Machine Learning Specialization Practice Support

Tag every missed question by category: algorithm selection, cost or loss function, gradient descent, feature scaling, regularization, bias and variance, metric choice, threshold selection, data leakage, neural network structure, tree behavior, clustering, anomaly detection, recommender systems, or workflow error.

Before choosing an answer, identify the task. Regression predicts a continuous value; classification predicts a class; clustering groups unlabeled examples; anomaly detection identifies unusual examples; recommender systems estimate user-item preference.

Many wrong answers misuse data splits. Training data fits parameters, validation data guides model selection, and test data estimates final performance. If a solution tunes repeatedly on the test set, it is leaking evaluation information.

Low training error and high validation error points toward variance. High training and validation error points toward bias. Diverging cost may point to learning rate or scaling. Strong accuracy with poor minority-class detection points toward metric or threshold problems.

Distractors are usually attractive because they sound like a familiar technique. Regularization is not the same as feature scaling. Logistic regression is classification, not continuous regression. Clustering is not supervised classification. Precision and recall answer different business risks.

After each DotCreds question, write a one-sentence explanation: “The model is overfitting because training error is low and validation error is high, so regularization, simpler features, or more data may help.” Short explanations expose fuzzy understanding.

If a concept stays abstract, reproduce it in a notebook. Change one variable and observe the result. ML understanding gets stronger when you can connect the practice question to a visible training curve, metric change, or prediction error.