New Framework Uses Temporal Dynamics to Forecast Micro-Video Popularity

The Challenge of Predicting Micro-Video Success#

Forecasting whether a short-form video will gain traction remains notoriously difficult. The engagement trajectory of micro-videos tends to fluctuate unpredictably—surging at one moment, then stalling the next. Existing prediction systems struggle with this volatility, particularly when trying to make accurate forecasts far into the future.

The core problem boils down to two separate gaps in how current methods work. First, they fail to recognize that user interactions like likes, comments, and shares feed into each other dynamically—earlier engagement can spark additional engagement, while momentum can fade over time. Second, most retrieval systems look for similar videos based only on static content features (what the video shows, sounds like, or says), overlooking the equally important signal of how a video's popularity curve actually unfolded.

How M³TR Works#

Researchers introduced M³TR, a framework that tackles both limitations simultaneously. The system pairs fine-grained temporal modeling with a sophisticated retrieval mechanism that understands popularity patterns.

At its heart sits a Mamba-Hawkes Process (MHP) module. This component treats user feedback as a sequence of self-exciting events—meaning it explicitly models how each interaction can trigger more interactions downstream, while also capturing how that excitement gradually decays. This approach preserves the long-range dependencies buried in user behavior.

That rich temporal understanding then fuels a retrieval engine that hunts for historically relevant reference videos. Rather than matching purely on visual, audio, or text similarity, the system searches across both multi-modal content and popularity trajectories. If a target video shows promise similar to how past videos evolved, their features can inform the prediction.

Results and Impact#

Testing on two real-world datasets revealed significant improvements. M³TR delivered state-of-the-art results, beating prior methods by up to 19.3% on normalized Mean Squared Error and demonstrating substantial gains on long-term prediction tasks—the hardest challenge in the field.