Bias detection and mitigation in ML Python - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding Bias Types in Machine Learning

Which of the following best describes sampling bias in a dataset?

A The model predictions are skewed because of incorrect feature scaling.

B The model favors one class over another because of imbalanced class labels during training.

C The dataset contains data points that are not representative of the real population due to how samples were collected.

D The dataset has missing values that cause errors during training.

💡 Hint

Think about how the data was gathered and if it truly reflects the whole group you want to study.

Sampling bias happens when the data collected does not fairly represent the entire population, leading to unfair or inaccurate model results.

❓ Predict Output

intermediate

Detecting Bias with Statistical Parity Difference

Given the following Python code calculating statistical parity difference, what is the printed output?

ML Python

group_0_positive_rate = 0.7
group_1_positive_rate = 0.5
statistical_parity_difference = group_0_positive_rate - group_1_positive_rate
print(f"Statistical Parity Difference: {statistical_parity_difference}")

A Statistical Parity Difference: 0.2

B Statistical Parity Difference: -0.2

C Statistical Parity Difference: 1.2

D Statistical Parity Difference: 0.0

💡 Hint

Subtract the smaller positive rate from the larger one.

Statistical parity difference is the difference between positive outcome rates of two groups. Here, 0.7 - 0.5 = 0.2.

❓ Model Choice

advanced

Choosing a Model to Mitigate Bias

You want to reduce bias in a classification task where sensitive attributes affect predictions. Which model approach is best to mitigate bias during training?

A Train a standard logistic regression without any fairness constraints.

B Use a decision tree with maximum depth to fit training data perfectly.

C Train a model only on the sensitive attribute to predict the target.

D Use an adversarial debiasing model that learns to predict while minimizing sensitive attribute information.

💡 Hint

Think about a model that tries to hide sensitive information while learning.

Adversarial debiasing trains a model to predict the target while an adversary tries to predict sensitive attributes. This reduces bias by removing sensitive info from predictions.

❓ Hyperparameter

advanced

Hyperparameter Impact on Bias Mitigation

In a fairness-aware model using a regularization term to penalize bias, which hyperparameter adjustment will most likely reduce bias?

A Increase the regularization strength to penalize bias more heavily.

B Decrease the regularization strength to allow more model flexibility.

C Increase the learning rate to speed up training.

D Decrease the batch size to improve gradient estimates.

💡 Hint

Think about how stronger penalties affect bias in the model.

Increasing regularization strength on bias penalty forces the model to reduce bias more, improving fairness.

❓ Metrics

expert

Interpreting Equalized Odds Metric

You have two groups in your dataset. The true positive rates (TPR) and false positive rates (FPR) for each group are:

• Group A: TPR=0.8, FPR=0.1
• Group B: TPR=0.6, FPR=0.1

Does this model satisfy equalized odds fairness?

A Yes, because both groups have the same false positive rate.

B No, because the true positive rates differ between groups.

C Yes, because the average of TPR and FPR is equal.

D No, because the false positive rates differ between groups.

💡 Hint

Equalized odds requires both TPR and FPR to be equal across groups.

Equalized odds means both true positive rates and false positive rates must be equal for all groups. Here, TPR differs, so fairness is not satisfied.