MLPipelineOptimizationStudy: End-to-End ML Pipeline Exploration

MLPipelineOptimizationStudy is a rigorous, reproducible investigation into the art of building high-performance machine learning pipelines. Each stage of the pipeline, from raw data to final predictions, is analysed independently and in combination to understand where gains are achieved.

Pipeline Stages Covered

Data Preprocessing

• Missing value imputation strategies (mean, median, KNN, iterative).
• Outlier detection and treatment.
• Class imbalance handling (SMOTE, class weights).
• Feature scaling (Standard, MinMax, Robust).

Feature Engineering

• PCA for dimensionality reduction with explained-variance analysis.
• t-SNE for high-dimensional data visualisation.
• Manual feature construction and selection (correlation, mutual information, RFE).

Model Selection

Systematic comparison of:

• Logistic Regression, Kernel SVM
• Random Forest, XGBoost, LightGBM, CatBoost
• Voting Ensembles (hard and soft)

Hyperparameter Tuning

• GridSearchCV and RandomizedSearchCV with cross-validation.
• Analysis of resource-accuracy trade-offs across search strategies.

Technology

Python, scikit-learn, XGBoost, LightGBM, CatBoost, Matplotlib, Seaborn. All experiments in reproducible Jupyter Notebooks.