Towards website domain name classification using graph based semi-supervised learning

In this work, we tackle the problem of classifying websites domain names to a category, e.g., mapping bbc.com to the "News and Media" class. Domain name classification is challenging due to the high number of class labels and the highly skewed class distributions. Differently from prior efforts that need to crawl and use the web pages’ actual content, we rely only on traffic logs passively collected, observing traffic regularly flowing in the network, without the burden to crawl and parse web pages. We exploit the information carried by network logs, using just the name of the websites and the sequence of visited websites by users. For this, we propose and evaluate different classification methods based on machine learning. Using a large dataset with hundreds of thousands of domain names and 25 different categories, we show that semi-supervised learning methods are more suitable for this task than traditional supervised approaches. Using graphs, we incorporate in the classifier aspects not strictly related to the labeled data, and we can classify most of the unlabeled domains. However, in this framework, classification scores are lower than those usually found when exploiting the page-specific content. Our work is the first to perform an extensive evaluation of domain name classification using only passive flow-level logs to the best of our knowledge.