Related Experiment Video
Updated: May 2, 2026
Generation of Electronic Cigarette Aerosol by a Third-Generation Machine-Vaping Device: Application to Toxicological Studies
Published on: August 25, 2018
Identification and Classification of Images in e-Cigarette-Related Content on TikTok: Unsupervised Machine Learning
Juhan Lee1, Dhiraj Murthy2, Rachel Ouellette1
1Department of Psychiatry, Yale University School of Medicine, New Haven, CT, United States.
Background:
Previous studies identified e-cigarette content on popular video and image-based social media platforms such as TikTok. While machine learning approaches have been increasingly used with text-based social media data, image-based analysis such as image-clustering has been rarely used on TikTok. Image clustering can identify underlying patterns and structures across large sets of images, enabling more streamlined distillation and analysis of visual data on TikTok. This study used image-clustering approaches to examine e-cigarette-related images on TikTok.
Methods:
We searched for 13 hashtags related to e-cigarettes in November 2021 (e.g., vape, vapelife). We scraped up to 1000 posts per hashtag depending on the number of available posts, for 12,599 posts in total. After randomly selecting 13% of posts and excluding non-English (N = 278), non-e-cigarette-related (N = 88), and unavailable posts (i.e., posts that the uploader deleted) (N = 286), N = 838 e-cigarette TikTok images were included in our image clustering model. Using quantitative (e.g., silhouette scores) and qualitative evaluations, we categorized clusters into overarching themes based on the types of e-cigarette content depicted within each cluster.
Results:
We identified N = 20 clusters, forming four overarching themes: (1) vapor clouds (e.g., vape tricks, vaping and exhaling vapor clouds, being captured as clouds from the mouth or nose or around the face); (2) devices (e.g., content presenting e-cigarette devices or individuals demonstrating use or modification of devices); (3) text (e.g., e-cigarette-related text inserted within images such as jokes); (4) other (i.e., e-cigarette-related images clustered based on other image characteristics such as color tones).
Conclusions:
This study using the state-of-the-art image-clustering method successfully identified various e-cigarette-related images on TikTok. This study suggests that novel methodologies can be helpful to tobacco regulatory agencies looking to conduct rapid surveillance of e-cigarette content on social media.
