Predicting the Virality of Videos

Dec 9, 2021
Code PDF
Predicting how viral a TikTok video will be.

  • Utilized a custom version of the ViT Video-Vision Transformer to classify whether an input video is viral or not.
  • The TikTok dataset was pre-processed to divide the data based on a thresholded view count.
  • Features like uploader name, duration, view count, tags and audio were combined and processed using an Audio-CNN model to generate an embedding, which was concatenated with the positional encoding of the transformer during each forward pass.
  • Achieved a classfication accuracy of 82%, based on the evaluation methodology provided in the YouTube-8M Large-Scale Video Understanding Challenge.

ML