Identifying Viral Short‑Video Content on Kuaishou: Models, Features, and Engineering Framework

The talk introduces the problem of early detection of potential viral short‑video素材 on the Kuaishou platform, emphasizing the need for timeliness, accuracy, reusability, interpretability, and automation.

Four primary素材 types are defined: hashtags (activities), music, magic‑effects (filters), and templates, each of which can become a viral driver when associated with many videos.

Three core viral characteristics are identified—rapid view growth, virus‑like replication, and high user engagement—and each is quantified using metrics such as view‑probability, acceleration, and first‑order differences.

The detection framework mirrors a recommendation system and consists of a coarse‑recall stage that extracts user, content, and interaction features, followed by a fine‑ranking stage that re‑orders candidates based on click, conversion, and dwell‑time objectives.

Three speed indicators (first‑order difference, acceleration, and viral‑speed ratio) are fed into a model that learns their optimal weights; a Gaussian Mixture Model (GMM) provides a soft latent variable z, which together with a shallow DNN yields the final viral probability.

Case studies demonstrate that the system can predict viral peaks hours ahead of actual view spikes, such as the "Winter Mask" activity and the "Falling Star" effect.

The Q&A covers practical advice for creators, labeling of training data, and why the system focuses on素材 rather than individual videos.