models/unsupervised/dimred/README.md

Dimensionality Reduction Models

This directory contains Python scripts defining dimensionality reduction techniques (e.g., PCA, t-SNE, UMAP). Each model file sets up a scikit-learn–compatible estimator or follows a similar interface, making it easy to swap in train_dimred_model.py.

Key Points:

• Estimator: Typically supports .fit_transform(X) for dimension reduction.
• Default Settings: e.g., PCA might default to n_components=2; t-SNE might set n_components=2 and perplexity=30; UMAP might define n_neighbors=15 or n_components=2.
• No Supervised Tuning: Usually we pick hyperparameters based on interpretability or domain. A manual approach or specialized metric can be used if needed.

Note: The train_dimred_model.py script handles dropping columns, label encoding, performing .fit_transform(X), and optionally saving a 2D/3D scatter plot if --visualize is used.

Available Dimensionality Reduction Models

• PCA
• t-SNE
• UMAP

Usage

To reduce data dimensions:

python scripts/train_dimred_model.py \
  --model_module pca \
  --data_path data/breast_cancer/data.csv \
  --select_columns "radius_mean, texture_mean, area_mean, smoothness_mean" \
  --visualize

This:

1. Loads pca.py, which defines a PCA(n_components=2) estimator by default.
2. Applies .fit_transform(...) to produce a 2D embedding.
3. Saves the model (dimred_model.pkl) and the transformed data (X_transformed.csv).
4. If --visualize is set and n_components=2, it scatter-plots the result.