Predictions to Ease Users' Effort in Scalable Sharing

Significant user effort is required to choose recipients of shared information, which grows as the scale of the number of potential or target recipients increases. It is our thesis that it is possible to develop new approaches to predict persistent named groups, ephemeral groups, and response times that will reduce user effort. We predict persistent named groups using the insight that implicit social graphs inferred from messages can be composed with existing prediction techniques designed for explicit social graphs, thereby demonstrating similar grouping patterns in email and communities. However, this approach still requires that users know when to generate such predictions. We predict group creation times based on the intuition that bursts of change in the social graph likely signal named group creation. While these recommendations can help create new groups, they do not update existing ones. We predict how existing named groups should evolve based on the insight that the growth rates of named groups and the underlying social graph will match. When appropriate named groups do not exist, it is useful to predict ephemeral groups of information recipients. We have developed an approach to make hierarchical recipient recommendations that groups the elements in a flat list of recommended recipients, and thus is composable with existing flat recipient-recommendation techniques. It is based on the insight that groups of recipients in past messages can be organized in a tree. To help users select among alternative sets of recipients, we have made predictions about the scale of response time of shared information, based on the insights that messages addressed to similar recipients or containing similar titles will yield similar response times. Our prediction approaches have been applied to three specific systems - email, Usenet and Stack Overflow - based on the insight that email recipients correspond to Stack Overflow tags and Usenet newsgroups. We evaluated these approaches with actual user data using new metrics for measuring the differences in scale between predicted and actual response times and measuring the costs of eliminating spurious named-group predictions, editing named-group recommendations for use in future messages, scanning and selecting hierarchical ephemeral group-recommendations, and manually entering recipients.Doctor of Philosoph