The paper tackles the Privacy, Sharing, and Interest (PSI) problem that hinders the widespread adoption of peer‑to‑peer interactive digital media. Traditional PSI mechanisms rely on explicit set‑membership specifications (buddy lists, security approvals) which are cumbersome and fail to capture the nuanced, context‑dependent nature of human relationships. The authors propose an automatic "mind‑reading" approach that continuously infers PSI parameters from real‑time analysis of users’ physical, social, and behavioral signals. Using large‑scale reality‑mining experiments with instrumented smartphones and wearable badges, they collect hundreds of thousands of hours of location, proximity, audio, and motion data. The paper introduces three core analytical techniques: (1) Eigenbehaviors – low‑dimensional principal‑component models of routine behavior that enable rapid classification of location context; (2) Influence Modeling – a coupled hidden‑Markov/influence model that captures conditional‑probability relationships between individuals to predict social affiliations and friendships; (3) Social Signal Classification – short‑duration behavioral displays (excitement, freeze, determined, friendly) extracted from voice and motion that predict interaction outcomes such as information exchange, negotiation success, or call‑center performance. Experiments demonstrate high inference accuracies (up to 96% for workgroup affiliation, 93% for company affiliation, 80‑90% for various interaction predictions) and show that limited user feedback can further improve performance. An illustrative scenario of media sharing at a soccer match shows how inferred location, social, and interest signals can automatically select appropriate peers and weight media content. The authors argue that such semi‑supervised automatic PSI can achieve comparable or better accuracy than manual settings while dramatically reducing user burden, thereby enabling broader adoption of interactive digital media.
Topics
Defining the PSI Problem and the Need for Automatic Inference
Introduces Privacy, Sharing, and Interest (PSI) as three intertwined dimensions of media distribution and explains why manual set‑membership approaches are insufficient.
Reality Mining with Smartphones and Wearable Badges
Describes the two sensor platforms—programmable smartphones and electronic badges—used to collect large‑scale location, proximity, audio, and motion data across multiple experiments.
Eigenbehaviors: Low‑Dimensional Modeling of Routine Behavior
Presents principal‑component (Eigenbehavior) analysis as a method to compress daily location sequences into a few components that capture the majority of variance, enabling rapid location‑context classification.
Influence Modeling: Predicting Social Relationships from Proximity
Explains the influence model (a coupled HMM framework) that uses conditional‑probability relationships between individuals to infer workgroup affiliation, friendships, and company membership.
Social Signal Classification: Thin‑Slice Behavioral Displays
Defines four short‑duration behavioral displays (excitement, freeze, determined, friendly) extracted from voice and motion, and shows how they predict outcomes in speed‑dating, negotiations, call centers, and sponsor meetings.
Application Example: Context‑Aware Media Sharing at Sports Events
Illustrates how inferred location, social, and interest signals can automatically select peers and weight media clips for a soccer‑match scenario, reducing user effort.
Evaluation Results: High Accuracy of Automatic PSI Inference
Summarizes quantitative findings—up to 96% accuracy for workgroup detection, 93% for company affiliation, 80‑90% for interaction predictions—and discusses the role of limited user feedback.
Semi‑Supervised PSI: Balancing Automatic Inference with User Input
Discusses how the system can request occasional user labels for novel or uncertain cases, achieving very high accuracy with minimal user burden.