Classifying Reddit comments by subreddit

Reddit.com is a website that is primarily organized by communities called subreddits, where users can post comments to. As subreddits can have very different cultures, we aim to classify comments by subreddit as a means of sentiment analysis. We use a publicly available reddit comment dataset over the year of 2016 and perform a classification on a selection of 20 subreddits among the top 50 by total comment word counts. We investigate the Recurrent Convolutional Neural Network (CNN) and Attention-Pooled CNN as more advanced models in the classification task. We find that the Recurrent CNN with GRU performs the best out of all the models tested, achieveing an F1 score of 0.53 and a Cohen’s Kappa score of 0.502 . Out of all the subreddits, we find that the Recurrent CNN model gives the highest F1 score on the ”anime” subreddit. Quantitative and qualitative analysis of the Recurrent CNN with GRU suggests that the model is able to reasonably approach the task. Further work suggests that the trained model can also be used to cluster subreddits based on comment similarity.