Skip to content

README.md

Latest commit

 

History

History
117 lines (80 loc) · 3.35 KB

README.md

File metadata and controls

117 lines (80 loc) · 3.35 KB

Machine Learning Playground

A collection of hands-on machine learning projects focused on model comparison, preprocessing, and real-world constraints.
Built primarily using Jupyter notebooks.

Index

Project 1: Regression Models Comparison

Comparison of multiple regression techniques under a unified pipeline to study the impact of regularization and optimization.

Models: Linear, Ridge, Lasso, Elastic Net, SGDRegressor, Polynomial

Key Work:

  • One-hot encoding + feature scaling (StandardScaler)
  • Unified training + evaluation pipeline
  • GridSearchCV for hyperparameter tuning

Insights:

  • Ridge stabilizes multicollinearity; Lasso performs feature selection
  • Elastic Net balances both worlds effectively
  • Polynomial models overfit quickly without regularization

Code: Regression_Pipeline_ipynb

Project 2: KNN on PIMA Diabetes Dataset

Built a KNN classifier with a focus on data cleaning and imputation for medical data.

Key Work:

  • Handled invalid zero values using median + KNN imputation
  • Scaled features for distance-based learning
  • Train/val/test split with pipeline-based modeling

Results:

  • Accuracy: ~57%
  • Strong class imbalance → poor recall for diabetic cases (~18%)

Insights:

  • Preprocessing > model choice for KNN
  • Distance-based models struggle on imbalanced datasets

Code: KNN.ipynb

Project 3: GloVe Embeddings Visualization

Explored semantic relationships in GloVe word embeddings using dimensionality reduction.

Key Work:

  • Used glove.6B.200d embeddings
  • PCA → t-SNE (2D visualization)
  • Tested multiple perplexity values (25–50)

Insights:

  • Semantic clustering visible at lower perplexities
  • “apple” positioned between fruit + tech contexts
  • Higher perplexity distorted semantic relationships

Code: glove-embeddings.ipynb

Project 4: XGBoost on Malaria Genomics

Baseline classification model for species prediction using SNP data (~48.5M features/sample).

Key Work:

  • MAF filtering to remove uninformative variants
  • 0/1/2 SNP encoding
  • Sparse processing + np.memmap for memory efficiency
  • TruncatedSVD for dimensionality reduction
  • XGBoost classifier

Results:

  • Accuracy: ~80%
  • Macro F1: ~0.60

Insights:

  • Dimensionality reduction is critical in genomics
  • XGBoost handles high-dimensional tabular data well
  • Performance limited by class imbalance + missing classes

Code Repository: XG Boost on Malaria Genomics

Tech Stack

scikit-learn • numpy • pandas • matplotlib • XGBoost • PyTorch

Notes

This repository focuses on learning through implementation, with emphasis on:

  • clean pipelines
  • fair model comparison
  • practical constraints (memory, data quality, imbalance)