Algoritmos de clasificación de emociones en base a registros multicanal.
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La presente memoria de título forma parte del proyecto FONDECYT 13220040, EMOCREA: descubre y recrea tus emociones en entornos de realidad virtual mediante técnicas neurocientíficas y de inteligencia artificial, cuyo objetivo es determinar las categorías emocionales más salientes para su posterior entrenamiento en medios educativos. El objetivo general de esta memoria es clasificar emociones a partir de señales fisiológicas, como GSR, PPG, respiración y temperatura corporal, las cuales son de carácter mínimamente invasivo y de fácil adquisición. La motivación de este trabajo radica en la importancia de clasificar emociones humanas con precisión, con aplicaciones directas en el ámbito educativo. Para ello, las señales fisiológicas ofrecen un enfoque más certero y objetivo en comparación con métodos como las expresiones faciales.
La metodología incluyó un experimento con 111 estímulos visuales clasificados en clústeres para evocar emociones. 64 participantes, con una media de edad de 15 años, calificaron cada estímulo en 3 categorías: aproximación, compromiso e identidad, utilizando una escala de 1 a 9. De estos, se seleccionaron 54 registros tras verificar la calidad. El procesamiento abarcó filtrado de ruido, normalización y segmentación de señales.
Se realizó un análisis estadístico de los datos conductuales obtenidos por cada participante, tanto de las respuestas proporcionadas durante el experimento como de los tiempos de respuesta. Adicionalmente, se aplicó un análisis de ANOVA para identificar diferencias significativas entre los datos. En los casos donde estas diferencias fueron significativas, se realizó un análisis posterior para identificar los clústeres entre los cuales las diferencias eran mayores.
Las señales fueron analizadas mediante algoritmos de Machine Learning, entre ellos árbol de decisión, regresión lineal, Naïve Bayes y SVM. Cada modelo fue evaluado utilizando métricas de pre dictibilidad positiva +P y sensibilidad Se, calculadas a partir de matrices de confusión. Los resultados mostraron que la efectividad de los modelos varió según la señal y el clúster de clasificación, alcanzando un desempeño aproximado del 33.8% en predictibilidad positiva y un 33.48% en sensibilidad, para los 3 clústeres.
En conclusión, el algoritmo desarrollado no logró clasificar emociones de manera efectiva utilizando características fisiológicas. Sin embargo, se identificaron desafíos importantes, como la sincronización de señales y el manejo del ruido en los datos, que representan áreas clave para futuras investigaciones.
This thesis is part of the FONDECYT project 13220040, EMOCREA: discover and recreate your emotions in virtual reality environments using neuroscientific and artificial intelligence techniques, which aims to determine the most prominent emotional categories for subsequent training in edu cational contexts. The primary objective of this report is to classify emotions based on physiological signals such as GSR, PPG, respiration, and body temperature, which are minimally invasive and easily obtainable. The motivation behind this work stems from the significance of accurately classifying hu manemotions, with direct applications in the educational sector. To achieve this, physiological signals provide a more precise and objective approach compared to methods like facial expressions. The methodology included an experiment with 111 visual stimuli classified into clusters to evoke emotions. 64 participants, with an average age of 15 years, rated each stimulus in 3 categories: ap proach, engagement, and identity, using a scale from 1 to 9. From these, 54 records were selected after verifying quality. The processing involved noise filtering, normalization, and signal segmentation. A statistical analysis was conducted on the behavioral data obtained from each participant, both from the responses provided during the experiment and from the response times. Additionally, an ANOVA analysis was applied to identify significant differences among the data. In cases where these differences were significant, a subsequent analysis was performed to identify the clusters among which the differences were greatest. The signals were analyzed using Machine Learning algorithms, including decision trees, logistic regression, Naïve Bayes, and SVM. Each model was evaluated using positive predictability metrics +P and sensitivity Se, calculated from confusion matrices. The results showed that the effective ness of the models varied according to the signal and classification cluster, achieving an approximate performance of 33.8% in positive predictability and 33.48% in sensitivity for the 3 clusters. In conclusion, the developed algorithm did not effectively classify emotions using physiological features. However, significant challenges were identified, such as signal synchronization and noise management in the data, which represent key areas for future research.
This thesis is part of the FONDECYT project 13220040, EMOCREA: discover and recreate your emotions in virtual reality environments using neuroscientific and artificial intelligence techniques, which aims to determine the most prominent emotional categories for subsequent training in edu cational contexts. The primary objective of this report is to classify emotions based on physiological signals such as GSR, PPG, respiration, and body temperature, which are minimally invasive and easily obtainable. The motivation behind this work stems from the significance of accurately classifying hu manemotions, with direct applications in the educational sector. To achieve this, physiological signals provide a more precise and objective approach compared to methods like facial expressions. The methodology included an experiment with 111 visual stimuli classified into clusters to evoke emotions. 64 participants, with an average age of 15 years, rated each stimulus in 3 categories: ap proach, engagement, and identity, using a scale from 1 to 9. From these, 54 records were selected after verifying quality. The processing involved noise filtering, normalization, and signal segmentation. A statistical analysis was conducted on the behavioral data obtained from each participant, both from the responses provided during the experiment and from the response times. Additionally, an ANOVA analysis was applied to identify significant differences among the data. In cases where these differences were significant, a subsequent analysis was performed to identify the clusters among which the differences were greatest. The signals were analyzed using Machine Learning algorithms, including decision trees, logistic regression, Naïve Bayes, and SVM. Each model was evaluated using positive predictability metrics +P and sensitivity Se, calculated from confusion matrices. The results showed that the effective ness of the models varied according to the signal and classification cluster, achieving an approximate performance of 33.8% in positive predictability and 33.48% in sensitivity for the 3 clusters. In conclusion, the developed algorithm did not effectively classify emotions using physiological features. However, significant challenges were identified, such as signal synchronization and noise management in the data, which represent key areas for future research.
Description
Tesis presentada para optar al título profesional de Ingeniero Civil Biomédico
Keywords
Algoritmos, Aprendizaje de máquina, Inteligencia artificial