I am a Master's student in Data Science at Lingnan University, Hong Kong, expected to graduate in July 2026. I have a strong passion for leveraging data-driven approaches to solve real-world problems, particularly in healthcare and smart tourism. With hands-on experience in deep learning frameworks like PyTorch, I have successfully delivered two significant projects: a CNN-based brain tumor classification system and an intelligent tourist diversion system. I am a proactive learner, team player, and results-oriented individual seeking opportunities to apply my skills in data science, big data analytics, and machine learning engineering.
Problem: Traditional brain tumor diagnosis relies heavily on radiologists' experience and is time-consuming. This project aims to develop an automated, accurate multi-class classification system for brain MRI images to assist clinical diagnosis.
Data: Brain Tumor MRI Dataset from Kaggle — 7,200 MRI images across 4 categories: Glioma, Meningioma, Pituitary Tumor, and No Tumor. Training set: 5,600 images, Testing set: 1,600 images (balanced distribution).
Method: Transfer learning using ResNet18 pre-trained on ImageNet. Applied data augmentation (random horizontal flip, rotation). Used cross-entropy loss and Adam optimizer. Built with PyTorch framework.
Outcome: Achieved 98% overall accuracy, with macro-average precision, recall, and F1-score all at 98%. AUC scores near 1.0 for all categories, demonstrating excellent generalization. Confusion matrix analysis showed minimal misdiagnosis risk.
My Contribution: As the team leader, I designed the model architecture, implemented the training pipeline, performed hyperparameter tuning, and evaluated model performance. I also contributed to code writing and team coordination.
🔗 View on GitHubProblem: Overcrowding in popular scenic spots during holidays degrades visitor experience and increases management pressure. Traditional manual scheduling cannot achieve dynamic, personalized route optimization.
Data: Simulated tourist data (500 visitors) including age, group size, budget, preferences, and planned duration. Scenic spot data includes 4 POIs (Waterfall, Temple, Flower Sea, Pavilion) with time and capacity constraints.
Method: Hybrid approach combining CSP (Constraint Satisfaction Problem) for feasible route generation, Minimax adversarial search for balancing tourist preferences and congestion control, Machine Learning (Decision Tree, Random Forest, Logistic Regression) for preference prediction, and Deep Learning (LSTM, GNN, Wide&Deep) for flow prediction and personalization.
Outcome: System successfully generates personalized, congestion-aware routes. Achieved real-time recommendations with balanced utility scores. Wide&Deep model provides accurate preference scoring. LSTM predicts crowd flow for proactive diversion.
My Contribution: As a key team member, I contributed to code design, model integration, and system implementation. I also prepared presentation materials and coordinated team efforts.
🔗 View on GitHub