Modeling Attrition in Organizations from Email Communication

Abstract

Modeling people's online behavior in relation to their real-world social context is an interesting and important research problem. In this paper, we present our preliminary study of attrition behavior in real-world organizations based on two online datasets: a dataset from a small startup (40+ users) and a dataset from one large US company (3600+ users). The small startup dataset is collected using our privacy-preserving data logging tool, which removes personal identifiable information from content data and extracts only aggregated statistics such as word frequency counts and sentiment features. The privacy-preserving measures have enabled us to recruit participants to support this study. Correlation analysis over the startup dataset has shown that statistically there is often a change point in people's online behavior, and data exhibits weak trends that may be manifestation of real-world attrition. Same findings are also verified in the large company dataset. Furthermore, we have trained a classifier to predict real-world attrition with a moderate accuracy of 60-65% on the large company dataset. Given the incompleteness and noisy nature of data, the accuracy is encouraging.