A COMPARATIVE STUDY OF MACHINE LEARNING MODELS FOR SENTIMENT ANALYSIS

Abstract

A comparative study investigates five models—Support Vector Machine with Histogram of Oriented Gradients (SVM with HOG), Custom Convolutional Neural Network (Custom CNN), LeNet-5, VGG16, and MobileNetV2—for classifying seven facial emotions (Angry, Disgust, Fear, Happy, Sad, Surprise, Neutral) on CK+48 and FER2013 datasets. The analysis assesses accuracy, F1-scores, and computational efficiency, tackling FER2013’s class imbalance (547 Disgust vs. 8,989 Happy samples) and noise. MobileNetV2 led FER2013 performance with 67.82% accuracy (F1-score: ~0.66), utilizing focal loss, Cutout, and Mixup to boost Disgust’s F1-score (~0.60). With ~2.4 million parameters and ~3-hour training, it suits real-time applications like mobile mental health monitoring or driver safety systems. Custom CNN achieved 99.32% accuracy (F1-score: ~0.99) on CK+48, leveraging the dataset’s 981 high-quality, balanced images, making it ideal for controlled settings like psychological research labs. VGG16 attained 67% accuracy (F1-score: ~0.64) on FER2013, benefiting from transfer learning but hindered by overfitting due to ~14.7 million parameters and ~4-hour training. SVM with HOG scored 64.86% accuracy, offering speed (~10 minutes) and noise robustness (~1.5% accuracy drop with Gaussian noise) but limited by handcrafted features. LeNet-5, with 49.47% accuracy (F1-score: ~0.45), struggled with FER2013’s noise and imbalance, highlighting shallow models’ inadequacy. FER2013’s low resolution (48x48) and imbalance caused errors in Disgust and Fear (F1-scores: ~0.50–0.60), driven by low samples and visual similarities (e.g., Fear misclassified as Sad/Surprise). The study emphasizes dataset quality, model complexity, and optimizations for effective FER. Future research should explore diverse datasets (e.g., AffectNet), Vision Transformers, video-based FER with 3D-CNNs, and ethical considerations like bias mitigation and federated learning to ensure fairness and enhance applications in healthcare, education, and human-machine interaction.