-time design with a GRU-only branch, or replacing cross-attention with simple concatenation, worsened performance.
Evidence remains tied to the benchmark
The evidence remains tied to two real-world datasets and the tested observation and prediction windows. EventCas was curated around hot events on Sina Weibo and retained only posts with more than five retweets, a design that may limit how representative it is of broader posting activity.
The main benchmark gains were reported across five random seeds, while the ablation comparisons used one fixed seed. The findings therefore concern the listed model configurations and do not establish how the approach would perform on other platforms, languages, event types or sampling schemes.
The authors state that the EventCas benchmark will be publicly released for future research. The supplied record does not say that the release is already available or provide access instructions.
Paper data and sources
Original title: Tabular Foundation Models for Multi-View Information Cascade Popularity Prediction
Authors: Wenting Zhu, Chenghua Gong, Sanchuan Guo et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text