Collection of ML/AI projects I worked on while learning. Each project taught me something different.
DeepLearning-Portfolio/ ├── graph-neural-networks/ │ ├── pancreatic-cancer-kg/ # Relational GAT (data leakage lol) │ ├── fake-news-detection/ # GAT experiments │ ├── tv-show-recommendation/ # Graph-based recommender │ └── steam-game-recommendation/ # Steam game recs │ ├── computer-vision/ │ ├── brain-tumor-classification/ # First time trying GradCAM │ ├── covid-analysis/ # Just for fun experiment │ └── car-recognition/ # First few-shot learning attempt │ ├── competitions/ │ ├── house-prices/ # Kaggle house prices │ ├── playground-series/ # Kaggle playground │ └── recent-competition/ │ └── data-analysis/ ├── microsoft-malware-prediction.ipynb # ⭐ SHOWCASE - 6 hours of pure analysis └── europa-salary-analysis/ # First visualization practice
My career-defining project. Spent 6 hours with pen and paper analyzing this dataset from Kaggle competition. This analysis changed my career direction back to data. Pure tabular data work - no fancy models, just deep understanding.
Pancreatic Cancer KG - Relational GAT for biomedical knowledge graph. Got 100% accuracy (definitely data leakage, need to fix). But learned a lot about graphs. These projects show I tried different things - graphs give best generalization but worst hardware performance. Very advanced topic. Most code here is AI-assisted.
Fake News Clickbait Detection - GAT + triplet loss experiments. Compared with classics like Naive Bayes.
TV Show - Graph-based rec systems. Mixed SBERT with manual features. Made cool network visualizations.
Brain Tumor Classification (BRICS 2025) - First time trying GradCAM. That's why I included this. MRI classifier with ConvNeXt, 97% accuracy. Medical imaging is interesting.
Car Recognition - First time solving few-shot learning. That's the reason I kept this one. Transfer learning experiments.
COVID Analysis - Just for fun. Experimented with CNN architectures to see what works.
Note: Competition folders are messy on purpose - it's a testing ground where the only goal is best score. That's the nature of competitions.
House Prices, Playground Series, Datathon - These show how obsessive you can get in competitions (reading 2 papers for 0.01 F1 improvement lol). But competitions are fun and you learn a lot about squeezing every drop of performance.
Recent Competition - CatBoost with heavy feature engineering.
To be honest, I forgot which competitions they were, so that's why the folder names are like that. I just picked them pretty randomly. (I wonder which one is my “recent competition”)
Microsoft Malware Prediction - See showcase section above. This is my best work (not technically, but in terms of the impact it leaves).
Europe Salary Analysis - First time doing visualization seriously. Learned different plotting libraries and techniques.
- Deep Learning: PyTorch, PyTorch Geometric
- ML: Scikit-learn, CatBoost, XGBoost and LightGBM
- NLP: SBERT, Transformers
- Data: Pandas, NumPy
- Viz: Matplotlib, Seaborn
- Graphs: Best generalization, worst hardware performance. Very advanced.
- Competitions: Teach you to be obsessive about details. Fun but brutal.
- Few-shot learning: Solved it first time in car recognition project.
- GradCAM: First experimented with it in brain tumor project.
- Tabular data: Microsoft Malware showed me this is where I want to focus.
All the projects here are learning journeys. My goal is not to showcase my work, but to share my learning journey. This means you can find everything here, including code errors, AI comments, etc.
For more: github.com/Tunahanyrd
Thought does not require language. Language is an expression of thought. Intelligence requires thought more than it requires language. ~Yann LeCun