Learning Predictive Models of Memory Landmarks - eScholarship

Learning Predictive Models of Memory Landmarks Eric Horvitz (horvitz@microsoft.com) Microsoft Research, One Microsoft Way Redmond, WA 98052 USA Susan Dumais (sdumais@microsoft.com) Microsoft Research, One Microsoft Way Redmond, WA 98052 USA Paul Koch (paulkoch@microsoft.com) Microsoft Research, One Microsoft Way Redmond, WA 98052 USA how we construct models that can be used to infer the likelihood that events will serve as memory landmarks, reviewing the extraction of data from subjects’ online calendars, the collection of assessments about landmarks with tools that enable subjects to label their calendar events, and the learning of models via Bayesian learning procedures. After reviewing the performance of the models, we describe, as a sample direction for the use of predictive models of memory landmarks in computing applications, a prototype, named MemoryLens Browser. MemoryLens Browser employs the inferences about landmarks in visualizations for browsing files and appointments. Finally, we review research directions aimed at enhancing coverage and discriminatory power of models of memory landmarks. Abstract We describe the construction of statistical models that provide inferences about the probability that subjects will consider events to be memory landmarks. We review methods and report results of experiments probing the classification accuracy and receiver-operator characteristics of the models. Then, we discuss opportunities for integrating models of memory landmarks into computing applications, and present a prototype time-line oriented content browsing tool. Introduction Studies of memory support the assertion that people make use of special landmarks or anchor events for guiding recall (Shum, 1994; Smith, 1979; Smith, Glenberg & Bjork, 1978)) and for remembering relationships among events (Davies & Thomson, 1988; Huttenlocher & Prohaska, Such landmarks include both public and autobiographical events. More generally, there has been significant study and modeling of episodic memory, where memories are considered to be organized by episodes of significant events, including such information as the location of an event, attendees, and information about events that occurred before, during, and after each memorable event (Tulving, 1983; Tulving & Thomson, 1980). Memory has been shown to also depend on the reinstatement of not only item-specific contexts, but also on more general context capturing the situation surrounding events. We believe that automated inferences about important memory landmarks could provide the basis for new kinds of personalized computer applications and services. Rather than focusing on specific machinery proposed as models for recall (e.g., Malmberg, Steyvers, Stephens, et al.,. 2002;. Raaijmakers & Shiffrin, 2002; Shiffrin & Steyvers, 1997), we set out to investigate the feasibility of directly learning models of memory landmarks via supervised learning. We focus here on the construction, testing, and application of predictive models of memory landmarks, based on events drawn from users’ online calendars. We first review experiments with the construction of personalized models of memory landmarks. We describe Accessing Events and Event Properties We will focus on the construction of models of memory landmarks derived from users’ online calendar information. Electronic encodings of calendars provide rich sources of data about events in users’ lives. People who rely on electronic calendars, often encode multiple types of events in an online format. Such items include appointments, holidays, and periods of time marked to indicate such activities as travel and vacation. In large enterprises that rely on computer-based calendaring systems, appointments and events are typically formulated, accepted, displayed and managed via schemas capturing multiple properties of the events. We developed a calendar event crawler that works with the Microsoft Outlook messaging and appointment management system. The crawler analyzes a user’s online calendar to create a case library of events and properties associated with each event. The calendar crawler extracts approximately 30 properties for each event. Most of these properties are obtained directly from the online data and metadata stored for events. These properties include the time of day and day of week of events, event duration, subject, location, organizer, number of invitees, relationships between the user and invitees, the role of the user (i.e., user was the organizer, a required invitee, or an optional invitee), response status of the user to appointment invitations (i.e.,