Discovery of Activity Patterns using Topic Models Paper by Tm Hunh, - - PowerPoint PPT Presentation

discovery of activity patterns using topic models
SMART_READER_LITE
LIVE PREVIEW

Discovery of Activity Patterns using Topic Models Paper by Tm Hunh, - - PowerPoint PPT Presentation

Discovery of Activity Patterns using Topic Models Paper by Tm Hunh, Mario Fritz and Bernt Schiele Presentation by Roland Meyer 2 Introduction Detect routines based on body movement Complex due to large variations in activities 3


slide-1
SLIDE 1

Discovery of Activity Patterns using Topic Models

Paper by Tâm Huỳnh, Mario Fritz and Bernt Schiele Presentation by Roland Meyer

slide-2
SLIDE 2

Introduction

  • Detect routines based on body movement
  • Complex due to large variations in activities

2

slide-3
SLIDE 3

Contributions

  • New method to recognize daily routines
  • Reusing an established method from text processing
  • Applicable without user annotation

3

slide-4
SLIDE 4

Topic Models

  • Used for text processing for classification
  • Collection of words (“Bag-of-words”)
  • Unsupervised

4

slide-5
SLIDE 5

Topic Models

5

slide-6
SLIDE 6

Daily Routine Modeling

6

slide-7
SLIDE 7

Data collection

  • 1 person
  • 16 days
  • 2 wearable sensors
  • Accelerometer
  • Realtime clock
  • 4 hours of memory

7

slide-8
SLIDE 8

Annotation

  • Online annotation
  • Periodic set of questions on cell phone
  • Time diary
  • Occasional snapshots
  • Offline annotation
  • User could correct / complement data
  • Used as ground truth

8

slide-9
SLIDE 9

Discovering activities

  • 34 distinct activities
  • Mean, variance, frequency from acceleration sensors
  • Combined with time-of-day
  • SVMs, HMMs, Naive Bayes evaluated as classifiers
  • 72.7% accuracy
  • Great variations
  • Problems with short and similar tasks

9

slide-10
SLIDE 10

Discovering topics

  • Latent Dirichlet Allocation on activity data
  • Sliding window of 30 min. over activity stream
  • 10 topics

10

slide-11
SLIDE 11

Discovering topics

11

slide-12
SLIDE 12

Results on Discovering topics

  • Precision and recall calculated for 6 of 7 day to cross-

validate results

  • Supervised classifier using HMMs to calculate baseline

12

slide-13
SLIDE 13

Unsupervised learning

  • Get rid of user annotations
  • Labels from data clustering

13

slide-14
SLIDE 14

Future work

  • Semi-supervision
  • Noise modeling
  • Include location information
  • More users with more diverse lives
  • Build applications
  • Use better sensors (more memory)

14

slide-15
SLIDE 15

Including location

  • “Discovering Daily Routines from Google Latitude with Topic Models”

by Laura Ferrari and Marco Mamei

  • “Discovering Human Routines from Cell Phone Data with Topic Models”

by Katayoun Farrahi and Daniel Gatica-Perez

15

slide-16
SLIDE 16

Including location

16

“Discovering Daily Routines from Google Latitude with Topic Models” - Laura Ferrari and Marco Mamei

slide-17
SLIDE 17

Including location

17

“Discovering Daily Routines from Google Latitude with Topic Models” - Laura Ferrari and Marco Mamei

slide-18
SLIDE 18

Including location

18

“Discovering Human Routines from Cell Phone Data with Topic Models” - Katayoun Farrahi and Daniel Gatica-Perez

slide-19
SLIDE 19

Reviews

  • Average score: 1.75 (accept)
  • Solid ground truth
  • Privacy not addressed
  • Spelling errors, graphs badly placed
  • No automation, data needs to be manually copied

19

slide-20
SLIDE 20