CS 229 A Project Report : Flame War Detection using Naïve Bayes classi cation techniques

Classifying text using multinomial naïve bayes is now a common technique, notably for spam e-mail ltering. Our goal with this project was to attempt to adapt the same technique to classifying a harder class of documents, ame wars, or very heated discussions in public internet forums. To do so we implemented a custom web scraper to build a corpus of data, accompanied with our implementation of a naïve bayes classier which we then augmented with several tweaks to attempt to gain a better recall measure on large training sets, including a technique described in [1] to leverage the large amount of unlabeled data we could gather.