Using file system content to organize e-mail

This paper is about using existing directory structures on the file system as models for e-mail classification. This is motivated by the aim to reduce the effort for users to organize their information flow. Classifiers were trained on categorized documents and tested on their performance on an unstructured set of e-mail correspondence related to the documents. Even though the documents and e-mails in our corpus belonged to the same categories, the classifiers showed very low accuracy on e-mail classification. More importantly, a learning curve experiment showed that initiating a model with documents can have a negative impact on the overall accuracy that could be achieved on e-mail classification. Features important for e-mail classification are inherently different than those important for document classification.