Discovering Similar Users on Twitter

This work studies the problem of discovering \similar" users at Twitter, where we dene two users to be similar if they produce content similar to each other. The discovery of top similar accounts for each Twitter user has a variety of applications at Twitter including user recommendations and advertiser targeting. Although the discovery of similar results is a well studied problem in information retrieval, the particular problem at Twitter has three novel challenges. The rst one is the heterogenous mix of signals that an effective technique could use: content analysis, social graph structure, user popularity, user interaction data etc. It is apriori unclear how could one blend all these dierent signals in an eective manner. Second, any technique needs to work eectively with very dierent kinds of users. This implies that the same framework needs to be applicable to users with millions of followers and users with very few or no followers. Finally, any proposed technique needs to be able to scale such that it can discover similar users for hundreds of millions of users, while keeping up with the highly dynamic nature of all the input signals. In this work, we share the machine learning based framework that we use to discover similar users at Twitter. We also evaluate the eectiveness of the framework on Twitter data.