What is ContextAwareness

1
What is Context-Awareness?

• Devices know users context and act accordingly
• Example A mobile senses meeting and rejects
  certain calls

2
The PDA

• HP hx4700
• Battery Time 4.5h
• 624MHz CPU
• 64MB RAM
• Light Sensor, WiFi, Bluetooth
• OS Windows Mobile 2003

3
Bayes Net Example

• Context , Features
• We want to know
• Bayes Net encodes conditional independence
• Using Bayes rule

4
Contexts

• Goal have as detailed data as possible
• Context categories
• Locations (21 contexts)
• Doings (25 contexts)
• Person states (2 contexts)
• PDA states (4 contexts)
• Time and Date (for future use)
• Total of 52 contexts excluding time and date

5
Context examples
Locations
Person states
Actions
PDA states
Street Car Bus Train Airplane Restaurant . . . . .

Music (listening) Film (watching) Shopping Brea
kfast Dinner "Fika" . . . . .
Busy Available
Pocket Table Hand Dock
6
Data Collection

• Data recorded on device in sets of 30 seconds
• 16 bit, 32kHz
• 818 recordings, 7 hours of recorded data

7
Feature extraction
8
Basic Audio Descriptors

• MPEG-7
• AudioPower
• AudioSpectrumEnvelope
• AudioSpectrumSpread
• AudioSpectrumCentroid
• AudioSpectrumFlatness
• Other descriptors
• Zero Crossings Rate
• Mel-Frequency Cepstrum Coeffecients (MFCC)

9
From Descriptors to Features

• Features statistical moments of descriptors
• Mean
• Standard Deviation
• Skewness
• Kurtosis
• Mean(Derivative)
• Further complexity reduction Compute only every
  2nd, 4th, 8th, block

10
Feature Selection

• Have 250 features how to select the best?
• How to account for computational complexity?
• Goal find best feature set for given complexity
  constraint

11
Review of Concepts

• Two groups of feature selection algorithms
• wrappers using classifier as performance measure
• filters classifier-independent
• Mutual information
  determines bound on classifier performance
• Evaluation of infeasible for all
  possible
• Solution max. dependency, min. redundancyin
  iterative algorithms using first order relations,
  e.g.,

How to account for complexity?
12
Our Solution

• Extend criterion to punish complexity
• Impossible to calculate analytical solution for
• Use grid search to find best

Lagrangemultiplier
13
Feature Selection

• Complexity-Constrained Feature Selection
  (Plasberg and Kleijn, ICCE 2007)

14
Classifier Performance Cross Evaluation

• 50 of data for training
• 50 of data for evaluation
• Training/evaluation sets completely separate and
  randomly chosen
• Evaluations for different complexity constraints
  and interval lengths
• Results averaged over 3 cross evaluations

15
Classifier Performance Best Performacne

• Consistent with previous research58 over 27
  similar contexts (Eronen et al, 2006)

16
Confusion matrix for class Locations
Transportation
17
Confusion matrix for class Actions
18
Confusion matrix, class PDA
Pocket
Table
Hand
Dock
19
Summary

• Audio Context-Awareness works on Pocket PC
• New complexity-constrained feature selection