# Knowledge base: Warsaw University of Technology

Back

## Detecting emotions in physiological signals using detrended fluctuaction analysis and machine learning

#### Abstract

Title of the thesis: Detecting emotions in physiological signals using detrended fluctuation analysis and machine learning. The point of the thesis is the question: “can we described what emotions felt participant during experiment using phycological signals?”. To find an answer, it is used machine learning (classification) to analyze data from Cybermotions experiment. At the beginning it is mentioned what classification is and descripted used classification models: k-nn, LDA, QDA and decision tree. Next there are shown measure rates to compare classifiers: accuracy, precision, recall, F1 and matthews correlation coefficient (MCC). Analyzed data are from part of the experiment where IAPS images were shown and participants respond to three questions about their emotional state in scale 1-7. During the experiment have been measured: feet sweatiness, pulse and movement of muscles near mouth and eyebrows. Used notes in thesis are: positive, average and arousal. To minimize vector’s length such features were counted: maximum, minimum, difference between maximum and minimum, mean, standard deviation, skew, kurtosis and DFA. There are described divisions from notes in scale 1-7 into two classes, research methodology and results. Scale is divided for three methods: GE3 (if note >=3: class +1, else: class -1), GE4 and GE5 (GE4, GE5 like GE3). Models have been trained 150 times for every division (every model from another dataset made by permutation of start dataset). The results of three measure rates (accuracy, f1, mcc) are shown on boxplots on pages: 29, 30 and 31. In 7 out of 9 cases by MCC rate the best classifier was decision tree. The most efficiency classification was decision tree for positive note GE5 with measure rates: accuracy = 0.58, f1 = 0.51, mcc = 0.16. The worst result has QDA. When numbers of vectors with positive notes decreased (GE5 > GE4 > GE3) quality of classification increased in positive and arousal scale. It could be concluded that positive and arousal emotions are easier to detect. There was build a program to show generated models with possibility of seeing raw data on charts, counted features, boxplots that compare models and confronting predictions with real classes. According to the research made on data from Cybermotions experiment the model of decision tree has been made. The classifier is able to detect human emotional state (positive – negative) based on measured phycological data.
Diploma type
Engineer's / Bachelor of Science
Diploma type
Engineer's thesis
Author
Title in Polish
Detekcja emocji w fizjologicznych szeregach czasowych z wykorzystaniem analizy fluktuacji oraz uczenia maszynowego
Supervisor
Jan Chołoniewski (FP/LPESS) Jan Chołoniewski,, Center of Physics in Economics and Social Sciences (FP/LPESS)Faculty of Physics (FP)
Janusz Hołyst (FP/LPESS) Janusz Hołyst,, Center of Physics in Economics and Social Sciences (FP/LPESS)Faculty of Physics (FP)
Certifying unit
Faculty of Physics (FP)
Affiliation unit
Center of Physics in Economics and Social Sciences (FP/LPESS)
Study subject / specialization
, Fizyka Techniczna
Language
(pl) Polish
Status
Finished
Defense Date
15-02-2019
Issue date (year)
2019
Reviewers
Julian Sienkiewicz (FP/LPESS) Julian Sienkiewicz,, Center of Physics in Economics and Social Sciences (FP/LPESS)Faculty of Physics (FP) Jan Chołoniewski (FP/LPESS) Jan Chołoniewski,, Center of Physics in Economics and Social Sciences (FP/LPESS)Faculty of Physics (FP)
Keywords in Polish
Uczenie maszynowe, klasyfikacja, wykładnik skalowania fluktuacji, emocje
Keywords in English
Machine learning, classification, scale fluctuation exponent, emotions
Abstract in Polish
File
• File: 1
urn:pw-repo:WUTd0ffb059296f42e996abdbad9aab6405