Why proper evaluation prevents overfitting in ML Python - Why Metrics Matter

Metrics & Evaluation - Why proper evaluation prevents overfitting

Which metric matters and WHY

To prevent overfitting, we focus on validation loss and validation accuracy. These metrics show how well the model works on new, unseen data. If the model does great on training data but poorly on validation data, it means it memorized the training set and won't work well in real life.

Confusion matrix example

    Actual \ Predicted | Positive | Negative
    -------------------|----------|---------
    Positive           |    40    |   10    
    Negative           |    5     |   45    
  

This confusion matrix shows a model tested on new data. It has some mistakes (10 false positives and 5 false negatives). If the model had zero mistakes on training but many on validation, it is likely overfitting.

Precision vs Recall tradeoff

Overfitting can cause high precision but low recall on new data. For example, a spam filter that perfectly blocks spam in training but misses many spam emails in real use has low recall. Proper evaluation helps find this problem early.

Good vs Bad metric values

Good: Training accuracy = 90%, Validation accuracy = 88% (close values show model generalizes well)

Bad: Training accuracy = 99%, Validation accuracy = 60% (big gap means overfitting)

Common pitfalls

• Accuracy paradox: High accuracy can be misleading if data is imbalanced.
• Data leakage: Using validation data in training hides overfitting.
• Overfitting signs: Big difference between training and validation metrics.

Self-check question

Your model has 98% accuracy on training but only 60% accuracy on validation. Is it good for production? No. This big gap means the model memorized training data and will likely fail on new data. You need better evaluation and techniques to reduce overfitting.

Key Result

Validation metrics reveal overfitting by showing performance gaps between training and unseen data.

Practice

1. Why is it important to evaluate a machine learning model on data it has never seen before?

easy

A. To check if the model can make good predictions on new data

B. To make the training process faster

C. To increase the size of the training data

D. To reduce the number of features in the data

Solution

1. Step 1: Understand model evaluation purpose

   Evaluating on unseen data tests if the model generalizes beyond training examples.

2. Step 2: Identify the goal of evaluation

   The goal is to ensure the model predicts well on new, real-world data, not just memorized training data.

3. Final Answer:

   To check if the model can make good predictions on new data -> Option A

4. Quick Check:

   Evaluation on new data = Check generalization [OK]

Hint: Test on new data to check real-world performance [OK]

Common Mistakes:

• Thinking evaluation speeds up training
• Confusing evaluation with feature selection
• Believing evaluation increases training data size

2. Which of the following is the correct way to split data for proper model evaluation?

easy

A. Split data into training and testing sets

B. Use only testing data for training

C. Use all data for training and testing

D. Train on test data and evaluate on training data

Solution

1. Step 1: Recall data splitting best practice

   Proper evaluation requires separating data into training and testing sets to avoid bias.

2. Step 2: Identify correct splitting method

   Splitting data into training and testing sets ensures the model is evaluated on unseen data.

3. Final Answer:

   Split data into training and testing sets -> Option A

4. Quick Check:

   Data split = Training + Testing [OK]

Hint: Always keep test data separate from training data [OK]

Common Mistakes:

• Training and testing on the same data
• Using test data for training
• Not splitting data at all

3. Consider this Python code snippet for evaluating a model:

from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score

X = [[1], [2], [3], [4], [5], [6]]
y = [0, 0, 0, 1, 1, 1]

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42)
model = LogisticRegression()
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
print(round(accuracy, 2))

What will be the printed accuracy?

medium

A. 0.0

B. 0.67

C. 0.5

D. 1.0

Solution

1. Step 1: Understand data split and model training

   Data is split with 33% test size, so test set has 2 samples. Model trains on 4 samples.

2. Step 2: Predict and calculate accuracy

   With logistic regression on this split and data, the model predicts 1 out of 2 test samples correctly, giving accuracy 0.5.

3. Final Answer:

   0.5 -> Option C

4. Quick Check:

   Correct predictions / total test samples = 0.5 [OK]

Hint: Check test set size and prediction correctness [OK]

Common Mistakes:

• Assuming accuracy is always less than 1
• Ignoring test set size
• Confusing train and test data

4. You trained a model and got 100% accuracy on training data but only 50% on test data. What is the most likely problem?

medium

A. The training data is too small

B. The model is underfitting the training data

C. The test data is too easy

D. The model is overfitting the training data

Solution

1. Step 1: Analyze accuracy difference

   Perfect training accuracy but low test accuracy suggests the model memorized training data but fails on new data.

2. Step 2: Identify overfitting

   This behavior is classic overfitting, where the model fits training data too closely and does not generalize.

3. Final Answer:

   The model is overfitting the training data -> Option D

4. Quick Check:

   High train accuracy + low test accuracy = Overfitting [OK]

Hint: Big train-test accuracy gap means overfitting [OK]

Common Mistakes:

• Confusing overfitting with underfitting
• Blaming test data difficulty
• Ignoring accuracy differences

5. You have a dataset with 1000 samples. You want to prevent overfitting and get a reliable estimate of model performance. Which evaluation strategy is best?

hard

A. Split data once into 90% train and 10% test

B. Use 5-fold cross-validation to train and evaluate

C. Train on all data and report training accuracy

D. Evaluate only on training data with early stopping

Solution

1. Step 1: Understand evaluation strategies

   Training on all data or evaluating only on training data risks overfitting and unreliable performance estimates.

2. Step 2: Compare splitting vs cross-validation

   Single split may give biased results depending on the split. 5-fold cross-validation averages results over 5 splits, reducing bias and variance.

3. Step 3: Choose best method for reliable evaluation

   5-fold cross-validation is best to prevent overfitting and get trustworthy performance estimates on limited data.

4. Final Answer:

   Use 5-fold cross-validation to train and evaluate -> Option B

5. Quick Check:

   Cross-validation = Reliable evaluation [OK]

Hint: Use cross-validation for reliable performance estimates [OK]

Common Mistakes:

• Using training accuracy as performance
• Relying on a single train-test split
• Ignoring overfitting risks