Transcript PowerPoint
Forecasting Presence and Availability Joe Tullio CS8803 Overview Why do this? Survey of projects Precursors/influences Coordinate Awarenex/Work rhythms Learning locations using GPS “Lighter” applications Augur Current incarnation Evaluation/future plans Motivation Why do this kind of prediction? Why now? Precursors Media spaces (CRUISER system) Portholes Beard et al – assigned priorities to events Priority was accorded a level of transparency So meeting scheduling involved overlaying calendars Worked well enough in the lab, but saw less success in the workplace. Why? Automatic meeting scheduling tools IM status – focus on current state of availability Coordinate (Horvitz et al) Preceded by Priorities Prioritize incoming notifications Relay to a mobile device if important enough Location was first determined by idle time Later added input from other sources Calendar, vision, audio levels Coordinate (continued) Intent: Answer broad range of queries “When will X return?” “When will X be available?” “Will X attend the meeting?” “When will X have access to a desktop machine?” Coordinate (continued) Method: collect lots of data Calendar, computer activity, devices used, email contents, meeting information, 802.11 location tracking Estimates of attendance augmented with handlabeling when necessary Employee directory establishes professional relationships between users Construct custom Bayesian networks appropriate to the query Example Rhythm modeling (Begole et al) Idea: people exhibit rhythms in their day-to-day work Capture those rhythms by recording email, IM, phone activity, computer use Visualize them and attempt to build models representing them Example Building the models Expectation maximization Discover transitions in activity Cluster similar periods of inactivity Refine Label transitions through simple matching Around 12 or 1 is lunch Recurring transitions named after calendar events, if they exist Location changes named after location, duh Other visualizations Gradient Compressed Probabilities Privacy How much to display, and to whom? Ideas: Expose more over time to simulate familiarization Expose only what is needed to answer a given question But how to explain or give context? Location Modeling Using GPS (Ashbrook and Starner) Location modeling as opposed to availability Uses? Encourage serendipitous meetings Intelligent interruption Meeting scheduling Step 1: find places Can’t just give people raw GPS coordinates Define a place as any location where one spends time t t chosen arbitrarily here Places become locations Use a clustering algorithm to group nearby places Also concept of sublocations Run clustering alg. On points within locations Example Adding time All these locations are time-stamped, so… Can identify order of places visited and predict transitions between places Markov model – one for each location, transitions to every other location Currently can predict where one will go next, but not when Can variance in arrival/departure indicate importance? Machine learning Most of these projects require a large corpus of data with discernable patterns of activity What happens when those patterns deviate or change? Incorporate learning or user interaction Broaden classes in accordance with their current fit to the data Coordinate – include more cases that are ‘relevant’ Rhythms/GPS – Weigh recent data more heavily Predicting interruptibility using sensors Hudson et al Goal: determine good time to interrupt Method: record people in their offices(A/V) Self-report interruptibility using ESM (~2/hr) Manually code situations (602 hours) Hypothesize which sensors would provide the most information about interruptibility Results Building models Simple 2-class classification problem Try: Decision trees (78.1%) Naïve Bayes (75.0%) Adaboost w/decision stumps (76.9%) Support-vector machines (77.8%) Predictions improve when tested per-subject as opposed to across subjects First few sensors account for most of the accuracy: Phone, talk, # of guests, sitting, writing, keyboard