Causal Email Threading

Viewing emails as parts of a sequence or a thread is a convenient way to quickly understand their context. It is therefore one of the most widely implemented and used features in mail clients, both web based and otherwise. Alas, current threading techniques are essentially limited to personal conversations which are human generated emails. In this paper we present the first, to our knowledge, system for threading machine generated emails which account for more than 60% of mail traffic. We present a three stage process. First, we identify email templates. These are almost identical messages sent to many users by the same vendor or service provider. Next, we build a template causality graph. Given this graph, one can assess for two emails what is the likelihood that one caused or triggered the other. Finally, using the causality graph, we find the most likely thread structure for each user’s inbound mail stream. We present thorough experimental results obtained by analyzing 2.5 million user inboxes.