classification-problems

This repository contains solutions for two Kaggle classification challenges completed as part of the AIT 511 Machine Learning course:

• COPD Risk Prediction (Binary Classification)
• Signal Cluster Classification (Multiclass Classification)

The goal of this project is to develop and compare machine learning models for both binary and multiclass classification tasks, while analyzing the performance of SVMs and Neural Networks across different dataset sizes.

---

Approach

Data Understanding & Preprocessing

Signal Cluster Dataset

The SignalCluster dataset consists of synthetic two-dimensional signals, each represented by:

• signal_strength — magnitude or intensity of the signal
• response_level — reaction or response measurement
• personality_cluster — multiclass target label

Key tasks performed:

• Visualizing clusters in 2D feature space
• Handling class imbalance (if present)
• Preparing data for both linear and nonlinear classifiers

COPD Dataset

This dataset contains clinical, physiological, and lifestyle measurements aimed at predicting whether a patient is at risk of COPD.

Notes:

• Some features required normalization or scaling
• Class imbalance affects performance, so F1 Score is used as the primary metric

Common Preprocessing Steps

• Missing value handling
• Categorical encoding (if applicable)
• Numerical feature scaling using StandardScaler
• Train-test split with stratification where needed

---

Modeling

Both datasets include implementations of:

Support Vector Machines (SVMs)

• Linear SVM
• RBF kernel SVM
• Hyperparameter tuning via Grid Search and cross-validation

Neural Networks (NN)

• Fully connected feed-forward networks
• Tuned hyperparameters such as:
  • Number of layers
  • Activation functions
  • Learning rate
  • Batch size and epochs

Additional COPD Requirement

For the binary COPD task:

• Created a 20% reduced subset of the data
• Trained SVMs and NNs on:
  • The full dataset
  • The reduced dataset
• Compared:
  • Generalization capability
  • Impact of dataset size
  • Sensitivity to data scarcity

---

Evaluation

Signal Cluster (Multiclass)

• Macro F1 Score
• Accuracy
• Visual comparison of linear vs nonlinear decision boundaries

COPD (Binary)

• F1 Score (primary metric)
• Precision, Recall, and Accuracy

Comparison of model performance between the full and reduced datasets