📊 Cross-Validation in Machine Learning

Welcome to this repository about Cross-Validation 📂, an essential technique in Machine Learning for evaluating a model's generalization performance. You'll find theory and a hands-on Python exercise.

---

🧭 Table of Contents

1. What is Cross-Validation?
2. Generalization and Expected Risk
3. Types of Validation
4. Practical Exercise
5. Advantages and Disadvantages
6. Presentation
7. Contributors

---

📌 What is Cross-Validation?

Cross-Validation is a technique to estimate how well a machine learning model will perform on unseen data. It splits the dataset into "folds" and repeatedly trains on some folds while testing on others.

👉 Goal: Minimize generalization error (Expected Risk) and prevent overfitting.

---

📚 Generalization and Expected Risk.

When we train a model, we want it to perform well not just on training data but also on new, unseen data — this is called generalization.

Concept	Definition
📈 Expected Risk	The model’s error on new, unseen data.
🏋️ Empirical Risk	The model’s error on training data.
🔥 Overfitting	When a model has low training error but high error on new data.
🔍 Cross-Validation	A technique to estimate the expected risk by splitting data and averaging results.

---

🧩 Types of Validation

Type	Description	Pros	Cons	When to Use
Hold-Out	Splits dataset into training and validation sets.	Simple, fast.	Sensitive to split randomness.	Large datasets or quick evaluations.
K-Fold (K=5)	Splits into K parts; each fold is test once.	Balances bias and variance.	Can be slow on large models.	Robust estimates without extreme costs.
Stratified K-Fold	Like K-Fold but preserves class proportions.	Respects class imbalance.	Only for classification.	Classification tasks with imbalanced classes.
LOOCV (Leave-One-Out)	Each point is a test set once.	Almost unbiased.	Very costly, high variance.	Small datasets or theoretical cases.
Group K-Fold	Like K-Fold but keeps groups (users, patients...) intact.	Prevents data leakage within groups.	Depends on group sizes.	Data with group dependencies.
Time Series Split	Respects time order; uses past data to predict future.	Essential for time series.	Less data for training at the start.	Forecasting and causal models.
Nested CV	Two-level validation: outer loop for model eval, inner for hyperparameter tuning.	Rigorous, prevents data leakage.	Very computationally expensive.	Model comparison or tuning in research.

---

🧪 Practical Exercise

• This repository includes a hands-on Cross-Validation exercise using the Iris dataset from Scikit-learn.
• ✅ Don’t forget to run the file and explore results across different models!

---

👍 Advantages and Disadvantages

✅ Advantages:

• More robust performance estimation.

• Reduces the risk of overfitting.

• Useful for small to medium datasets.

⚠️ Disadvantages:

• More computationally expensive.

• Slower for large datasets and models.

• May need adaptation for specific data types (time series, classification, etc.).

---

🎬 Presentation

You can access the presentation on Cross Validation at the following Canva link:

Cross Validation

---

👥 Contributors

• This repository was created by:

• Fernando García

• Nhoeli Salazar

---

Thank you for visiting! If you find this helpful, give it a ⭐️ on the repo. 🚀