Using Deep Learning to Predict the Popularity of Instagram Posts By Plastic Surgeons
Instagram has overtaken Facebook as the social media platform of choice for 18- to 34-year olds. With the rise of social media use, there has been a commensurate increase in resources being devoted to understanding how information propagates across these platforms. Previous studies have described virality and influence in social media marketing, by modelling the spread of content and seeking to maximize the number of users reached. More recently, studies have used artificial intelligence (AI) and a combination of image features, text analysis, and social context to predict the popularity of images online/ on social media. In commercial settings, accurately predicting popularity can help value sponsored content. The aim of this study was to predict the popularity of images posted by plastic surgeons and quantify the social context and content- specific factors that contribute to their popularity. More generally, we sought to answer the question: “What makes a plastic surgery- related image popular?”
Methods/Technique
:A list of US- listed plastic surgeons, current as of December 1st, 2019, was generated from the ASAPS webpage. Instagram accounts (and all posts associated to all accounts) associated to individual surgeons were identified manually. For prediction purposes, we deployed a random forest machine learning algorithm on the dataset to train it to predict the log- scaled popularity of Instagram posts. We used Spearman’s rank correlation (ρ) to quantify relationships between popularity and 1) social context input features (number of followers, followers- to- following ratio, number of posts, mean number of likes/ comments, Instagram verified status, Instagram engagement rate, and timestamp analysis and 2) content- specific input features (object classification, dominant color analysis, and caption length).
Results/Complications
:Across 2,183 US- based ASAPS members, we identified accounts associated to 58.2% (n= 1,272) of plastic surgeons. Across all accounts, we identified 395,537 posts. The mean number of likes was 133 (range 0- 45,588). When we used indicators as input features to predict Instagram likes, mean number of likes (ρ= 0.64) and DL- assisted object classification (ρ= 0.19) achieved the highest Spearman rank correlations of all social context and content- specific indicators respectively. When combined, all social context indicators achieved a rank correlation of 0.22, while the combination of all content- specific indicators achieved a rank correlation of 0.67. The combination of all content- specific and social context indicators generated a rank correlation of 0.74. Our deep learning content analysis revealed that “swimming trunks”, “lab coat”, and “gas mask” had a strong positive impact on popularity. “Convertible car”, “desk”, and “screen” had a moderately positive impact on popularity.
Conclusion
:While previous studies evaluated virality of text and images unrelated to plastic surgery on the internet, this study achieved a significant rank correlation between the predicted and actual popularity of Instagram posts by ASAPS- member plastic surgeons.