Learning semantic information from Internet Domain Names using word embeddings

摘要

Word embeddings is a well-known set of techniques widely used in Natural Language Processing (NLP). These techniques are able to learn words' semantic based on the distributional hypothesis which states that words that are used and occur in the same contexts tend to purport similar meanings. This paper explores the usage of word embeddings in a new scenario to create a Vector Space Model (VSM) for Internet Domain Names (DNS). Our goal is to find semantically similar domains only using information of DNS queries without any knowledge about the content of those domains. The results presented here have practical applications in many engineering activities including websites recommendations, identification of fraudulent or risky sites, parental-control systems and anomaly detection in network traffic analysis (among others). We use the distributional hypothesis to learn the semantic of domain names from users' web navigation patterns, validating empirically that domain names that occur in the same web sessions tend to have similar semantic. We also test different word embeddings techniques: word2vec, app2vec (considering time intervals between DNS queries), and f astText (which includes sub-word information). Due to the characteristics of domain names, we found f astText as the best option for building a VSM for DNS, being 10.5% superior than word2vec with Skip-Gram which was the next best technique considering the Mean Average Precision at k (MAP@k) metric, which compares the most similar domains in our VSM with the most similar domains provided by a third party source, namely, similar sites service offered by Alexa Internet, Inc.