From context aware to socially awareness computing - IEEE Pervasive Computing, 11 (1), 32–41. (2012
- 1. From Context Awareness to Socially Aware
Computing
Paul Lukowicz
DFKI Kaiserslautern,
Germany
Fredrick Awuor
Alex Pentland
MIT Mdeia Lab
Alois Ferscha
Johannes Kepler Univ. of Linz,
Austria
IEEE Pervasive Computing, 11 (1), 32–41. (2012)
- 2. Outline
Introduction
Beyond locomotion and location
Challenges
Towards Solution
Realizing opportunistic recognitions
Socially Aware Computing
Sensing Social Interactions
Sensing Social Phenomena
Discussion
Comment
31-Mar-15
2
- 3. Introduction
Context Awareness – Notion of systems adapting their functionality to user’s activities
and situation in the environment
Marc Weiser’s dream: mobile phones featuring acceleration and light sensors and offering ability to
recognize simple context such as device position (on the table or in the pocket), mode of location, and
user location – and adapt their functionality accordingly
Key directions that are driving recognition research:
Moving beyond locomotion and location to incorporate complex activity recognition into
real-world applications
Realizing opportunistic recognitions; and
Moving from ”single-system, single user” perspective toward large-scale ensembles of
networked systems interacting with communities of users
31-Mar-15
3
- 4. Beyond Modes of Locomotion and Location
Context is more than simply recognizing motion state and location.
Why less impact of complex activity recognition to real world applications?
Challenges: #1. Recognizing compound activities
Recognizing modes of locomotion is well-defined constrained problem (there’s limited
number of nonoverlapping, discriminant states (labels) with labels such as standing,
sitting, driving, cycling, walking, running
Assumption: People walk, stand or sit but not two at the same time (exceptions such as sitting while driving exit)
Activities of daily living (ADL –based approach) is motivated by contextual activities and
include things such as housekeeping, meal preparation, watching TV.
Recognition labels here represent complex compound activities, some of which can be
composed of numerous, diverse lower-level activities
Many of the activities can be executed in parallel (housekeeping and watching TV) or
interleaved
31-Mar-15
4
- 5. Beyond Modes of Locomotion and Location
Challenges: #2. Accommodating high variability
Modes of Locomotion have a distinct signature in an easily deployable sensor modality
(acceleration or motion sensors in general).
Activates like housekeeping, there is simply too much variability in the way they can be
performed.
We can rely on highly instrumented environment
Putting RFID tags on every relevant object in a household and RFID reader on a user’s wrist to provide
recognition
Smart apartments with binary sensors indicating the opening of every cabinate or removal of every
object
Problems:
How practical is such extensive instrumentation
The transfer of results from one environment to another isn’t straightforward. One person might use an electric
kettle to warm water for tea, while another gets hot water from coffee machine
31-Mar-15
5
- 6. Beyond Modes of Locomotion and Location
Challenges: #3. User-specific training
Modes of locomotion can be trained reasonably well in a user-independent way. That is,
smartphone app can be pretrained and delivered ready to use.
For complex activities, good results are mostly achieved only with user (and environment)
specific training.
31-Mar-15
6
- 7. Beyond Modes of Locomotion and Location
Towards a Solution:
#1. Activity models and high-level reasoning
Complex structure of human activity: Activity descriptions (and range of variations) can be
mined from the web or contributed by users
Also, models of high-level activities can contribute to increasing the accuracy of the low-level
actions on which they build by providing additional boundary conditions and prior probabilities
Finding a set of activities that covers a significant portion of relevant applications and models
that optimally support the recognition process.
#2. Need for standard datasets for training and evaluation
#3. Experimental procedures and description standards
Much of activity-recognition work published to date is inherently difficult to replicate.
Observing a set of generally accepted “best practice guidelines” could significantly reduce
the problem
#4 Better Evaluation Metrics: Example: Evaluation on sensor-frame basis, even basis etc
Bottom line: Need to put more focus on incremental improvements of algorithms
31-Mar-15
7
- 8. The Need for Opportunistic Systems
Assuming we can reliably recognize complex activities in real-life environments, will that
automatically lead to wide-scale adoption of context aware systems?
State-of-the-art approaches to context recognition mostly assume fixed, narrowly defines
system configurations dedicated to often equally narrowly defined tasks.
Thus, for each application, the user must deploy and place specific sensors at certain
well-defined locations in the environment or his body (purposefully deployed)
For universal context awareness, this approach is not realistic
We need systems that can exploit devices that just “happen” to be in the environment
Sensors are invisibly becoming omnipresent in our environment eg miniaturized motion
sensors unobtrusively integrated into clothing
Note: widespread sensor availability doesn’t necessarily mean that an application can
assume that required sensors will continuously be available in the same configuration.
31-Mar-15
8
- 9. The Need for Opportunistic Systems
Assumption:
As user moves around, s/he is at times in highly instrumented environments where a lot of information
sources are available.
At other times, s/he might stay in places with little or no intelligent infrastructure.
Methodological approach for context recognition reverses:
Sensors no longer purposefully deployed to satisfy data needs for recognition algorithms and
applications.
Instead, sensors are there and recognition algorithms and applications can make appropriate use of
them as needed
This is called opportunistic recognition system (everything is a sensor)
Consider every source of data as an abstract sensor
Miniaturized unit of hardware
Online sensors such as (Twitter feed, facebook updates)
Harvest sensors (which can collect and locally store data)
Playback sensors (which can generate data streams form repositories)
31-Mar-15
9
- 10. The Need for Opportunistic Systems
Abstracting all kinds of information sources into generic sensor categories providing
standardized access interface is the key to opportunistic sensing
Example: goal-oriented sensor assemblies that configure spontaneously to achieve a
common activity (context recognition goal), without requiring a predefined sensor
infrastructure or fixed recognition goal to be defined at design time
Sensor interface abstraction need to be preserved (spontaneous sensor ensemble configuration)
Another Approach: Have sensors and systems accumulate and exchange experiences
about successful sensing missions for the recognized activities in nonsupervised learning
procedure.
“Transfer learning” (then next, “informed” sensing)
31-Mar-15
10
- 11. Socially Aware Computing
Implication so far:
The environment will be full of users equipped with devices that can reliably perform complex
activity recognition.
Every device will be “willing” and able to identify and exploit potential information sources in its
neighborhood
What happens if, instead of considering each device for itself, we look at them as a
collaborating collective? Social aware computing
Novel, more powerful methods for monitoring and analyzing social interactions
Subsuming the information form many individual interactions into models of aggregate human
behavior and social dynamics
Developing collaborative, “social” algorithms for the recognition process itself
Social aware computing is related to recent trend or leveraging people’s mobile phones
for large-scale, mobile sensing
31-Mar-15
11
- 12. Socially Aware Computing
Social aware computing is more than scale of sensing. It allows fundamentally new types
of applications, extending the notion of context aware into a different domain.
#1: Sensing Social Interactions:
Early work focused on looking at signal from a system of a single user to establish the type of
social activity in which s/he could be engaged in
User location, motion patterns, and vocal behavior would be used to determine the type of
social context in which the user was engaged (eg at a meeting, party or alone)
Leveraging collaborative analysis to “reality mine” information from many users collected
continuously during everyday interactions can reveal a much more detailed and subtle
picture of social interactions
31-Mar-15
12
- 13. Socially Aware Computing
Example: Consider reality mining mobile phone – phone GPS data, calls, and email
records to better understand “traffic” within an organization.
Analyzing these digital traces creates a detailed picture of face-to-face, voice, and
digital communication patterns.
31-Mar-15
13
A German bank achieved significant
improvements in performance just by
relocating a previously orphaned
“customer service” group (labeled
“CService,” mostly communicating by
email in the original setup) to facilitate
more face to face communication. This
simple, inexpensive change made sure
that that everyone was “in the loop.”
Red: Face-to-Face Communication
Blue: Email Communication
- 14. Socially Aware Computing
#2. Sensing Social Phenomena:
Further exploration leads to more than social interactions of individuals, but second-by-
second models of group dynamics and reactions over extended periods of time,
providing dynamic, structural, and content information.
Idea: Harness these streams of personal data and use them to create and drive dynamic
models of aggregate human behavior.
Example:
Public health beyond static demographics (patterns of movement between the places a person
lives, eats, works, hang out ie mobile behavior pattern)
Using mobile traces, we can also cluster together people with similar behavior patterns to
discover largely independent subgroups in a population (subpopulation)
Knowing mobility behavior of different subpopulations provides a far more accurate picture of
their preferences and risks than does standard demographic information
31-Mar-15
14
Pervasive sensor systems collaborating in complex, dynamic way could help us discover complex social dynamics
- 15. Discussion
Envisions moving towards reliable, practicable recognition of complex contexts and
activities and performing such recognition using dynamically changing, a priori, unknown
system configurations in virtually new real-life setting
Then we can recognize social contexts, community level situations, and collective human
behaviors
This will facilitate socially aware and adaptive computing as the logical next step from
context awareness
31-Mar-15
15
- 16. Context to Socially Aware Computing: Challenge
Challenge? Privacy and Data Ownership
Insightful approach: You own your data
You have the right to possess your data
You control the use of your data
You have the right to dispose or distribute your data
Reality Mining
31-Mar-15
16
Editor's Notes
- Late Marc Weiser (1952-1999)is considered the father f ubiquitous computing, chief scientist at Xerox PARC, US
- To understand this, consider diseases of behavior, such as diabetes. It makes sense tht the types of restaurants you go to, the frequency of waling, and dimilar patterns of behaviou strongly influences your risk for diabetes.
- To understand this, consider diseases of behavior, such as diabetes. It makes sense tht the types of restaurants you go to, the frequency of waling, and similar patterns of behaviou strongly influences your risk for diabetes.
- To understand this, consider diseases of behavior, such as diabetes. It makes sense tht the types of restaurants you go to, the frequency of waling, and similar patterns of behaviou strongly influences your risk for diabetes.