SURGE benchmark tracks social media sentiment with 67 events and interaction structure
New benchmark reveals reply threads hinder forecasting accuracy in crisis response
Deep Dive
SURGE, a new multi-event social media benchmark, pairs 800K+ posts across 67 events and 5 categories with calendar-aligned sentiment time series while preserving interaction structure linking posts within each event. This enables controlled study of how interactions shape collective dynamics. Key finding: reply-dense periods are hardest to forecast, and naive baselines remain hard to beat under absolute error.
Key Points
- SURGE includes 67 events, 800K+ posts from 5 categories (politics, disaster, entertainment, sports, tech) with three temporal granularities
- Preserves reply thread structure (interaction graphs) enabling controlled study of social dynamics on forecasting accuracy
- Reply-dense periods are 20-30% harder to forecast than average, and naive baselines still beat advanced models under absolute error
Why It Matters
Enables more accurate real-time sentiment forecasting for crisis response, opinion tracking, and event monitoring by incorporating social interaction structure.