TikTok Video Virality Study
A computer-vision study of whether video content can predict engagement. The project combines a custom dataset, two definitions of virality, and fine-tuning of a pretrained video model.
View source on GitHubThe question
Could a model estimate a TikTok video’s engagement from its visual content? This collaborative academic study explored that question through both regression and binary classification. The report was co-authored with Dylan McIntosh, Samuel Bernsen, Karen Ngo, and Spencer Au.
Defining the target
Engagement is not a single label. The continuous target combines transformed likes, views, shares, and comments, while the classification target marks videos above the median on every engagement measure. Each definition encodes a different interpretation of popularity.
Training and evaluation
The implementation adapts an R(2+1)D-18 model pretrained on Kinetics-400. Video preprocessing samples fixed-length clips and pads shorter inputs. The report evaluates regression errors and classification metrics, then checks individual predictions.
The classifier’s reported ROC-AUC was 0.4908, and the regression model produced the same prediction across the examined examples. The experiment did not establish a useful virality predictor. It did show why inspecting model outputs matters alongside aggregate scores.