Evaluating User Actions as a Proxy for Email Significance
Email remains a critical channel for communicating information in both personal and work accounts. The number of emails people receive every day can be overwhelming, which in turn creates challenges for efficient information management and consumption. Having a good estimate of the significance of emails forms the foundation for many downstream tasks (eg email prioritization); but determining significance at scale is expensive and challenging.
In this work, we hypothesize that the cumulative set of actions on any individual email can be considered as a proxy for the perceived significance of that email. We propose two approaches to summarize observed actions on emails, which we then evaluate against the perceived significance. The first approach is a fixed-form utility function parameterized on a set of weights, and we study the impact of different weight assignment strategies. In the second approach, we build machine learning models to capture users’ significance directly based on the observed actions. For evaluation, we collect human judgments on email significance for both personal and work emails. Our analysis suggests that there is a positive correlation between actions and significance of emails and that actions per-formed on personal and work emails are different. We also find that the degree of correlation varies across people, which may reflect the individualized nature of email activity patterns or significance. Subsequently, we develop an example of real-time email significance prediction by using action summaries as implicit feedback at scale. Evaluation results suggest that the resulting significance pre-dictions have positive agreement with human assessments, albeit not at statistically strong levels. We speculate that we may require personalized significance prediction to improve agreement levels.