Machine Learning Fundamentals & Optimization

A comprehensive machine learning portfolio project covering Deep Learning optimization, Supervised Learning (Classification & Regression), and Unsupervised Learning (Clustering).

Table of Contents

1. Part 1: Deep Learning Optimization
2. Part 2.1: Classification (Credit Card Fraud)
3. Part 2.2: Regression (Bikeshare Demand)
4. Part 2.3: Clustering (Spotify Audio Features)
5. Dependencies & Installation

---

Part 1: Deep Learning Optimization

Goal: Train a Convolutional Neural Network (CNN) on the MNIST dataset to reach exactly 0.97 validation accuracy as quickly as possible (measured in wall-clock time).

• Baseline Model: Standard CNN architecture taking ~23.65 seconds.
• Optimization 1 (Sub-Epoch Evaluation): Implementing a custom Keras Callback (TimeAndAccuracyStopper) to evaluate validation accuracy every 50 batches instead of waiting for the epoch to end.
• Optimization 2 (Parameter Reduction): Drastically reducing the number of Conv2D filters and Dense neurons to cut computational overhead, successfully reaching 97% accuracy in 12.33 seconds (processing batches twice as fast at 6ms/step).

---

Part 2.1: Classification (Credit Card Fraud)

Goal: Identify fraudulent credit card transactions in a highly imbalanced dataset (99.8% legitimate vs. 0.2% fraud).

• Dataset: Kaggle Credit Card Fraud Detection.
• Problem: Standard accuracy metrics are useless. A naive baseline model missed 36% of all fraudulent transactions (Recall: 0.64).
• Techniques & Preprocessing: * Implemented Class Weights (class_weight='balanced') to algorithmically penalize the model for missing fraud.
  • Applied SMOTE (Synthetic Minority Over-sampling Technique) to mathematically generate synthetic fraud examples and balance the training data.
• Results: Both SMOTE and Class Weighting models successfully captured 92% of all fraud (Recall: 0.92), proving that handling imbalanced data is strictly necessary for real-world deployment.

---

Part 2.2: Regression (Bikeshare Demand)

Goal: Predict the continuous exact number of daily bike rentals based on environmental and meteorological factors.

• Dataset: UCI Capital Bikeshare Demand.
• Preprocessing: Dropped strictly correlated features to prevent data leakage (casual and registered) and applied One-Hot Encoding to categorical variables like season and weathersit.
• Models Compared: * Linear Regression: Baseline parametric model (MAE: 581.18 bikes, R²: 0.8436).
  • Random Forest Regressor: Complex non-parametric ensemble model (MAE: 452.67 bikes, R²: 0.8762).
• Results: The Random Forest successfully captured non-linear weather relationships (e.g., sharp drops in rentals during snow), reducing daily prediction error by nearly 130 bicycles.

---

Part 2.3: Clustering (Spotify Audio Features)

Goal: Discover hidden groupings (clusters) within a large library of music based purely on mathematical audio profiles (Danceability, Energy, Valence, Acousticness, Instrumentalness), without using genre labels.

• Dataset: Kaggle Spotify Tracks Audio Features (Author: Maharshi Pandya).
• Techniques: * StandardScaler: To ensure equal geometric distance weighting.
  • K-Means Clustering: Segmented the data into 4 distinct clusters.
  • PCA (Principal Component Analysis): Reduced the 5-dimensional audio profile down to 2 dimensions for visual validation.
• Evaluation: * Silhouette Score: 0.268 (Confirmed distinct but realistically overlapping musical genres).
  • Davies-Bouldin Index: 1.204 (Confirmed clusters are compact and well-separated).
• Results: Successfully isolated distinct "vibes" pure through math:
  • Cluster 0: Rock/Metal (High Energy, Low Valence)
  • Cluster 1: Acoustic/Folk (Low Energy, High Acousticness)
  • Cluster 2: EDM/Techno (High Energy, High Instrumentalness)
  • Cluster 3: Pop/Upbeat (High Energy, High Valence)

---

Dependencies & Installation

To run these notebooks, ensure you have the following libraries installed:

pip install pandas numpy matplotlib seaborn scikit-learn imbalanced-learn tensorflow