Bias detection and mitigation in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Bias detection and mitigation

Which metric matters for Bias Detection and Mitigation and WHY

When checking for bias in a model, we look beyond overall accuracy. We want to see if the model treats all groups fairly. Key metrics include Demographic Parity (does each group get positive outcomes equally?), Equal Opportunity (does each group have equal true positive rates?), and Disparate Impact (ratio of positive outcomes between groups). These metrics help us find if the model favors or harms certain groups unfairly.

Confusion Matrix Example for Two Groups

Imagine a model predicting loan approvals for two groups: Group A and Group B.

    Group A Confusion Matrix:
      TP=80  FP=20
      FN=10  TN=90

    Group B Confusion Matrix:
      TP=40  FP=10
      FN=30  TN=120

    Total samples Group A = 80+20+10+90 = 200
    Total samples Group B = 40+10+30+120 = 200
    

From these, we calculate metrics like True Positive Rate (Recall) for each group:

• Group A Recall = 80 / (80 + 10) = 0.89
• Group B Recall = 40 / (40 + 30) = 0.57

This shows Group B has a lower chance of correctly getting a positive outcome, indicating bias.

Precision vs Recall Tradeoff in Bias Context

Suppose a hiring model favors Group A over Group B. If we only optimize for overall precision, the model might reject many qualified candidates from Group B (low recall for Group B). To fix this, we might accept a slight drop in precision to improve recall for Group B, making the model fairer.

Example:

• High precision but low recall for Group B means many qualified candidates are missed.
• Improving recall for Group B may lower precision but reduces unfair rejection.

What "Good" vs "Bad" Metric Values Look Like

Good bias metrics mean similar performance across groups:

• Recall difference between groups < 0.05 (5%)
• Disparate Impact ratio close to 1 (between 0.8 and 1.25 is often acceptable)
• Equal Opportunity difference small

Bad values show large gaps, like:

• Recall difference > 0.2 (20%)
• Disparate Impact < 0.8 or > 1.25
• One group has very low true positive rate compared to others

Common Pitfalls in Bias Metrics

• Ignoring base rates: Different groups may have different real-world rates, so metrics must consider context.
• Accuracy paradox: High overall accuracy can hide poor performance on minority groups.
• Data leakage: Sensitive attributes leaking into features can cause hidden bias.
• Overfitting mitigation: Over-correcting bias can reduce overall model usefulness.
• Single metric focus: Using only one fairness metric can miss other bias types.

Self-Check Question

Your model has 98% accuracy overall but only 12% recall on fraud cases in a minority group. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most fraud cases in that group, which is harmful. High overall accuracy hides this problem because fraud is rare. You need to improve recall for that group to reduce bias and protect users.

Key Result

Fairness metrics like recall differences and disparate impact reveal bias beyond overall accuracy.

Practice

1. What is the main goal of bias detection in machine learning models?

easy

A. To find unfair treatment of certain groups by the model

B. To increase the speed of model training

C. To reduce the size of the dataset

D. To improve the model's accuracy on all data

Solution

1. Step 1: Understand bias detection purpose

   Bias detection aims to identify if a model treats some groups unfairly compared to others.

2. Step 2: Compare options to goal

   Only To find unfair treatment of certain groups by the model matches this goal; others focus on speed, size, or accuracy, which are unrelated.

3. Final Answer:

   To find unfair treatment of certain groups by the model -> Option A

4. Quick Check:

   Bias detection = find unfair treatment [OK]

Hint: Bias detection finds unfairness in model decisions [OK]

Common Mistakes:

• Confusing bias detection with improving accuracy
• Thinking bias detection speeds up training
• Mixing bias detection with data size reduction

2. Which of the following is a common technique to mitigate bias in training data?

easy

A. Oversampling underrepresented groups

B. Reducing the number of features

C. Increasing the learning rate

D. Using dropout layers in neural networks

Solution

1. Step 1: Identify bias mitigation methods

   Oversampling underrepresented groups balances data, reducing bias.

2. Step 2: Evaluate other options

   Reducing features, increasing learning rate, and dropout affect model complexity or training, not bias directly.

3. Final Answer:

   Oversampling underrepresented groups -> Option A

4. Quick Check:

   Bias mitigation = oversampling minority data [OK]

Hint: Balance data by oversampling small groups [OK]

Common Mistakes:

• Confusing feature reduction with bias mitigation
• Thinking learning rate affects bias
• Mixing dropout with bias correction

3. Consider this Python code snippet for bias mitigation:

from sklearn.utils import resample

# Original data
X = [[1], [2], [3], [4]]
y = [0, 0, 1, 1]

# Minority class samples
X_minority = [x for x, label in zip(X, y) if label == 1]
y_minority = [1, 1]

# Oversample minority class
X_upsampled, y_upsampled = resample(X_minority, y_minority, replace=True, n_samples=4, random_state=42)

print(len(X_upsampled), len(y_upsampled))

What will be the output of the print statement?

medium

A. 2 2

B. 4 4

C. 8 8

D. Error due to wrong resample usage

Solution

1. Step 1: Understand resample parameters

   resample duplicates minority samples to reach n_samples=4, so output lists have length 4.

2. Step 2: Check print output

   Both X_upsampled and y_upsampled have length 4, so print outputs '4 4'.

3. Final Answer:

   4 4 -> Option B

4. Quick Check:

   Oversampling size = 4 [OK]

Hint: resample with n_samples sets output length [OK]

Common Mistakes:

• Assuming output length stays original size
• Thinking resample doubles data always
• Expecting error from resample usage

4. You have a model that shows bias against a group. The code below tries to fix it but causes an error:

def mitigate_bias(data, labels):
    # Remove samples from majority group
    filtered_data = [d for d, l in zip(data, labels) if l != 'majority']
    filtered_labels = [l for l in labels if l != 'majority']
    return filtered_data, filtered_labels

X = [1, 2, 3, 4]
y = ['majority', 'minority', 'majority', 'minority']

X_new, y_new = mitigate_bias(X, y)
print(len(X_new), len(y_new))

What is the error in this code?

medium

A. Function returns wrong variable names

B. Syntax error in list comprehension

C. No error, code runs correctly

D. filtered_labels length does not match filtered_data length

Solution

1. Step 1: Analyze the filtering logic

   filtered_data is filtered using zip(data, labels) with condition l != 'majority', selecting elements at indices 1 and 3. filtered_labels is filtered directly from labels with the same condition, resulting in elements at indices 1 and 3. However, filtered_data and filtered_labels are filtered differently: filtered_data uses zip to filter data and labels together, but filtered_labels filters labels alone, which can cause length mismatch if data and labels are not aligned.

2. Step 2: Check lengths

   In this specific example, both filtered_data and filtered_labels have length 2, but this is coincidental. The approach is error-prone and can cause mismatches in other cases.

3. Step 3: Correct approach

   Both filtered_data and filtered_labels should be filtered using zip together to ensure matching lengths.

4. Final Answer:

   filtered_labels length does not match filtered_data length -> Option D

5. Quick Check:

   Separate filtering risks length mismatch [OK]

Hint: Filter data and labels together using zip to keep lengths equal [OK]

Common Mistakes:

• Mistakenly thinking separate comprehensions cause length mismatch
• Believing zip and direct label filter select differently
• Assuming the code has a bug without tracing execution

5. You want to reduce bias in a model predicting loan approvals. You have data with a sensitive attribute 'gender'. Which combined approach best mitigates bias while keeping model performance?

hard

A. Increase model complexity to capture more patterns

B. Remove the 'gender' attribute and train the model normally

C. Use reweighting to balance training samples and add fairness constraints during model training

D. Only oversample the majority group to increase data size

Solution

1. Step 1: Understand bias mitigation strategies

   Reweighting balances data impact; fairness constraints guide model to treat groups fairly.

2. Step 2: Evaluate other options

   Removing 'gender' may hide bias but not fix it; increasing complexity can worsen bias; oversampling majority increases imbalance.

3. Final Answer:

   Use reweighting to balance training samples and add fairness constraints during model training -> Option C

4. Quick Check:

   Combine reweighting + fairness constraints = best bias mitigation [OK]

Hint: Combine data balancing and fairness rules for best results [OK]

Common Mistakes:

• Removing sensitive attribute hides but does not fix bias
• Increasing complexity ignores bias issues
• Oversampling majority worsens bias