Troubleshooting & Problem Resolution Flashcards

A classification model scores 99% accuracy on training data but 68% on the test set. What is the most likely problem?

Answer: Overfitting

A large gap between training and test performance is the classic sign of overfitting.

Which action is MOST appropriate to reduce overfitting in a decision tree?

Answer: Prune the tree or limit its depth

Pruning or limiting depth reduces model complexity and improves generalization.

A k-means run gives different clusters each time it is executed on the same data. What is the most likely cause?

Answer: Random centroid initialization

K-means starts from random centroids and can converge to different local optima, so fixing the seed or using multiple restarts helps.

A fraud detection model with only 1% positive cases predicts 'not fraud' for everything and reports 99% accuracy. What is the best fix to investigate first?

Answer: Address class imbalance and evaluate with precision/recall

With imbalanced classes accuracy is misleading, so resampling or class weights plus precision/recall are needed.

Model performance drops sharply a few months after deployment even though the code has not changed. What is the most probable cause?

Answer: Concept or data drift

Changes in the underlying data distribution or relationships degrade a deployed model over time.

Which symptom most strongly suggests target leakage in a model?

Answer: Suspiciously near-perfect performance that collapses in production

Leaked features carry information unavailable at prediction time, giving unrealistically high validation scores.

A linear regression shows a very high R² but coefficient signs that contradict domain knowledge. Which issue should be checked first?

Answer: Multicollinearity among predictors

Highly correlated predictors make coefficients unstable and can flip their signs.