Cross-validation (K-fold) in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Cross-validation (K-fold)

Which metric matters for Cross-validation (K-fold) and WHY

Cross-validation helps us check how well a model will work on new data. The main metric we look at is the average performance across all folds, like average accuracy or average error. This shows if the model is stable and not just lucky on one set of data.

We also look at the variation (like standard deviation) of the metric across folds. Low variation means the model performs consistently.

Confusion matrix example across K-folds

Imagine a binary classification with 3 folds. Each fold has its own confusion matrix:

Fold 1:
TP=40, FP=10
FN=5,  TN=45

Fold 2:
TP=38, FP=12
FN=7,  TN=43

Fold 3:
TP=42, FP=8
FN=6,  TN=44

Total samples per fold = 100
    

We calculate precision and recall for each fold, then average them:

• Precision fold 1 = 40 / (40 + 10) = 0.80
• Recall fold 1 = 40 / (40 + 5) = 0.89
• Precision fold 2 = 38 / (38 + 12) = 0.76
• Recall fold 2 = 38 / (38 + 7) = 0.84
• Precision fold 3 = 42 / (42 + 8) = 0.84
• Recall fold 3 = 42 / (42 + 6) = 0.88

Average precision = (0.80 + 0.76 + 0.84) / 3 = 0.80

Average recall = (0.89 + 0.84 + 0.88) / 3 = 0.87

Precision vs Recall tradeoff in Cross-validation

Cross-validation helps us see if a model keeps a good balance between precision and recall on different data parts.

For example, if one fold has very high precision but low recall, and another fold the opposite, the average will show this imbalance.

This helps us choose a model that works well overall, not just on one slice of data.

Real-life example: For a spam filter, we want high precision (few good emails marked spam). Cross-validation shows if this holds across all data parts.

What "good" vs "bad" metric values look like for Cross-validation

Good:

• Average accuracy or F1 score is high (e.g., above 85%)
• Low variation between folds (e.g., standard deviation below 5%)
• Consistent precision and recall across folds

Bad:

• High average accuracy but very different scores in folds (one fold 95%, another 60%)
• High variation means model is unstable
• Precision very high but recall very low or vice versa in some folds

Common pitfalls with Cross-validation metrics

• Data leakage: If data from the test fold leaks into training folds, metrics look too good.
• Overfitting: High training scores but low cross-validation scores mean model memorizes training data.
• Ignoring variation: Only looking at average metric hides instability.
• Wrong metric choice: Using accuracy on imbalanced data can be misleading; prefer precision, recall, or F1.

Self-check question

Your model has 98% accuracy but 12% recall on fraud cases in cross-validation. Is it good for production? Why or why not?

Answer: No, it is not good. The very low recall means the model misses most fraud cases, which is dangerous. High accuracy can be misleading if fraud cases are rare. You need to improve recall to catch more fraud.

Key Result

Cross-validation metrics should show both high average performance and low variation across folds for a reliable model.

Practice

1. What is the main purpose of K-fold cross-validation in machine learning?

easy

A. To evaluate model performance on different subsets of data

B. To increase the size of the training dataset

C. To reduce the number of features in the dataset

D. To speed up the training process

Solution

1. Step 1: Understand K-fold cross-validation concept

   K-fold cross-validation splits the data into K parts and trains the model K times, each time using a different part as test data.

2. Step 2: Identify the main goal

   This process helps evaluate how well the model performs on unseen data by testing on different subsets.

3. Final Answer:

   To evaluate model performance on different subsets of data -> Option A

4. Quick Check:

   Cross-validation = model evaluation [OK]

Hint: K-fold tests model on all data parts for reliable results [OK]

Common Mistakes:

• Thinking it increases training data size
• Confusing it with feature selection
• Assuming it speeds up training

2. Which of the following is the correct way to import KFold from scikit-learn in Python?

easy

A. from sklearn.cross_validation import KFold

B. import KFold from sklearn.cross_validation

C. from sklearn.model_selection import KFold

D. import KFold from sklearn.model_selection

Solution

1. Step 1: Recall the current scikit-learn module for KFold

   KFold is part of the model_selection module in scikit-learn.

2. Step 2: Check correct import syntax

   The correct syntax is from sklearn.model_selection import KFold. Other options use outdated or incorrect modules or syntax.

3. Final Answer:

   from sklearn.model_selection import KFold -> Option C

4. Quick Check:

   Correct import = from sklearn.model_selection import KFold [OK]

Hint: Use model_selection module for KFold import [OK]

Common Mistakes:

• Using deprecated sklearn.cross_validation module
• Wrong import syntax like 'import KFold from ...'
• Confusing module names

3. What will be the output of the following Python code snippet?

from sklearn.model_selection import KFold
import numpy as np

X = np.array([10, 20, 30, 40, 50])
kf = KFold(n_splits=5)
for train_index, test_index in kf.split(X):
    print(test_index)

medium

A. [0 1 2 3]
[1 2 3 4]
[0 2 3 4]
[0 1 3 4]
[0 1 2 4]

B. [0 1]
[2 3]
[4]

C. [4]
[3]
[2]
[1]
[0]

D. [0]
[1]
[2]
[3]
[4]

Solution

1. Step 1: Understand KFold with n_splits=5 on 5 elements

   With 5 splits, each fold uses 4 elements for training and 1 element for testing.

2. Step 2: Check test indices printed in each fold

   Each fold's test index is a single element index from 0 to 4, printed one per line.

3. Final Answer:

   [0] [1] [2] [3] [4] -> Option D

4. Quick Check:

   Each fold tests one unique index [OK]

Hint: With 5 splits, test sets have 1 unique index each [OK]

Common Mistakes:

• Expecting multiple indices per test set
• Confusing train and test indices
• Misunderstanding KFold split behavior

4. Identify the error in this KFold usage code snippet:

from sklearn.model_selection import KFold
kf = KFold(n_splits=5)
X = [1, 2, 3, 4]
for train_idx, test_idx in kf.split(X):
    print(train_idx, test_idx)

medium

A. n_splits cannot be greater than the number of samples

B. X must be a numpy array, not a list

C. train_idx and test_idx should be swapped in the loop

D. Missing shuffle=True in KFold initialization

Solution

1. Step 1: Check n_splits vs data size

   n_splits=5 but X has only 4 samples (5 > 4).

2. Step 2: Understand KFold requirement

   KFold requires n_splits to be less than or equal to the number of samples, otherwise it raises a ValueError.

3. Final Answer:

   n_splits cannot be greater than the number of samples -> Option A

4. Quick Check:

   n_splits ≤ samples [OK]

Hint: n_splits must not exceed sample count [OK]

Common Mistakes:

• Using n_splits larger than dataset size
• Thinking list input causes error
• Swapping train and test indices

5. You want to use 10-fold cross-validation on a dataset with 100 samples. However, your model training is very slow. Which strategy can you use to still evaluate your model reliably but reduce training time?

hard

A. Use only one fold for testing and ignore others

B. Use 5-fold cross-validation instead of 10-fold

C. Use 10-fold cross-validation but shuffle=False

D. Use the entire dataset as training without validation

Solution

1. Step 1: Understand the trade-off between folds and training time

   More folds mean more training rounds, increasing time. Reducing folds reduces training rounds.

2. Step 2: Evaluate options for reducing training time

   Using 5-fold halves the number of training rounds compared to 10-fold, still providing reliable evaluation.

3. Step 3: Check other options

   Using entire data without validation loses evaluation, shuffle=False doesn't reduce folds, ignoring folds loses reliability.

4. Final Answer:

   Use 5-fold cross-validation instead of 10-fold -> Option B

5. Quick Check:

   Fewer folds = faster training [OK]

Hint: Reduce folds to speed up cross-validation [OK]

Common Mistakes:

• Skipping validation entirely
• Ignoring folds reduces reliability
• Thinking shuffle affects speed