Comparative Machine Learning and Deep Learning Project | UMBC Graduate Studies | 2023
This project explores facial-emotion classification using multiple machine-learning and deep-learning approaches and examines how predicted emotional states can be incorporated into a simple context-aware recommendation workflow.
Rather than relying on a single modeling technique, the project evaluates several approaches to image classification, including convolutional neural networks and traditional supervised-learning algorithms.
The repository represents graduate-level work completed during my master's studies at the University of Maryland, Baltimore County (UMBC) in 2023.
Historical project note: The underlying notebooks originate from the 2023 academic project. Repository documentation has subsequently been expanded to improve clarity, reproducibility, and technical presentation. The original project history has been retained.
- Explore facial-image data for emotion-classification tasks.
- Compare traditional machine-learning techniques with deep-learning approaches.
- Investigate feature-based and image-based classification strategies.
- Evaluate ensemble-learning approaches.
- Build an application-oriented workflow connecting emotion prediction with song recommendation.
- Develop practical experience across the machine-learning lifecycle.
A CNN-based approach was explored for learning image representations directly from facial-image data.
Relevant notebook: cnn_model-final.ipynb
Support Vector Machine classification was investigated as a traditional supervised-learning approach.
Relevant notebook: saved_svm_model.ipynb
Random Forest models were explored as an ensemble tree-based classification approach.
Relevant notebooks:
random_forest_model.ipynbrandom_forest_model with Vgg.ipynb
Gradient-boosted tree methods were evaluated using XGBoost.
Relevant notebook: XGBoost.ipynb
A separate boosting experiment explores another ensemble-learning strategy for classification.
Relevant notebook: Boosting_model.ipynb
The project also explores combining multiple classification approaches using an ensemble-voting strategy.
Relevant notebook: Ensemble-Voting XG-GG-RF-SVM.ipynb
Relevant notebook: Data Exploration.ipynb
Exploratory analysis supports understanding data distribution, class representation, image quality, preprocessing requirements, and potential sources of bias before model development.
The repository also includes RUN THIS APP.ipynb.
Conceptual workflow:
Facial Image
↓
Image Preprocessing
↓
Emotion Classification
↓
Predicted Emotional Context
↓
Recommendation Logic
↓
Song Recommendation
Data Collection / Input
↓
Data Exploration
↓
Preprocessing
↓
Feature Representation
↓
Model Training
↓
CNN / SVM / RF / XGBoost / Boosting
↓
Model Comparison
↓
Ensemble Evaluation
↓
Application Integration
| Notebook | Purpose |
|---|---|
Data Exploration.ipynb |
Exploratory analysis |
cnn_model-final.ipynb |
Convolutional neural-network experiment |
saved_svm_model.ipynb |
Support Vector Machine experiment |
random_forest_model.ipynb |
Random Forest classification |
random_forest_model with Vgg.ipynb |
Random Forest experiment using VGG-related representations |
XGBoost.ipynb |
XGBoost classification |
Boosting_model.ipynb |
Boosting-based model experiment |
Ensemble-Voting XG-GG-RF-SVM.ipynb |
Multi-model ensemble/voting experiment |
RUN THIS APP.ipynb |
Application-oriented emotion/recommendation workflow |
- Python
- Jupyter Notebook
- Machine Learning
- Deep Learning
- Convolutional Neural Networks
- Support Vector Machines
- Random Forest
- Gradient Boosting
- XGBoost
- Ensemble Learning
- Image Classification
- Exploratory Data Analysis
- Model Evaluation
- Recommendation Workflows
The project demonstrates an early focus on comparative model evaluation, ensemble methods, image-based machine learning, and converting predictive outputs into downstream recommendation workflows.
Emotion-recognition systems have important limitations. Facial appearance does not provide a definitive representation of an individual's emotional state, and models can be affected by dataset composition, demographic representation, image conditions, labeling methodology, and cultural differences.
Systems of this type should therefore be treated as experimental or assistive rather than authoritative assessments of an individual's emotional state.
This project was developed as graduate academic work rather than as a production facial-analysis system. Potential limitations include dataset dependence, class imbalance, variation in image conditions, generalization to unseen populations, and the difference between measured model performance and real-world reliability.
The notebooks preserve the original experimentation workflow. Because this repository originates from a 2023 academic environment, package versions and external datasets may differ from current environments. Dependency and data-access documentation can be expanded later after the original notebooks are verified.
This project was developed during graduate studies at UMBC in 2023 as part of broader work in data science, machine learning, analytics, and applied AI.
It represents an earlier stage in a continuing technical progression from predictive analytics and machine learning toward enterprise automation and AI-assisted decision-support systems.
This repository is provided for educational and technical demonstration purposes. It should not be interpreted as a production-ready emotion-recognition or behavioral-assessment system.