Conceptual
Browse Instruct Curate

Multi-Modal Video Popularity Prediction

Predicting short-video engagement (view, like, comment, and share counts) by fusing several complementary representations of each video: features from video-classification transformer backbones, a text description generated by a video-to-text vision-language model and encoded with BERT, and engineered tabular metadata features. Per-metric neural networks are ensembled with a gradient-boosted (XGBoost) model on the tabular features, and their predictions averaged to minimize mean absolute percentage error.