One-vs-rest and one-vs-one strategies in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - One-vs-rest and one-vs-one strategies

Which metric matters for this concept and WHY

When using one-vs-rest (OvR) or one-vs-one (OvO) strategies for multi-class classification, metrics like accuracy, precision, recall, and F1-score matter. This is because these strategies break a multi-class problem into multiple binary problems. We need to measure how well each binary classifier performs and then combine results. Macro-averaged precision, recall, and F1-score help us understand performance equally across all classes, especially if classes are imbalanced.

Confusion matrix or equivalent visualization (ASCII)

For OvR, each class has its own binary confusion matrix. For example, with 3 classes (A, B, C), the OvR confusion matrix for class A looks like:

      Predicted A | Not A
    -----------------------
    A     |  TP  |  FN
    Not A |  FP  |  TN
    

For OvO, each pair of classes has a binary confusion matrix. For classes A and B:

      Predicted A | Predicted B
    ---------------------------
    A     |  TP  |  FN
    B     |  FP  |  TN
    

All these binary results combine to decide the final multi-class prediction.

Precision vs Recall tradeoff with concrete examples

In OvR or OvO, each binary classifier faces a tradeoff between precision and recall:

• Precision: How many predicted positives are actually correct? Important if false alarms are costly.
• Recall: How many actual positives are found? Important if missing a class is costly.

Example: For a disease detection with multiple diseases (classes), using OvR:

• If you want to avoid wrongly labeling healthy people as sick (false positives), focus on high precision.
• If you want to catch all sick people (true positives), focus on high recall.

Choosing OvR or OvO affects how these tradeoffs appear because OvO compares pairs directly, often improving precision but increasing complexity.

What "good" vs "bad" metric values look like for this use case

Good metrics for OvR/OvO multi-class classification:

• Accuracy: High (close to 1.0) means most samples are correctly classified.
• Macro F1-score: High (above 0.8) means balanced performance across all classes.
• Precision and Recall: Both should be reasonably high (above 0.7) for each class to avoid bias.

Bad metrics:

• High accuracy but low recall on some classes means the model misses many samples of those classes.
• High precision but low recall means the model is too strict and misses positives.
• Very low F1-score (below 0.5) indicates poor balance and unreliable classification.

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced. For example, if one class dominates, predicting it always yields high accuracy but poor real performance.
• Data leakage: If test data leaks into training, metrics look unrealistically good.
• Overfitting: Very high training metrics but low test metrics show the model memorizes training data but fails to generalize.
• Ignoring class imbalance: Not using macro-averaged metrics can hide poor performance on minority classes.

Self-check: Your model has 98% accuracy but 12% recall on fraud. Is it good?

No, this model is not good for fraud detection. Although 98% accuracy sounds high, the recall of 12% means it only finds 12% of actual fraud cases. This is bad because missing fraud is costly. The model likely predicts most samples as non-fraud, inflating accuracy but failing its main goal. Improving recall is critical here.

Key Result

In one-vs-rest and one-vs-one strategies, balanced precision, recall, and F1-score across classes are key to reliable multi-class classification.

Practice

1. What is the main idea behind the one-vs-rest strategy in multi-class classification?

easy

A. Train one model per class to separate that class from all others combined.

B. Train one model for every pair of classes.

C. Train a single model to classify all classes at once.

D. Train models only for the most frequent classes.

Solution

1. Step 1: Understand one-vs-rest approach

   One-vs-rest means creating one model per class. Each model learns to separate its class from all other classes combined.

2. Step 2: Compare with other options

   One-vs-one trains models for every pair, not per class. Single model for all classes is not one-vs-rest. Training only on frequent classes is unrelated.

3. Final Answer:

   Train one model per class to separate that class from all others combined. -> Option A

4. Quick Check:

   One-vs-rest = One model per class [OK]

Hint: One-vs-rest means one model per class vs all others [OK]

Common Mistakes:

• Confusing one-vs-rest with one-vs-one
• Thinking one-vs-rest uses one model for all classes
• Assuming one-vs-rest trains only on frequent classes

2. Which of the following correctly describes the number of models trained in the one-vs-one strategy for a problem with 4 classes?

easy

A. 4 models

B. 6 models

C. 1 model

D. 8 models

Solution

1. Step 1: Calculate number of pairs for 4 classes

   One-vs-one trains a model for every pair of classes. Number of pairs = 4 choose 2 = 4*3/2 = 6.

2. Step 2: Verify other options

   4 models is one per class (one-vs-rest). 1 model is single multi-class. 8 models is incorrect count.

3. Final Answer:

   6 models -> Option B

4. Quick Check:

   Pairs for 4 classes = 6 [OK]

Hint: Number of one-vs-one models = n*(n-1)/2 [OK]

Common Mistakes:

• Using number of classes instead of pairs
• Confusing one-vs-one with one-vs-rest counts
• Calculating pairs incorrectly

3. Consider a dataset with 3 classes: A, B, and C. Using one-vs-rest, how many models are trained and what does each model learn?

medium

A. 6 models; each separates pairs of classes.

B. 3 models; each separates one class from one other class only.

C. 1 model; separates all three classes at once.

D. 3 models; each separates one class from the other two combined.

Solution

1. Step 1: Count models in one-vs-rest for 3 classes

   One-vs-rest trains one model per class, so 3 models total.

2. Step 2: Understand model learning in one-vs-rest

   Each model learns to separate its class from all other classes combined (not just one other class).

3. Final Answer:

   3 models; each separates one class from the other two combined. -> Option D

4. Quick Check:

   One-vs-rest with 3 classes = 3 models [OK]

Hint: One-vs-rest trains one model per class vs all others [OK]

Common Mistakes:

• Thinking one-vs-rest trains models per pair
• Assuming only one model is trained
• Confusing one-vs-rest with one-vs-one

4. You implemented one-vs-one for a 5-class problem but only trained 4 models. What is the likely mistake?

medium

A. You trained models only for the most frequent classes.

B. You trained one model per class instead of pairs.

C. You forgot to train models for all pairs; should be 10 models.

D. You trained a single multi-class model.

Solution

1. Step 1: Calculate expected number of one-vs-one models for 5 classes

   Number of pairs = 5 choose 2 = 5*4/2 = 10 models expected.

2. Step 2: Identify mistake from training only 4 models

   Training only 4 models means some pairs were missed. Possibly forgot to train all pairs.

3. Final Answer:

   You forgot to train models for all pairs; should be 10 models. -> Option C

4. Quick Check:

   One-vs-one for 5 classes = 10 models [OK]

Hint: One-vs-one needs n*(n-1)/2 models; check count [OK]

Common Mistakes:

• Counting models as number of classes
• Confusing one-vs-one with one-vs-rest
• Training incomplete pairs

5. You have a 4-class problem with unbalanced data. Which strategy is better to handle this imbalance and why?

hard

A. One-vs-one, because training on pairs reduces imbalance impact between classes.

B. Neither, use a single multi-class model only.

C. One-vs-rest, because each model focuses on separating one class from all others, helping with imbalance.

D. Train only on the largest class to simplify the problem.

Solution

1. Step 1: Understand imbalance effect on one-vs-rest

   One-vs-rest models separate one class vs all others combined, which can cause imbalance if one class is small and others are large.

2. Step 2: Understand one-vs-one advantage

   One-vs-one trains models on pairs of classes, so imbalance is less severe per model, improving learning on minority classes.

3. Step 3: Evaluate other options

   Single multi-class model may struggle with imbalance. Training only on largest class ignores others.

4. Final Answer:

   One-vs-one, because training on pairs reduces imbalance impact between classes. -> Option A

5. Quick Check:

   One-vs-one handles imbalance better [OK]

Hint: One-vs-one handles imbalance better by focusing on pairs [OK]

Common Mistakes:

• Assuming one-vs-rest always better for imbalance
• Ignoring imbalance effects on combined classes
• Choosing single model ignoring class distribution