Mobile Personal Sensing: A new driver for high performance transaction systems
Deborah Estrin, Jeff Burke, Mark Hansen, Katie Shilton, Min Mun
(and based on collaborations with many colleagues at UCLA/CENS)

April 5, 2009

Mobile phones can be programmed to provide high resolution,
location-based, time series data, which can be automatically uploaded
into web-based datastores, activity annotated, and used to draw a
wide array of inferences regarding exposure to health risks, habitual
health patterns, consumer-reported outcomes, and quality of life.
Processing of these data, both onboard the mobile device and via
powerful web services, results in a continuous personal data stream
consisting of a temporally-linked series of location and activity
events. The resulting personal data streams are of qualitatively
higher resolution and validity when compared to self-reports and
can be fed into multi-scale modeling and analyses to further transform
individualized measurements into meaningful Endings.

We have found the following workflow to be common across our Mobile Personal
Sensing systems: 
	- Data are captured based on static participant configuration,
	  automated capture, as well as prompted and unprompted capture
	  by the participant and uploaded to the individual's Personal
	  Data Vault which may be thought of as a private and encrypted
	  container in the cloud.  

	- Within the vault the data are run through data cleaning
	  pre-processing, annotation such as activity classification,
	  and filtering steps to create a personal data stream,
	  comprised of a time-activity-location time series and
	  geocoded media.

	- Depending on the applications subscribed to by the user,
	  data are then passed through a dynamic web of application
	  specific processing steps which transform the data based
	  on application specific inferences.

	- The outputs of the inferences are passed to the user-facing
	  components of the system, such as data visualization and
	  feedback to the mobile device, which may in turn trigger
	  additional data capture.

Two elements of MPS make it relevant as an important driver for
high performance transaction systems:  the combination of ubiquitous
(24x7x3*10^9) data capture and leveraged web/cloud data processing.
In some cases, the data collected on a mobile device are enough to
reveal an interesting pattern on their own. However, when processed
through a series of external and cross-user data sources, models,
and algorithms, simple data can be used to infer complex phenomena
about individuals or groups. Mapping and other interactive capabilities
of today's web enhance the presentation and interpretation of these
patterns for participants.

A third element of MPS introduces important constraints on the
solution space and warrants particular consideration, namely the
inherently intimate nature of the data captured with mobile personal
sensing and the resulting concerns for protecting individual privacy,
for personal data stream ownership, and for visibility into the web
of processing that is used to contextualize and interpret the data.
We are exploring the design of (a) Personal Data Vaults to support
personal-data ownership and selective sharing, and (b) Trace-audit
mechanisms to provide visibility into shared-data handling.

In previous work we have described our experience developing one
such application and the transaction processing needed to support
even a small scale system [M.  Mun, et al, =D2PEIR, the Personal
Environmental Impact Report, as a Platform for Participatory Sensing
Systems Research. To appear in ACM Mobisys 2009, Krakow, Poland,
June 22-25, 2009.] Through participation in this HPTS 2009 workshop
we look forward to exploring: (a) the implications of these 24x7x6*10^9
systems for scaling of high performance transaction systems more
generally, and (b) the implications of the privacy sensitive nature
of the applications for personal-data handling and visibility.