Comparing Supervised Machine Learning Algorithms for Emotion Recognition Performance

Description

I compared 8 machine learning algorithms for human emotion recognition using Scikit-learn's python library. I used the Extended Cohn-Kanade Dataset (CK+), which is a database developed by Carnegie Mellon Univer- sity and the University of Pittsburgh, and it contains facial expression data that can be used to study automatic facial expression detection. Specifically, I used the Active Appearance Models Landmarks (AAMs), Facial Action Coding System (FACS) Action Unit (AU) Labels, and the emotion labels for the video sequence.

Algorithms

• Linear Discriminant Analysis (LDA)
• Support Vector Machine (SVM)
• Stochastic Gradient Descent (SGD)
• K-Nearest Neighbors k = 5 (KNN5)
• K-Nearest Neighbors k = 10 (KNN10)
• K-Nearest Neighbors k = 15 (KNN15)
• Naive Bayes (NB)
• Decision Tree (DT)
• AdaBoost (AB)
• Label propagation (LP)

Features and Classification

• AAM as feature, emotion label as classification
• AU as feature, emotion label as classfication
• AAM as feature, AU as classification

Results