ROC curve and AUC in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - ROC curve and AUC

Which metric matters for ROC curve and AUC and WHY

The ROC curve shows how well a model can separate positive and negative cases at different thresholds. The AUC (Area Under the Curve) summarizes this ability into one number between 0 and 1. A higher AUC means the model is better at ranking positive cases higher than negative ones. This is important when you want to understand the model's overall discrimination power, regardless of a specific cutoff.

Confusion matrix or equivalent visualization

The ROC curve is built from many confusion matrices at different thresholds. Here is an example confusion matrix at one threshold:

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP)  | False Positive (FP) |
      | False Negative (FN) | True Negative (TN)  |
    

By changing the threshold, TP, FP, TN, and FN change, creating different points on the ROC curve.

Precision vs Recall tradeoff with concrete examples

While ROC and AUC focus on true positive rate (Recall) and false positive rate, precision is not directly shown. For example:

• In a spam filter, you want high precision to avoid marking good emails as spam.
• In a cancer detector, you want high recall to catch as many cancer cases as possible.

ROC and AUC help you see how well the model balances catching positives (Recall) and avoiding false alarms (False Positive Rate) across all thresholds.

What "good" vs "bad" metric values look like for ROC and AUC

AUC values range from 0 to 1:

• Good: AUC close to 1 means the model perfectly ranks positives above negatives.
• Average: AUC around 0.7 to 0.8 means the model has decent discrimination.
• Bad: AUC near 0.5 means the model is no better than random guessing.
• Very bad: AUC below 0.5 means the model is worse than random (it ranks negatives higher than positives).

Metrics pitfalls

• Accuracy paradox: High accuracy can happen if data is imbalanced, but AUC shows poor ranking ability.
• Data leakage: If test data leaks into training, AUC will be unrealistically high.
• Overfitting: AUC on training data may be high but low on new data, showing poor generalization.
• Ignoring class imbalance: AUC is robust but can still be misleading if positive cases are very rare.

Self-check question

Your model has 98% accuracy but an AUC of 0.55. Is it good for production? Why or why not?

Answer: No, it is not good. The high accuracy likely comes from many negative cases, but the AUC near 0.5 means the model cannot distinguish positives from negatives well. It will fail to rank positive cases correctly, making it unreliable.

Key Result

AUC summarizes model's ability to rank positives above negatives; closer to 1 is better, 0.5 means random guessing.

Practice

1. What does the ROC curve primarily show in a classification model?

easy

A. The trade-off between true positive rate and false positive rate at different thresholds

B. The accuracy of the model on the training data

C. The number of correct predictions made by the model

D. The loss value during model training

Solution

1. Step 1: Understand ROC curve components

   The ROC curve plots the true positive rate (sensitivity) against the false positive rate (1 - specificity) at various threshold settings.

2. Step 2: Identify what ROC shows

   This curve shows how well the model can distinguish between classes by varying the decision threshold.

3. Final Answer:

   The trade-off between true positive rate and false positive rate at different thresholds -> Option A

4. Quick Check:

   ROC curve = True Positive Rate vs False Positive Rate [OK]

Hint: ROC curve plots TPR vs FPR at thresholds [OK]

Common Mistakes:

• Confusing ROC with accuracy
• Thinking ROC shows loss values
• Assuming ROC shows only one threshold

2. Which of the following Python code snippets correctly imports the function to compute AUC from sklearn?

easy

A. import roc_auc_score from sklearn.metrics

B. from sklearn.metrics import auc_roc_score

C. from sklearn.metrics import roc_auc_score

D. import sklearn.metrics.roc_auc_score

Solution

1. Step 1: Recall correct import syntax

   In Python, functions are imported using 'from module import function' syntax.

2. Step 2: Identify correct function name

   The correct function name to compute AUC is 'roc_auc_score' from 'sklearn.metrics'.

3. Final Answer:

   from sklearn.metrics import roc_auc_score -> Option C

4. Quick Check:

   Correct import syntax = from sklearn.metrics import roc_auc_score [OK]

Hint: Use 'from sklearn.metrics import roc_auc_score' [OK]

Common Mistakes:

• Using wrong import syntax
• Misspelling function name
• Trying to import module as function

3. Given the following code, what is the printed AUC value?

from sklearn.metrics import roc_auc_score
true_labels = [0, 0, 1, 1]
predicted_scores = [0.1, 0.4, 0.35, 0.8]
auc = roc_auc_score(true_labels, predicted_scores)
print(round(auc, 2))

medium

A. 0.75

B. 0.85

C. 0.65

D. 0.95

Solution

1. Step 1: Understand roc_auc_score calculation

   The AUC measures how well the predicted scores rank positive samples higher than negative ones.

2. Step 2: Calculate AUC for given data

   Using the pairs, the AUC is 0.75 as the positive samples are ranked correctly most of the time but not perfectly.

3. Final Answer:

   0.75 -> Option A

4. Quick Check:

   AUC for given scores = 0.75 [OK]

Hint: Use roc_auc_score to get AUC directly [OK]

Common Mistakes:

• Confusing AUC with accuracy
• Rounding incorrectly
• Mixing up true labels and scores

4. Identify the error in the following code that tries to plot an ROC curve:

from sklearn.metrics import roc_curve
fpr, tpr, thresholds = roc_curve(y_true, y_pred)
plt.plot(fpr, tpr)
plt.show()

medium

A. roc_curve function is not imported correctly

B. y_pred should be predicted probabilities, not class labels

C. plt.plot should use thresholds instead of fpr

D. Missing import statement for numpy

Solution

1. Step 1: Check input requirements for roc_curve

   The roc_curve function requires true labels and predicted scores or probabilities, not class labels.

2. Step 2: Identify the error in inputs

   Using predicted class labels instead of probabilities causes incorrect ROC curve calculation.

3. Final Answer:

   y_pred should be predicted probabilities, not class labels -> Option B

4. Quick Check:

   Use probabilities for roc_curve inputs [OK]

Hint: Use predicted probabilities, not labels, for roc_curve [OK]

Common Mistakes:

• Passing class labels instead of probabilities
• Forgetting to import matplotlib.pyplot as plt
• Plotting thresholds instead of FPR and TPR

5. You have two models with these AUC scores: Model A = 0.82, Model B = 0.79. You want to improve Model B's performance. Which approach is best to increase the AUC score?

hard

A. Reduce the size of the training dataset to avoid overfitting

B. Tune the classification threshold to balance true positive and false positive rates

C. Use accuracy as the main metric instead of AUC

D. Increase the number of training epochs without changing data or model

Solution

1. Step 1: Understand what affects AUC

   AUC measures the model's ranking ability and is threshold-independent.

2. Step 2: Choose the best method to improve AUC

   Tuning the classification threshold does not improve AUC because AUC is threshold-independent; instead, improving the model's discriminative ability or training is needed.

3. Step 3: Evaluate options

   Increasing training epochs (Increase the number of training epochs without changing data or model) can help but may cause overfitting if not managed properly. Reducing dataset size (Reduce the size of the training dataset to avoid overfitting) usually harms performance. Using accuracy (Use accuracy as the main metric instead of AUC) does not improve AUC. Tuning threshold (Tune the classification threshold to balance true positive and false positive rates) affects operating point but not AUC.

4. Final Answer:

   Increase the number of training epochs without changing data or model -> Option D

5. Quick Check:

   More epochs can improve model performance and AUC [OK]

Hint: More training epochs improve AUC [OK]

Common Mistakes:

• Thinking threshold tuning affects AUC
• Reducing data size harms performance
• Switching to accuracy doesn't improve AUC