# Dank or Not? -- Analyzing and Predicting the Popularity of Memes on Reddit

**Authors:** Kate Barnes, Tiernon Riesenmy, Minh Duc Trinh, Eli Lleshi, Nóra Balogh, Roland Molontay  
**arXiv:** [2011.14326](https://arxiv.org/abs/2011.14326)  
**Date:** 2020-11-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2011.14326v2/

## Key points

- Dataset of 129,326 memes analyzed from Reddit in March 2020 during COVID-19 pandemic
- Aim of the study: gain insights into thoughts and sentiments of internet users, perform content-based predictive analysis on viral memes
- Machine learning methods used to predict meme success based on content alone with moderate accuracy (AUC score of 0.68)
- Image-related and textual attributes both have significant incremental predictive power over each other for meme popularity
- Negative correlation between amount of text required to read and upvotes; shorter texts perform better
- Neutral memes performed better than extreme ones; negative sentiments performed better than positive sentiments
- Humor in memes often associated with surprise rather than arousal
- Current events played a significant role in meme content during the COVID-19 pandemic
- Study provides valuable insights into internet users' thoughts during the pandemic and offers predictive models for meme virality

## Summary

In this study, the researchers analyzed a dataset of 129,326 memes collected from Reddit in March 2020 during the height of the COVID-19 pandemic. The aim of the study was to gain insights into the thoughts and sentiments of internet users during this time as well as to perform a content-based predictive analysis on what makes a meme go viral. Using machine learning methods, the researchers found that they were able to predict the success of a meme based on its content alone with moderate accuracy. Their best performing machine learning model achieved an AUC (area under the curve) score of 0.68 in predicting viral memes. Additionally, the researchers investigated the incremental predictive power of image-related attributes compared to textual attributes on meme popularity and found that both image-related and textual attributes had significant incremental predictive power over each other. The study also examined the relationship between extracted text features and normalized upvotes and suggested that users may respond better to memes with shorter texts as there was a negative correlation between the amount of text required to read and upvotes; this aligns with previous research on social media engagement with brand posts. Furthermore, sentiment analysis in memes revealed that neutral memes performed better than extreme ones; among extreme sentiments, negative sentiments performed better than positive sentiments. These results contradicted previous findings on viral online news content but suggested that humor in memes is often associated with surprise rather than arousal. They created a word cloud visualization showing these prominent words indicating that current events played a significant role in meme content during this time. Overall, this study provides valuable insights into internet users' thoughts during the COVID-19 pandemic and offers predictive models for meme virality. The findings contribute to understanding internet memes as a form of contemporary social communication and highlight the importance of both image-related and textual attributes in meme popularity.