Simple Model Assemblages for Website Identification

We address the task of identifying the originating website given some anonymized web traffic for the purposes of gaining new insights for improving network security and optimization. We construct a data set of TCP streams sourced from 5 different websites: github.com, google.com, reuters.com, wikipedia.org, and youtube.com. With this, we develop and test 15 different machine learning models based on simple and ensemble model designs. We determine that a bagging ensemble model built using a decision tree provides the best performance with an accuracy score of 0.83 on the testing set, demonstrating the capabilities of ensemble models and suggesting that web traffic data may be distinctive and generally non-overlapping. Limitations to this paper are then analyzed and addressed with potential future avenues for improvement.