ContextAware Mobile Computing: Learning ContextDependent Personal Preferences from a Wearable Sensor

1
Context-Aware Mobile ComputingLearning
Context-Dependent PersonalPreferences from a
Wearable Sensor Array

• Authors Andreas Krause, Asim Smailagic, Daniel P
  Siewiorek, CMU
• IEEE Transactions on Mobile Computing, Vol. 5,
  No. 2, February 2006
• Presenter here Aravind Krishna Kalavagattu

2
A.SenseWear armband B.BlueSpoon
Headset C.Backpack emcompassing Life-Book P-1120
with 802.11b transceiver D.Tungsten W smart
phone E.External antenna for GPS receiver.
3
Background

• Clustering
• the partitioning of a data set into subsets
  (clusters), so that the data in each subset
  (ideally) share some common trait - often
  proximity according to some defined distance
  measure
• Algorithms
• K-means
• k-means algorithm is an algorithm to cluster
  objects based on attributes into k partitions

4
Background

• Classification
• The learner is required to learn (to approximate)
  the behavior of a function which maps a vector
  into one of several classes by looking at several
  input-output examples of the function.
• Training Set
• Training and generating a model
• Testing Set
• Data is assigned labels according to the learnt
  model from training

5
Classification (example..)
6
Background

• KSOM (Kohonen Self-Organizing Maps)
• In the training phase weights of the whole
  neighborhood are moved in the same direction,
  similar items tend to excite adjacent neurons.
  Therefore, SOM forms a semantic map where similar
  samples are mapped close together and dissimilar
  apart.
• Bayesian Networks
• Ayans slides..
• Key Idea
• If we can make use of independences within the
  random variables effectively, we can avoid
  requiring more probability numbers
• Depends on the way Joint probability distribution
  can be factored
• If all variables are independent, the number of
  parameters is linear in the number of variables
• Example

7
Bayesian Networks

• Bayesian network is a generative model
• Joint Probability Distribution
• Conditional Dependencies
• In techniques used to learn Bayes nets, the order
  in which the variables are introduced is
  important, and it reflects the causality between
  the random variables

8
Motivation

• Exploit context information to significantly
  reduce demands on human attention.
• Example Configure settings of your mobile
  proactively.
• Use a BodyMedia SenseWear to monitor the user and
  learn with minimum supervision
• Defining context is the key!
• Having predefined thresholds to define contexts
  and generic behavior wont satisfy the user
• User interests vary and we need personalization !
• Personalize the application to each user by
  learning their preferences rather than have
  predefined thresholds.

9
Novelty and Contributions

• Context Identification
• Offline and Online approaches to define a context
  abstraction from various sensor readings
• Using state-of-the-art machine learning
  techniques
• Preference Learning
• Using Bayesian networks to effectively represent
  the causal dependencies and learn the
  probabilities using user interactions, for
  further inference.
• Design and implementation of a wearable study
  platform realizing these methods
• A survey justifying the importance of
  personalization
• A platform for implementing the system using a
  wearable sensor array.
• Shows that context-aware wearable computers are
  feasible for real life applications
• Improvements in wearability and usability.

10
Key Ideas and Details

• Two step approach gathers data using wearable
  sensors.
• Context identifier identifies typical contexts
  and classifies acquired signals.
• -Preference Learner relates users device
  interactions to their current context and provide
  feedback plus configuration of the system
  variables

11
Key Ideas and Details
12
Context Identifier
13
(No Transcript)
14
Preference Learning

• Creating a generative model relating the context-
  and system variables
• Technique Bayesian Network
• Efficient method to compute joint PDF
• Can handle incomplete data
• Can incorporate dynamics
• Issues
• Algorithms for parameter- / structure learning
• K2
• Priors

15
Experimental Design

• Motivation of machine learning approach
• Survey among college phone users (preliminary)
• Threshold analysis
• Evaluation of Context Identification method
• Self-report study
• Real-time movement identification /
  classification
• Evaluation of Preference Learning method
• SenSay training
• Self-report study

16
System Architecture

• Three major components
• A laptop
• PDA
• Wearable sensor array

17
Experimentation
18
User studies
19
Software Architecture

• Sensor fusion process modeled as a directed
  acyclic graph
• Sensors and User Interactions are sources
• Preprocessing steps are internal nodes
• Clustering / Learning algorithms are sinks
• Configurable using XML
• Object oriented implementation (Java)
• Extendable with new sensors / preprocessing steps

20
Software details (contd..)

• Event based communication
• Distribution of events over the network or
  streaming into a database (different speeds)
• Infrastructural sensors can connect upon
  availability
• High level of concurrency
• Maintenance / Reliability
• Acoustic feedback in case of error
• Tap into sensor fusion graph
• Runs 10 hours without recharging

21
Drawbacks

• Practical deployment seems very costly
• A PDA, laptop and sensors per person
• Motivating scenario could have been better, than
  just adjusting the cell phone settings.
• From a machine learning perspective, the reasons
  for choosing the specific approach over other
  alternatives is not clear
• Experimental analysis with other possible
  techniques would have been better
• Temporal independence among the observed evidence
  is assumed
• Dynamic Bayes Nets would be more effective
• They seem to agree with this, but leave it as
  future work
• User study is preliminary and not representative
  enough.
• Though the sample is small, the population could
  have been diverse (different age groups,
  different work schedules and habits)
• In experimentation for learning user preferences,
  a study with metrics like precision-recall would
  have been appropriate to understand how the
  systems learning is compatible with the actual
  user preferences.

22
Relation to the class

• Context-aware computing
• New type of sensors
• Location Management
• Related to Midterm paper
• Context-aware migratory services
• Application of machine learning techniques for
  mobile computing problems

23
Relation to our project

• Context-aware caching scheme for real-time health
  monitoring systems
• Monitoring patients in a community setting
• Hierarchical Tree topology
• Sensors report to a PDA
• Intermediate Hubs
• Doctor sits at the server
• Context defined in terms of environmental
  conditions, sensor readings and users health
  history plus vulnerabilities
• Parameters like TTL and Priority are set based on
  the context, to manage resources and ensure
  real-time caching.

24
Conclusions

• Enable a wearable computer to learn
• about individual user states using sensors
• This process should not require supervision by
  the user
• Let the computer learn to associate user states
  with user preferences
• This paper showed (at least a small-scale)
  practical implementation framework to do so!
• Machine Learning techniques are effectively used

25
References

• Andreas Krause, Asim Smailagic, Daniel P.
  Siewiorek, Context-Aware Mobile Computing
  Learning Context-Dependent Personal Preferences
  from a Wearable Sensor Array, IEEE Transactions
  on Mobile Computing, Vol. 5, No. 2, February 2006
• Lecture Slides
• CSE494 (Information Retrieval Course, Dr. Rao
  Kambhampati)
• CSE471 (Introduction to Artificial Intelligence,
  Dr. Rao Kambhampati)
• K. Van Laerhoven, Combining the Kohonen
  Self-Organizing Map and K-Means for On-Line
  Classification of Sensor Data, Artificial Neural
  NetworksICANN 2001, G. Dorffner, H. Bischof, and
  K. Hornik, eds., pp. 464-470, 200
• Wikipedia
• Clustering
• Classification
• A. Krause, D.P. Siewiorek, and A. Smailagic,
  Unsupervised, Dynamic Identification of
  Physiological and Activity Context, Proc.
  Seventh Intl Symp. Wearable Computers, Oct. 2003.