DEVELOPMENT OF FACE EMOTION RECOGNITION MACHINE LEARNING MODEL FOR AIVUE JOB INTERVIEW APPLICATION

• Main Author: Richardo, Ryo
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76033
• Institution: Institut Teknologi Bandung

The recruitment process is one of the most important stages for a company. The recruitment process takes a relatively long time, especially for the interview stage, thus raising an urgency to improve the efficiency of the recruitment process. Based on that requirement, we developed AIVue, an application that able to manage candidate, do asynchronous interview, and showing candidate’s face emotion from the interview. This paper will discuss the Convolutional Neural Network (CNN) machine learning model such as ZFNet, GoogLeNet, and ResNet which used to extract candidate’s face emotion inside the AIVue application. ResNet has become the favorite model because of its residual block and skip connection that could handle overfitting if using large number of layers. GoogLeNet also estimated to have a similar performance with ResNet because of its inception module that could stack convolutional layers in parallel. Unlike those two models, ZFNet does not have any unique concepts that could improve performance. The result of the testing process consistent with the hypothesis, where ResNet become the best model with 62% accuracy and 52% f1 score, followed by GoogLeNet with 61% accuracy and 51% f1 score. Other than that, every model also spends a relatively short time to predict face emotion in a range of 10 to 11 seconds for a 10-minute interview video. Based on the testing result, ResNet is chosen to be implemented for AIVue application. To further improve the performance for the next development stage, it is recommended to add more variation for the dataset, use the right optimizer and learning rate, and explore more machine learning models which could be a better fit for face emotion recognition.