multi modal affective data analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf ·...

68
MultiModal Affective Data Analytics Mykola Pechenizkiy SDAD 2012 @ ECMLPKDD2012 2 September 2012 Bristol, UK http://www.win.tue.nl/stressatwork

Upload: others

Post on 18-Jun-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Multi‐Modal Affective Data Analytics

Mykola Pechenizkiy

SDAD 2012 @ ECMLPKDD2012 2 September 2012

Bristol, UK

http://www.win.tue.nl/stressatwork

Page 2: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Affective data  Social media

• Social media leads to masses of affective data related to peoples’ emotions, sentiments and opinions

• In the recent past was used mainly for marketing needs– Web analytics  Social Media

• Whatever the incentive was to study this, sentiment classification has become much more accurate

3

Page 3: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Multilingual Sentiment Classification

4

Page 4: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Rule‐based polarity detectionRule‐based emission model: 8 kinds of rules:

Emission 5

Page 5: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

SentiCorrHow much positive and negative content do we read or write?

6

Page 6: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Mobile SentiCorr AppWhat a fantastic idea, now ifThis app is esigned to make someone else (or a computer) read our e‐mails for us and “protect” us from WHAT??? How lazy can we get? Like someone commented on CNN reactions, if we are getting upset by tones and/or scoldings in e‐mails, we certainly have bigger issues that need to be dealt with. C’mon, guys, go invent something useful. Not to mention, does it detect irony? Will it weed out the liars?Pleeeeezzzeeee, what a WASTE OF SOMEONE’S COLLEGIATE TIME AND ENERGY. Don’t we have houses to clean and poor people to feed and old folks to help with their shopping? Go do something useful with your time, inventors of his app!!!

Great idea! Get it on iOSsoon (anonymous)“Stress is often made worse by the anticipation of an unpleasant event and actually dissipated once you tackle the problem directly”

Pamela BriggsBritish Psychological Society

7

Page 7: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

OLAP Style Exploration of Data Summaries

8

Page 8: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Exploration of Individual Cases, e.g. e‐Mails

9

Page 9: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Sentiment vs. Fact Classification  

News in media or business are considered to be sentiment neutral,– but they often contain positive or negative information, e.g.

“You will be fired in 3 months because of the serious budget cuts.”

– no sentiment, but negative informationSimilarly, in work‐related correspondence there could be stressing information:– How can we identify it?

10

Page 10: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Sentiment discovery: State‐of‐the‐art• Sentiment analysis/classification is mature!

– Commercial products, free services, open‐source, variety of apps, evolves in many directions

• Several great overviews:– Sentiment Analysis in Practice ‐ ICDM2011 tutorial by Tiger Zhang (eBay Research Labs)

• http://web.cs.dal.ca/~yongzhen/publication/paper/ICDM2011_SentimentAnalysisInPracticeTutorial.pdf

– Modeling Opinions and Beyond in Social Media by Bing Liu (UIC)

• http://kdd2012.sigkdd.org/sites/images/summerschool/Bing‐Liu.pptx

Page 11: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Outline

• Framework for Stress Analytics:–Data management, OLAP support– Shape‐based Query‐by‐Example

• Stress detection from speech and GSR–Predictive features and classification– From controlled experiments to real life

Page 12: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

What is stress? Is it a bad thing?

Page 13: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress in NL according to Coosto.nl

Not really job related

Page 14: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Impact of Stress at WorkWHO: by 2020 Top 5 diseases will be stress related. USA: health care expenditures are ~50% greater for workers who report high levels of stress at work (J. Occup. Env. Med, 40:843‐854).

the Netherlands: (TNO, 2006; TU/e Cursor 2012):• The direct costs of stress are 4 billion Euro per year.  • Every year 150.000  300.000 employees become ill because of stress at work.

• 1/7 disabled because of stress at work.• In TU Delft, 53% of surveyed students indicated that they experienced huge stress during their studies.

15

Page 15: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

What do organizations try (not) to do?

Discuss psychological load (28%)

Change work processes (17%)

Source: (TNO, dossier Werkdruk)

Extend regulations (9%)

Reduce workload (33%)

Improve managers’ skills (13%)

Improve work/life balance (14%)

Page 16: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

What can go wrong?• They are not always aware of the problem ordon’t know the exact cause

• People do not always want to share what they experience with others

• Not always timely enough• Expensive to organize meeting with psychologists, interventions

• The individual causes are different and notalways well understood

• Giving practical advises is not trivial

Page 17: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Types of Stress and StressorsDifferent types of stress:• Survival stress – a response to a physical danger• Environmental stress ‐ noise, crowding, pressure from work or family

• Internal stress ‐ worrying about things we can't control; putting ourselves in situations we know will cause us stress (addicted to stress – expanding todolist with more and more conference deadlines)

• Fatigue and overwork ‐ in a long term perspective

Stress affects both body and mind

18

Page 18: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Types of Stress and StressorsThree kinds of stress:• Acute: caused by an acute short‐term stress factor.• Episodic acute: occurs more frequently & periodically.• Chronic: caused by long‐term stress factors ‐ harmful.

Factors causing stress@work:• long work hours, work overload, time pressure, difficult, demanding or complex tasks, high responsibility, lack of breaks, lack of training

• conflicts, underpromotion, job insecurity, lack of variety, and poor physical work conditions (limited space, temperature and lighting conditions)

19

Page 19: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Concept

20

Be‐eep!Be‐eep!

Page 20: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

StressAnalyticsMake people aware of their stress and stressors

Overview of stressors Exploration of relations

Access to evidence, i.e. annotated, measured stressEmpowerment by awareness (+ implicit/explicit advice)

Page 21: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Our approach to StressAnalytics

What, When, Where, with Whom

Physiological signs

OLAP cube

Pattern Mining

22

Page 22: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Our approach to Stress Analytics• Make a person aware of what is happening

– how they spend their time and when and from where the stress comes in

• Provide valuable input for pattern mining/knowledge discovery– Much richer data sources

• Visual analytics– Interactive exploration of stress‐related data– Collecting subjective data/labels from a person through the interaction

23

Page 23: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

GSR, temp., voice, heart rate, facial expressions

Physiological signsrelated data

External user‐related data

KPI, E‐mail, calendar, social media, news

environmentExternal 

environment

Zoom in&out, slice&diceOLAP

Zoom in&out, slice&dice

Pattern mining, prediction, query‐by‐exampleData Mining

Feature extraction, peak/change detection, classification

Raw data, objective evidence

GUIExploration, Interaction, Visual Analytics

temperature, lighting, noise, airconditioning

Page 24: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Evidence: physiol. signals & external sourcesGSR, Temperature, Speech, Facial expressions, Sentiment in text

Page 25: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Alignment of Information Sources

• What person reads and writes: SentiCorr• What person does in general according to agenda• Environment context (lighting, noise, temp etc.)• Annotate data from video, sound, text processing, and vital signs

• …• What person does with the computer

– http://wakoopa.com/

Different aspect with pre‐processing, storing, managing

26

Page 26: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress Data Cube/OLAPQuick data summaries wrt predefined dimensions

27

Page 27: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress Analytics Visualization• OLAP‐style exploration: selecting multi‐dimension, zoom‐in, zoom‐out.

• Navigating to the evidences: i.e. raw data:– GSR, skin temperature, speech, and email

• Shape‐based time‐series similarity search–State‐of‐the‐art UCR‐Suite (Keogh et al.)

• Demo:http://www.win.tue.nl:8080/saw_analytics/stress_v

isualization.jsp

Page 28: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

OLAP system, a Star Schema

Page 29: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Shape‐Based Query‐by‐Example 

• Find a similar shape time‐series with s

• Given a subsequence of GSR time series s

Query

Result

Page 30: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Shape‐based QBE• Euclidean Distance:

• Dynamic Time Warping (DTW)

State‐of‐the‐art UCR‐Suite (Keogh et al.)

Page 31: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

How to measure stress

Determine stress levelbased on observed sweat production

32

Page 32: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Detection and Categorization of Stress

Based on GSR data alone ‐ not as easy as the following figure may suggest:

33

Page 33: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Challenges in Stress Detection• All kinds of noise, e.g. loosing contact with the skin

• Activity (exercising) , environment (cold/hot) context and personal differences may impact GSR we observe

34

Page 34: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Interpretation isn’t straightforward

35

Page 35: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Detection as Classification

• Total number of GSR response.• The sum of GSR amplitude.• The sum of rising time response.• The sum of energy response.

• Mean, SD, min and max of GSR.• Mean, min and max of peak height.

GSR features

Page 36: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Adding more data to disambiguate

• Skin and room temperature, noise, accelerometer, voice, face, … 

37

Page 37: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

e.g. activity recognition can helpWriting vs. typing vs. walking vs. teaching vs …

Analyzing accelerometer data only (wrist band) 38

Page 38: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Uncontrolled and semi‐controlled

• Philips Research employees wearing the device during their working hours

• Students passing the written and multiple choice exams

• Students presenting demos/posters with course project results

• More to come via HumanCapitalCare

39

Page 39: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Experiment demo

Page 40: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

41

Page 41: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Measuring GSR in (un)controlled settings

• Philips prototype 

• Self‐made, the LEGO Mindstorms NXT

42

Page 42: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Multi‐Source Affective Data Classification

Stress/Emotion classification from text, GSR & speech

GSR & other sensors

Facial expression analysis

43

Page 43: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Automatic Stress Detectionfeature enrichment ensemble learning

speech GSR

speechfeatures

GSRfeatures

combinefeatures

classification

speech GSR

speechfeatures

GSRfeatures

ensemble

classificationclassification

speech

speechfeatures

classification

GSR

GSRfeatures

classification

speech model GSR model

Page 44: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress and Skin Conductance

StressChanges in Autonomic Nervous System (ANS)

activation of sweat glands

Changes of  the amount of the produced sweatChanges of skin conductance

• Relax  skin is drier skin conductance is lower

• Stress  sweat increasesskin conductance is higher

Page 45: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

GSR features

• Total number of GSR response.• The sum of GSR amplitude.• The sum of rising time response.• The sum of energy response.

• Mean, SD, min and max of GSR.• Mean, min and max of peak height.

Page 46: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Change detection approachOnline settings

Page 47: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Preprocessing steps

50

Page 48: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress and Speech

Stress Respiration Rateincreases

Increased subglottal pressure

Increased Pitch

Voice is a good indicator of stress [scherer, 1986]

Page 49: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Speech Features• Voiced and unvoiced speech

Page 50: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Speech Features• Pitch / Fundamental frequency

Page 51: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Speech FeaturesMel Frequency Cepstral Coefficients (MFCCs)are coefficients that approximate human perception auditory response.

Audio(temporal)

frequency filtered frequency

log frequency

FFT Mel scale filter

logs power

DCT representation

DCTStore the first coefficientsMFCCs

Page 52: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Classification Methods

• Support Vector Machine (SVM) – State of the arts.

• Decision Tree classifier.• K‐means using Vector Quantization (VQ). This method is chosen as a baseline.

• Gaussian Mixture Model (GMM). This method works well for speaker recognition task.

• Change detectors: ADWIN, thresholding

Page 53: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress Dataset

• Three types of GSR patterns.

• First type:• Second type:• Third type:

Page 54: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Aligning of data sources

Instance 1 Instance 2 Instance 3

Instance 1 Instance 2 Instance 3

60 seconds

GSR

speech

Page 55: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress Dataset: Speech Features

Page 56: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress Model using GSR features

• SVM outperformed other methods.• Recognizing light vs heavy workload is harder than between recovery vs heavy workload.

46.12

55.54 53.21

70.5174.9

66.82

79.66 80.72

70.673.45

77.81

62.52

0

10

20

30

40

50

60

70

80

90

Recovery vs workloads Recovery vs heavy workload Light vs heavy workload

Accuracy (p

ercent)

k‐means

GMM

SVM

Decision Tree

10-times 10-fold CV (not subject independent)

Page 57: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress Model using speech features

• SVM outperforms the other classifier.• K‐means and GMM do not perform well for speech.• MFCC is a good indicator for stress detection.

49.6555.39

49.17 50.6

58.82 56.78 59.08

52.3

62.08

92.39 92.56 91.69

55.6

68.86 70.69 71.47

0

10

20

30

40

50

60

70

80

90

100

Pitch MFCC MFCC‐Pitch RASTA

Accuracy (p

ercent)

k‐means

GMM

SVM

Decision Tree

Page 58: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

1‐subject‐leave‐out cross‐validation(subject independent model)

79.66 80.72

70.674.84 75

63.04

0

10

20

30

40

50

60

70

80

90

Recovery vs workloads Recovery vs heavy workload Light vs heavy workload

Accuracy (p

ercent)

GSR Tasks

10‐times 10‐fold CV 1‐Subject‐Leave‐Out CV

62.08

92.39 92.56 91.69

53.04

67.82 70 72.17

0

10

20

30

40

50

60

70

80

90

100

Pitch MFCC MFCC‐Pitch RASTA PLP

Accuracy (p

ercent)

Speech Features

10‐times 10‐fold CV 1‐subject‐leave‐out CV

It is better to address the problem of stress detection using a subject dependent model

Page 59: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Fusion ApproachesFeature

enrichmentEnsemble learning

Page 60: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Fusion of GSR and Speech90.73 91.34

69.04

92.43 92.47

70.17

0

10

20

30

40

50

60

70

80

90

100

MFCC and GSR MFCC‐Pitch and GSR Pitch and GSR

Accuracy (p

ercent)

Enriching Feature Space Logistic Regression as MetaLearner

Light vs. heavy workload, balanced data

Page 61: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Kappa Agreement for Classifiers

• Measure agreement between two model using Cohen’s Kappa test.

• Kappa = 1 complete agreement.• Kappa = 0 complete disagreement.

Page 62: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Stress detection summary• Speech is more reliable (in lab settings) than GSR, but more subject dependent.

• SVM is performing better on both GSR and Speech signal.

• ADWIN & thresholding detectors do well on GSR

• Combining GSR and Speech is not trivial:– Speech and GSR predictions are highly independent (low kappa value)

– This diversity may be exploited with dynamic integrations methods

Page 63: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Further directions• Extend the notion of stress (positive and negative) in the stress analytics framework.

• Stress analytics  affective data analytics• Collect more data to enable OLAP  KDD part of the framework.

• Combine with other signals, such as facial expression, heart rate, nutrition.

• Long path from lab setting to real‐life situation; but both are needed.

Page 64: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Is Acute Stress Good or Bad?

69

Page 65: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

What is the Relaxation Then?

70

Page 66: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Is “Normal” Condition Good or Bad?

What if someone’s patterns looks like NNNNNNNNNNNNNNNN ……

71

Page 67: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Summary• The fun parts come from

– The fact that not much is known about stress–Playing with heterogeneous/multi‐modal data–Multi‐disciplinary (data collection, data management, data mining, visual analytics)

– Engineering approach to data mining

• How to show the utility – i.e. what we do–helps to understand better stress as a phenomenon, and the stressors, and how to  

–helps people at the end 72

Page 68: Multi Modal Affective Data Analytics - win.tue.nlmpechen/talks/sdad2012_invited.pdf · Stress/Emotion classification from ... • K‐means using Vector Quantization (VQ). This method

Take home messages• Lab settings vs. real world• Availability and quality of the signal 

– Voice recorded– Someone’s else voice recorded– Noise and missing data, uncertainty– A person cannot speak (during the meeting while someone else is speaking)

• Ground truth, labels, subjective vs. objective• A large problem space

– If you know how to help us with any part on StressAnalytics – talk to me 73