An Email Attachment is Worth a Thousand Words, or Is It?

TL;DR

This paper proposes a novel social network analysis method based on shared email attachments, revealing stronger social ties and organizational structures within email archives, exemplified by the Enron dataset.

Contribution

It introduces using shared email attachments as edges in social network analysis, highlighting their significance and providing insights aligned with organizational hierarchy.

Findings

Attachments constitute 80-90% of email archive disk space.

Shared attachments network reveals organizational structure.

Analysis shows different centrality measures highlight distinct social roles.

Abstract

There is an extensive body of research on Social Network Analysis (SNA) based on the email archive. The network used in the analysis is generally extracted either by capturing the email communication in From, To, Cc and Bcc email header fields or by the entities contained in the email message. In the latter case, the entities could be, for instance, the bag of words, url's, names, phones, etc. It could also include the textual content of attachments, for instance Microsoft Word documents, excel spreadsheets, or Adobe pdfs. The nodes in this network represent users and entities. The edges represent communication between users and relations to the entities. We suggest taking a different approach to the network extraction and use attachments shared between users as the edges. The motivation for this is two-fold. First, attachments represent the "intimacy" manifestation of the relation's strength. Second, the statistical analysis of private email archives that we collected and Enron email corpus shows that the attachments contribute in average around 80-90% to the archive's disk-space usage, which means that most of the data is presently ignored in the SNA of email archives. Consequently, we hypothesize that this approach might provide more insight into the social structure of the email archive. We extract the communication and shared attachments networks from Enron email corpus. We further analyze degree, betweenness, closeness, and eigenvector centrality measures in both networks and review the differences and what can be learned from them. We use nearest neighbor algorithm to generate similarity groups for five Enron employees. The groups are consistent with Enron's organizational chart, which validates our approach.