Front-End modeling for emotional state recognition

Abstract

Morphological biometrics have proved to be important contributors to e-security and e-health alike. One of its subdivisions, behavioral biometrics, focuses on understanding an individual based on several activities, in this case, a user with drawing and handwritten tasks. Following up the EMOTHAW methodology, this work focuses on extracting and generating new features from the original raw attributes of the aforementioned database. These techniques range from signal processing, physics, and statistics from which important feature vectors and models are created. The signal processing techniques focuses on calculating the logarithmic energy of a signal and then generate both the spectral and cepstral domains to create new numerical features for the resulting vector. In the physics department, traditional kinematic variables are calculated from the position of each task, and lastly the statistical features includes momentum, descriptive statistics, and the different means available in the literature (arithmetic, harmonic, etc.). Furthermore, the implementation of a combined dimensionality reduction (PCA) and feature selection (a novel correlation-based filtering algorithm) pipeline methodology is key for the continuous improvement of the results presented in this research thesis project, going from 60.5 - 71.6 % to 100% accuracy on a binary problem, and reaching accuracy of 82.45% in a ternary problem. This was accomplished by generating synthetic observations to compensate the imbalance distribution of classes intrinsic to health databases with added Gaussian White Noise to a sample number of real observations. Finally, by implementing a Machine Learning frame of several classification algorithms with the library H2O, a considerable testing efficiency was reached, guarantying the best performance.