The scope of this project is to see if we can classify the posts of 2 different communities (subrredits) based solely on how people comment and interact with each other.
In the demo, I extract data from reddit and make a graph from the comments section of each post, where each comment author is a node and replies between them are edges. Then I compute graph embeddings to get features for each graph and classify them with a logistic regression.
Taking the "food" and "hardware" communities, I uploaded approximately 850 posts from each subreddit (/subreddit/hot thread), dividing the dataset into 80% train and 20% test, I got an ~86% accuracy on the test dataset.
After the model is trained we can take any post from the 2 communities used, paste the link into the coresponding text box and make a prediction for it. For representative posts (number of nodes > 10) the model gives very good results, with predicted probabilities > 90%.
I use praw to extract data, networkx to create graphs, karate club to compute graph embeddings, sklearn for logistic regression and streamlit for frontend.
Credits to: Zademn