Installs into .claude/skills of the current project.
Are you the author of 572 Usage Guide Fcc82e24?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/tools-only-572-usage-guide-fcc82e24)
# Classification Models Usage Guide
## Overview
Build classification models for biomarker discovery and diagnostics using RandomForest, XGBoost, and logistic regression with sklearn-compatible APIs.
## Prerequisites
```bash
pip install scikit-learn xgboost pandas matplotlib
```
## Quick Start
Tell your AI agent what you want to do:
- "Build a random forest classifier on my expression data"
- "Train an XGBoost model to predict disease status from my omics data"
- "Use logistic regression with L1 penalty to find sparse biomarkers"
- "Evaluate my classifier with ROC curves and AUC"
## Example Prompts
### Basic Classification
> "Train a random forest classifier on my gene expression matrix (expression.csv) with labels in metadata.csv. Show me the ROC curve and top 20 important features."
> "Build an XGBoost classifier to predict tumor vs normal from my RNA-seq counts."
### Biomarker Selection
> "Use L1-regularized logistic regression to find a sparse set of genes that classify my samples. Which genes have non-zero coefficients?"
### Multi-class
> "I have expression data with 4 cancer subtypes. Train a classifier and show the confusion matrix."
### Model Comparison
> "Compare RandomForest, XGBoost, and logistic regression on my dataset. Which has the best AUC?"
## What the Agent Will Do
1. Load expression matrix and sample labels
2. Split data into train/test sets (stratified)
3. Build preprocessing pipeline with scaling
4. Train classifier and generate predictions
5. Report metrics (accuracy, AUC, classification report)
6. Visualize ROC curve and feature importances
## Tips
- Always use stratified splits for imbalanced classes
- StandardScaler is essential for logistic regression; optional for tree models
- Use class_weight='balanced' when classes are imbalanced
- XGBoost parameters: use random_state not seed, n_jobs not nthread
- L1 penalty gives sparse solutions; L2 handles correlated features better
- For small datasets, consider nested cross-validation instead of train/test split
