Combining textual and non-textual features for e-mail importance estimation

In this work, we present a binary classification problem in which we aim to identify those email messages that the receiver will reply to. The future goal is to develop a tool that informs a knowledge worker which emails are likely to need a reply. The Enron corpus was used to extract training examples. We analysed the word n-grams that characterize the messages that the receiver replies to. Additionally, we compare a Naive Bayes classifier to a decision tree classifier in the task of distinguishing replied from non-replied e-mails. We found that textual features are well-suited for obtaining high accuracy. However, there are interesting differences between recall and precision for the various feature selections.