A/B testing models in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - A/B testing models

Which metric matters for A/B testing models and WHY

A/B testing compares two models to see which one works better in real life. The key metrics depend on the goal. For example, if the goal is to increase clicks, use click-through rate (CTR). If it is to improve sales, use conversion rate. These metrics show the real impact of each model on user behavior.

Besides these, statistical significance (like p-value) is important to know if the difference is real or just by chance.

Confusion matrix or equivalent visualization

In A/B testing, we often track outcomes like success or failure for each model. Here is an example confusion matrix for Model A and Model B showing user actions:

      Model A Confusion Matrix:
      -------------------------
      |           | Clicked | Not Clicked |
      |-----------|---------|-------------|
      | Positive  |   80    |     20      |
      | Negative  |   30    |    870      |
      -------------------------

      Model B Confusion Matrix:
      -------------------------
      |           | Clicked | Not Clicked |
      |-----------|---------|-------------|
      | Positive  |   90    |     10      |
      | Negative  |   40    |    860      |
      -------------------------
    

These matrices help calculate metrics like CTR and conversion rate for each model.

Precision vs Recall tradeoff with concrete examples

A/B testing focuses less on precision and recall and more on overall impact metrics like conversion rate or revenue. However, if the models are classifiers, precision and recall matter.

For example, if Model A catches more true positives (high recall) but also more false positives (low precision), it might annoy users. Model B might be more precise but miss some opportunities. A/B testing helps find the balance that leads to better user experience and business results.

What "good" vs "bad" metric values look like for A/B testing

Good: Model B increases conversion rate from 5% to 7% with a p-value < 0.05, meaning the improvement is real and not by chance.

Bad: Model B shows a 1% increase but with a p-value > 0.1, so the change might be random. Or Model B increases clicks but decreases sales, which is bad for business.

Common pitfalls in A/B testing metrics

• Small sample size: Leads to unreliable results that can mislead decisions.
• Multiple testing: Running many tests increases chance of false positives.
• Ignoring user segments: Different groups may react differently, hiding true effects.
• Data leakage: Mixing data from before and after the test can bias results.
• Stopping test too early: Can cause wrong conclusions due to random fluctuations.

Self-check question

Your A/B test shows Model B has 98% accuracy but only 12% recall on fraud detection. Is it good for production? Why or why not?

Answer: No, it is not good. High accuracy can be misleading if fraud cases are rare. Low recall means the model misses most frauds, which is risky. You want high recall to catch as many frauds as possible.

Key Result

A/B testing metrics focus on real user impact like conversion rate and statistical significance to decide which model performs better.

Practice

1. What is the main purpose of A/B testing models in machine learning?

easy

A. To compare two models by testing them on separate groups

B. To train a single model on all data

C. To reduce the size of the dataset

D. To increase the number of features in the model

Solution

1. Step 1: Understand the goal of A/B testing

   A/B testing is used to compare two different models or versions by testing them separately.

2. Step 2: Identify the correct purpose

   The purpose is to find which model performs better by testing on separate groups, not to train on all data or change dataset size.

3. Final Answer:

   To compare two models by testing them on separate groups -> Option A

4. Quick Check:

   A/B testing = compare models [OK]

Hint: A/B testing means comparing two models on different groups [OK]

Common Mistakes:

• Thinking A/B testing trains one model on all data
• Confusing A/B testing with feature engineering
• Believing A/B testing reduces dataset size

2. Which of the following is the correct way to split data for A/B testing?

easy

A. Train both models on the entire dataset simultaneously

B. Use the same data for both models without splitting

C. Split data randomly into two groups and test each model separately

D. Split data by sorting values and assigning first half to model A

Solution

1. Step 1: Recall proper data splitting for A/B testing

   Data should be split randomly to avoid bias and ensure fair comparison.

2. Step 2: Evaluate options

   Only random splitting into two groups is correct; using same data or sorted split causes bias.

3. Final Answer:

   Split data randomly into two groups and test each model separately -> Option C

4. Quick Check:

   Random split = fair test [OK]

Hint: Always split data randomly for fair A/B testing [OK]

Common Mistakes:

• Using same data for both models
• Splitting data by sorting values
• Training both models on full data

3. Consider this Python code for A/B testing two models on separate data splits:

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_A, X_test_A, y_train_A, y_test_A = train_test_split(X[:3], y[:3], test_size=1, random_state=42)
X_train_B, X_test_B, y_train_B, y_test_B = train_test_split(X[3:], y[3:], test_size=1, random_state=42)

model_A = LogisticRegression().fit(X_train_A, y_train_A)
model_B = LogisticRegression().fit(X_train_B, y_train_B)

pred_A = model_A.predict(X_test_A)
pred_B = model_B.predict(X_test_B)

acc_A = accuracy_score(y_test_A, pred_A)
acc_B = accuracy_score(y_test_B, pred_B)

print(acc_A, acc_B)

What will be the printed output?

medium

A. 0.0 0.0

B. 1.0 1.0

C. 1.0 0.0

D. 0.0 1.0

Solution

1. Step 1: Analyze data splits and training

   Each model trains on 2 samples and tests on 1 sample. The test samples are [3] for A and [6] for B with labels 0 and 1 respectively.

2. Step 2: Predict and calculate accuracy

   LogisticRegression will perfectly separate these simple points, so predictions match true labels, giving accuracy 1.0 for both.

3. Final Answer:

   1.0 1.0 -> Option B

4. Quick Check:

   Perfect split and simple data = accuracy 1.0 [OK]

Hint: Small simple splits often give perfect accuracy [OK]

Common Mistakes:

• Assuming accuracy is zero due to small data
• Confusing train and test splits
• Ignoring random_state effect

4. You run an A/B test comparing two models but notice both groups have very different user ages. What is the likely problem?

medium

A. Data was not split randomly, causing bias

B. Models are overfitting the training data

C. Test size was too large

D. Models have too many features

Solution

1. Step 1: Identify issue from uneven user ages

   Unequal age distribution means groups are not similar, indicating non-random splitting.

2. Step 2: Understand impact on A/B testing

   Non-random splits cause bias, making test results unreliable.

3. Final Answer:

   Data was not split randomly, causing bias -> Option A

4. Quick Check:

   Unequal groups = non-random split [OK]

Hint: Unequal groups mean split was not random [OK]

Common Mistakes:

• Blaming overfitting for group differences
• Thinking test size causes group bias
• Confusing feature count with data split

5. You want to run an A/B test comparing two recommendation models on a website. Which approach best ensures a fair test?

hard

A. Assign users based on their signup date to different models

B. Show both models to all users and average their recommendations

C. Train both models on the same user data and pick the one with higher training accuracy

D. Randomly assign users to Model A or Model B groups and collect click data separately

Solution

1. Step 1: Understand fair A/B testing for users

   Random assignment ensures groups are similar and results are unbiased.

2. Step 2: Evaluate other options

   Showing both models to all users or assigning by signup date causes bias; training accuracy alone is not enough.

3. Final Answer:

   Randomly assign users to Model A or Model B groups and collect click data separately -> Option D

4. Quick Check:

   Random user assignment = fair A/B test [OK]

Hint: Random user assignment ensures unbiased A/B test [OK]

Common Mistakes:

• Showing both models to same users
• Using signup date for group assignment
• Choosing model by training accuracy only