Data_projects_TripleTen

This repository contains selected data science and machine learning projects I completed as part of the TripleTen bootcamp. Each project solves a real-world business challenge using data-driven techniques.

---

Table of Contents

1. Age Verification with Computer Vision
2. IMDB Sentiment Analysis
3. Sweet Lift Taxi Demand Forecasting
4. Insurance Benefit Prediction & Privacy

---

1. Age Verification with Computer Vision

Goal: Build and evaluate a model to estimate a person’s age from a facial photograph to help Good Seed supermarket comply with alcohol sales laws.
Metric: MAE ≤ 8
Best Model: ResNet50-based CNN (MAE = 6.53)
Techniques: EDA, ImageDataGenerator, ResNet50 with GlobalAveragePooling & Dropout, trained on GPU.
Conclusion: The ResNet50 model surpassed the benchmark with MAE = 6.53. It generalized well across age groups, especially near critical legal thresholds like 18 years.

---

2. IMDB Sentiment Analysis

Goal: Classify IMDB movie reviews as positive or negative.
Metric: F1 Score ≥ 0.85
Best Model: NLTK + TF-IDF + Logistic Regression (F1 = 0.89)
Other Models: spaCy + TF-IDF + Logistic Regression (F1 = 0.88), spaCy + LGBM (F1 = 0.83)
Exclusion: BERT was tested on a small subset but excluded due to performance issues.
Conclusion: NLTK-based preprocessing with TF-IDF and Logistic Regression yielded the best results. The model met the F1 score requirement and maintained strong generalization.

---

3. Sweet Lift Taxi Demand Forecasting

Goal: Predict the number of taxi orders for the next hour.
Metric: RMSE ≤ 48
Best Model: Tuned CatBoost (RMSE = 43.12)
Other Strong Models: LightGBM (RMSE = 43.26), Random Forest (RMSE = 46.08), SARIMA (RMSE = 47.67)
Techniques: Time series decomposition, lag features, rolling means, SARIMA, ensemble models with hyperparameter tuning.
Conclusion: CatBoost was the top-performing model, generalizing well across time with minimal error.

---

4. Insurance Benefit Prediction & Privacy

Goal: Solve four ML tasks for Sure Tomorrow Insurance:

1. Customer Similarity: PCA + kNN revealed patterns in income/age/family size; scaling was essential for clustering.
2. Eligibility Classification: Scaled kNN reached F1 = 0.943, outperforming dummy models even under class imbalance.
3. Benefit Count Prediction: Linear regression with and without scaling achieved RMSE = 0.36; age was the top predictor.
4. Privacy-Preserving Obfuscation: Feature matrix transformation using invertible matrix P retained exact model performance post-obfuscation.
   Conclusion: Machine learning supports marketing, eligibility prediction, and secure data transformation. Model performance was preserved even after applying obfuscation, ensuring ethical AI practices.

---

📬 Contact

Feel free to reach out or explore more of my work:
📧 aa.mutapcic@gmail.com | LinkedIn