SVM v2 result

F_.v1.v2.Test

t c=32.0, g=0.125 CV rate=94.8725
Training...
Output model: F_v1.v2.Train.model
Scaling testing data...
Testing...
Accuracy = 46.9498% (1647/3508) (classification)
Output prediction: F_v1.v2.Test.predict


answer
H, A, S, F, P |Predict
48 69 103 27 264 |0
9 7 2 63 78 |1
0 24 46 3 7 |2
6 7 15 46 77 |3
287 187 184 449 1500 |4

=============
F0_.v1.v2.Test

t c=32.0, g=0.125 CV rate=94.7875
Training...
Output model: F0_v1.v2.Train.model
Scaling testing data...
Testing...
Accuracy = 52.1095% (1828/3508) (classification)
Output prediction: F0_v1.v2.Test.predict


answer
H, A, S, F, P |Predict
47 44 50 1 159 |0
8 2 0 46 56 |1
0 23 72 24 12 |2
1 5 12 26 18 |3
294 220 216 491 1681 |4

J48 v1 result

F_v1.csv(10 CV)
=== Run information ===

Scheme: weka.classifiers.trees.J48 -C 0.25 -M 2
Relation: F_v1
Instances: 6274
Attributes: 103
[list of attributes omitted]
Test mode: 10-fold cross-validation

=== Classifier model (full training set) ===

=== Summary ===

Correctly Classified Instances 5823 92.8116 %
Incorrectly Classified Instances 451 7.1884 %
Kappa statistic 0.8832
K&B Relative Info Score 552448.6144 %
K&B Information Score 9940.3112 bits 1.5844 bits/instance
Class complexity | order 0 11285.6028 bits 1.7988 bits/instance
Class complexity | scheme 327256.4897 bits 52.1607 bits/instance
Complexity improvement (Sf) -315970.8869 bits -50.362 bits/instance
Mean absolute error 0.0316
Root mean squared error 0.1657
Relative absolute error 12.8354 %
Root relative squared error 47.2221 %
Total Number of Instances 6274

=== Detailed Accuracy By Class ===

TP Rate FP Rate Precision Recall F-Measure Class
0.923 0.006 0.942 0.923 0.933 _S
0.845 0.012 0.851 0.845 0.848 _A
0.941 0.013 0.925 0.941 0.933 _F
0.861 0.017 0.856 0.861 0.858 _H
0.948 0.07 0.949 0.948 0.949 _P

=== Confusion Matrix ===

a b c d e <-- classified as
553 4 14 1 27 | a = _S
8 382 7 8 47 | b = _A
1 15 859 5 33 | c = _F
1 6 8 570 77 | d = _H
24 42 41 82 3459 | e = _P

Number of Leaves : 278

Size of the tree : 555



================
F0_v1.csv (10 CV)

=== Run information ===

Scheme: weka.classifiers.trees.J48 -C 0.25 -M 2
Relation: F0_v1
Instances: 6274
Attributes: 103
[list of attributes omitted]
Test mode: 10-fold cross-validation

=== Classifier model (full training set) ===

Number of Leaves : 262

Size of the tree : 523


Time taken to build model: 34.48 seconds

=== Stratified cross-validation ===
=== Summary ===

Correctly Classified Instances 5742 91.5206 %
Incorrectly Classified Instances 532 8.4794 %
Kappa statistic 0.8615
Mean absolute error 0.0364
Root mean squared error 0.1782
Relative absolute error 14.7968 %
Root relative squared error 50.7918 %
Total Number of Instances 6274

=== Detailed Accuracy By Class ===

TP Rate FP Rate Precision Recall F-Measure Class
0.93 0.008 0.922 0.93 0.926 _S
0.812 0.012 0.838 0.812 0.825 _A
0.92 0.015 0.911 0.92 0.916 _F
0.802 0.016 0.854 0.802 0.827 _H
0.945 0.092 0.935 0.945 0.94 _P

=== Confusion Matrix ===

a b c d e <-- classified as
557 7 10 1 24 | a = _S
9 367 16 8 52 | b = _A
5 12 840 5 51 | c = _F
1 8 8 531 114 | d = _H
32 44 48 77 3447 | e = _P

SVM v1 result

F_v1.test.5

Best c=128.0, g=0.5 CV rate=93.2173
Training...
Output model: F_v1.train.5.model
Scaling testing data...
Testing...
Accuracy = 94.8791% (1334/1406) (classification)
Output prediction: F_v1.test.5.predict


answer
H, A, S, F, P |Predict
129 0 0 1 13 |0
0 88 1 0 2 |1
1 0 124 0 2 |2
0 1 2 188 3 |3
20 11 3 12 805 |4


=========
F0_v1.test.5

Best c=32.0, g=0.5 CV rate=93.0753
Training...
Output model: F0_v1.train.5.model
Scaling testing data...
Testing...
Accuracy = 94.8791% (1334/1406) (classification)
Output prediction: F0_v1.test.5.predict


answer
H, A, S, F, P |Predict
129 0 0 1 12 |0
0 86 1 0 3 |1
1 0 124 0 2 |2
0 0 2 189 2 |3
20 14 3 11 806 |4

論文開工

最近忙私事
忙到壓力很大,
但是論文總是開工了!
加油吧...

專心做一件事情的感覺
舒服的多
比壓力小的多

Recording Data

Today, I am trying to parse the recording data.
I count the duration of my experiment and the file. I found that they did not match!! There are 16 times than the samples I collect. I need to find out what the others stand for.
Sigh....

[Reading] XPod: a Human Activity Aware Learning Mobile Music Player

http://ebiquity.umbc.edu/paper/html/id/335/XPod-A-Human-Activity-Aware-Learning-Mobile-Music-Player

Sandor Dornbush, Jesse English, Tim Oates, Zary Segall, and Anupam Joshi
Jan 08, 2007

I think the work is an on-going one. In the paper the show on Jan 08, 2007 (a coming day), I think the writers take out the part of emotion, which is the one I am interested in. The try to uses different method, including decision tree, AdaBoost, SVM, KNN, neural networks, to classify 5 different state. They collect GSR, acceleration (2D), skin temperature, BVP, time, song information and beats per minute to predict how a user would rate a song in the future. They have 565 training instances. The result consider states is a little better than the one without states. More precisely speaking, states are considering the physiological information gather from the sensors. The result range is from 31.87% to 46.72, and mean square error is from 0.17 to about 0.45.
I think the result is not very well. I think they should collect more training data. Some more interesting points should be added into XPod.

Before the holiday

What should me do .
-- problem definition
-- complete survey related work
-- solultion (optional)
-- experiment result

[Reading] A Study in Users' Physiological Response to an Empathic Interface Agent

Helmut Prendinger, Christian Becker, and Mitsuru Ishizuka.
A study in users' physiological response to an empathic interface agent. [pdf]
International Journal of Humanoid Robotics, Vol. 3, No. 3, Sept. 2006, pp 371-391.
http://www.worldscinet.com/191/03/0303/S0219843606000801.html


Game: Skip-Bo
Emotion recognition: SC, EMG, & game sate(time) as input, 2-axes (arousal & valence) as output. (ref29)
Method: ANOVA
Subject: 32 (14m, 18f)
Hypothesis:

• If the virtual game opponent behaves ``naturally'' in that if follows its own goals and expresses associated positively or negatively valenced affective behaviors, users will be less aroused or stressed than when the agent does not do so.
• If the game opponent is oriented only toward its own goals and displays associated behaviors, users will be less aroused or stressed than when the agent does not express any emotion at all.
  

Agent's behavior:

• non-emotional
• self-centered emotional
• negative empathic
• positive empathic
  

Agent's abilities: (in PAD space)

• auditory speech
• facial
• body gesture

Result:

• The positive empathic condition was experienced as significantly more arousaling or stressful than the negative empathic condition.
  
• Users seemingly do not respond significantly different when empahtic agent behavior is absent.
  

Notes:

• EMG high, negative valence more
• global baseline -> for individual differences
  

a good word to use

  empathy
      n : understanding and entering into another's feelings

This is what I want the computer to know and to do.

[listen] Brooky's proposal

• Learn ability - features number v.s training set size (how many data should I have to train a trusty model?)
• Technical detail of methods - such as SVN, DBN
• CRF how to use.
  

Go, go, go!

Experiment Prework

The induction has been record as many sound files. I have make some movie clips for different mood induction. I will finish the flash for connecting database and different playing tomorrow or on Thursday.

Something need to do is to ask SY to borrow the bio-sensors. I need time....!!!
So many things need to do.

Experiment Design

The early version has been done, but there still have something needs to be done.

• make the self-evaluation table. By degree
  
• 平靜＆自我評估的順序誰先誰後？還是整個做完才做自我評估
• 表情以及動作的enhancement要如何放入整個場景
• 平靜的詳細指導
• 事後處理的說明
• music choice
  

This is tomorrow's list..

[Reading] From Physiological Signals to Emotion: Implementing and Comparing Selected Methods for Feature Extraction and Classification

Wagner, J. Jonghwa Kim Andre, E.

This paper appears in: Multimedia and Expo, 2005. ICME 2005. IEEE International Conference on
Publication Date: 6-8 July 2005
On page(s): 940- 943
ISBN: 0-7803-9331-7
Digital Object Identifier: 10.1109/ICME.2005.1521579
Posted online: 2005-10-24 09:39:34.0

elicitation method: music
elicited emotion: joy, anger, sadness, pleasure (arousal & valance)
measurement: SC, EMG, ECG, RSP
analysis method: kNN, LDF(linear discriminative, function), multi-layer perceptron
result: 9x%, better then the MIT one.

[Reading] Emotion Analyzing Method Using Physiological State

Mera, Kazuya and Ichimura, Takumi

Book Series	Lecture Notes in Computer Science
Publisher	Springer Berlin / Heidelberg
ISSN	0302-9743
Subject	Computer Science
Volume	Volume 3214/2004
Book	Knowledge-Based Intelligent Information and Engineering Systems
DOI	10.1007/b100910
Copyright	2004
ISBN	978-3-540-23206-3
Pages	195-201

1. using uttrance in dialoge to fine the emotion (by ECG)
2. using the method proposed by R. Picard to recognize the emotion (extrating 20 features)
3. using NN to combine 1 + 2

Condiser 20 emotions which are classified into 6 emotion groups :

• Well-Being: joy, distress
• Fortunes-of-Others: happy-for, gloating, resentment, sorry-for
• Prospect-based: hpe, fear
• Confirmation: satisfaction, relief, fears-confirmed, disappointment
• Attribution: pride, admiration, shame, disliking
• Well-Being/Attribution: gratitude, anger, gratification, remorse
  

A method combine context and physiological signal processing to detect the clients emotion.

[Readings] Emotion recognition form physiological signals using wireless sensors for presence techonologies

Journal	Cognition, Technology & Work
Publisher	Springer London
ISSN	1435-5558 (Print) 1435-5566 (Online)
Subject	Computer Science and Engineering
Issue	Volume 6, Number 1 / February, 2004
Category	Original Article
DOI	10.1007/s10111-003-0143-x
Pages	4-14
Online Date	Thursday, February 19, 2004

Fatma Nasoz¹ , Kaye Alvarez², Christine L. Lisetti¹ and Neal Finkelstein

subject: 29
elicitation method: movie
method: kNN(71%), DFA(74%), MBP(83%)
elicited emotioin: sadness anger surprise fear frustration amusement (by 1995 Gross Levenson)
Measure: GSR, HR, tempreture
Data saving: 3-dimensional array, real number (see in page 12)

[readings] Multimodal Emotion Recognition & Expressivit Analysis

Multimedia and Expo, 2005. ICME 2005. IEEE International Conference on

Kollias, S. Karpouzis, K.


The paper present the framwork of multimodal emotion recognition and expressivity analysis. First, it introduces technologis that involve in emotion reocognition, such as speech, visaul, physiological signal processing, etc. Then, it mentions the important issues in emotional speech analysis, visual analysis, ECA expressivity, and physiological signal analysis in sections. Finally, some key problem identified for integration of sigle-mode emotion analysis techniques are listed.
I think this is a high level paper which shows up and lists the problem in the field of emotion recognition. The cited papers are distribute in three fields, which are speech, facial expression, and physiological signal. I will read more and part 3, the physiological signal process with emotion recognition.

[Readings]Evolving Pet Robot with Emotional Model

Kubota, N. Nojima, Y. Baba, N. Kojima, F. Fukuda, T.

This paper appears in: Evolutionary Computation, 2000. Proceedings of the 2000 Congress on
Publication Date: 2000
Volume: 2, On page(s): 1231-1237 vol.2
ISBN: 0-7803-6375-2
References Cited: 23
Digital Object Identifier: 10.1109/CEC.2000.870791

In daily, many pet robots have been existed. Writers show a pet robot with emotion model. They discribe the details in the paper. First, they have some sensors in the robot. People can interact with the robot via the sensors. The sensors' signal inputs feed into emotional model which is desing by themself. (The model include emotion, mood, and feelings.) Robots are express their emotion by actions such as moving, dancing, and sounding. Moreover, they use potential field to represent a trick. They try to teach the robot the ``trick''. This is the step of behavioral learning. Finally, they use GA to optimal fuzzy controllers and path planning for mobile robots.

In this paper, everything the robot feels is defined by human. I think the writers model the robot's behavior with emotional concept. But all the thing about the model is predefine without animal phsychology. It really work, but it is pre-define. That is another side of emotional using. I am not focusing this part now.

[Readings] Aritficail Emotion and Its Recognition, Modeling and Applications: An Overview

This paper really an overview. The writers using many reference to discribe the whole picture of artificial emotion. They using the view of medical science. But in the book I recently read, it is just one part of theory in the whole emotion psychology. Moreover, they discribe differenct emotion recognition methods. In the emotion modeling, analytic and symbolic method and botom-up method are both mentioned.

I think this is really an overview paper and no technical details inside. What I should do is to read the reference and to find the recognition method.

By the way, it is really strange that the papar is writen in simplify Chinese and published in Proceeding of the 5th World Cngress on Intelligent Control and Automatin. The conference was held in China. I think this is the reason why it is writen in Chinese.

Self-Cam: Feedback from wat would be your social partner

Teeters, Alea and Kaliouby, Rana El and Picard, Rosalind
(MIT Media Lab, Affective computin)

Self-cam can record 24 feature points of personal face and infer the wearer's hidden affective-cognitive states. It can identify 20 facial and head movements and 11 gestures. It uses Bayesian Network to infer the affective-cognitive states. The states includes agreeing, disagree, interested, confused, concentrating and thinking. Wearer can explore his/her apperence in the interaction.

This is a little equipment that can be used to record the feature point. I may used it to detect some personal activities and emotions. As the result, I can provide feed back to help people to do some interaction. (CHI?)

I think it looks like that poeple like to take photos themself. In many camara, the producer provides the fucntion of "self-photoing". It take advantage human's instinct. I think this may be workable in the recognition of emotions.

Ideas about app

• Music can affect the emotion
• Pictures (photo) can, too
• An emotion treatment?
  
• An enviroment make you comfortable
• The ``role'' of emotion in the app -> should approve the system