Learning Individuals’ Patterns and Contextual Events with Mobile Data Streams
Personal mobile devices such as smartphones are powerful tools for people to collect data about their lives. Not only are smartphones capable and connected computing and sensing devices, but their portability makes them proxies for their users in terms of data as they become an increasing part of users’ routines. People do many tasks for communication and information consumption on their smartphones, and the phone can automatically track user context like location and motion. This enables them to help users to learn about themselves and their routines, and also provides their applications in domains like health and productivity with useful data to make services more relevant, adaptable, and intelligent.
Useful data from phones includes both immediate context information and higher level analyses about user patterns. This research includes investigations into both of these context types. In the first phase we developed techniques for detecting and classifying users’ current transport mode. Our solution, MyClassifier, addresses the difficulty of creating a classifier that performs accurately for a variety of users with different gait. We used an online, semi-supervised learning method to adapt a generic transport mode classifier to users’ own data, increasing the accuracy when the default classifier was initially inaccurate.
The second phase of the research explored the use of transport mode and other immediate context streams available on smartphones to learn higher-level aspects of personal context. Our solution, RoutineSense, predicts the consistency of individuals’ daily routines from the data that smartphones can passively collect. Routine regularity is an important metric both for its intrinsic value as it relates to psychological health factors, and as an input for other personalized applications. Researchers in the health sciences measure lifestyle consistency with the Social Rhythm Metric (SRM), a score based on the consistency of when certain routine daily events, such as dinner time, occur.
The events traditionally used to calculate SRM were designed to be significant recurring events in everyone’s routines. RoutineSense uses a novel method of identifying passively-detectable recurring events (called “landmark” events) which, along with their expected time offsets with events of interest, can be used to predict the times of these daily events. Unlike other methods tailored to detecting a specific event, this method can work on any recurring event of interest in a user’s routine. This general approach, while potentially less accurate than targeting specific EOIs with a separate method, allows it to scale to more events, and could support user-defined events in future applications. The events traditionally used to calculate the SRM were chosen in advance to be appropriate for self-report manual data entry scenarios. These specific measurements can be difficult to capture accurately without daily user input, even with our landmark event method. However, the landmark events mobile phones can detect in users routines from their smartphones’ passive context data streams provide an alternative measure of daily rhythms. Rather than recreate the SRM-5 by predicting the times of the traditional set of 5 events designed for human recall, RoutineSense chooses the most representative landmark event corresponding to each traditional event, and uses these landmarks directly as a surrogate set of events to create the “SRM-P” (passive SRM), allowing users to track their routine regularity without the continuous requirement of daily surveys.