All projects

Predicting the Virality of Videos

A transformer that forecasts whether a short-form video will go viral.

Dec 2021 Machine Learning

A multimodal transformer predicting whether a TikTok video will become viral from its content and metadata.

  • Used a custom Video-Vision Transformer (ViT) to classify whether an input video is viral.
  • Pre-processed the TikTok dataset by thresholding view counts.
  • Combined uploader name, duration, view count, tags and audio via an Audio-CNN embedding, concatenated with the transformer's positional encoding each forward pass.
  • Achieved 82% accuracy under the YouTube-8M Large-Scale Video Understanding evaluation methodology.
TransformersViTMultimodalAudio-CNN