Domain 2 of 5 · Chapter 2 of 3

ML Core Concepts

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing Microsoft Azure AI Fundamentals premium course — no separate purchase.

Included in this chapter:

  • Features (x), labels (y), and why you split the data before you train
  • Train → evaluate → deploy → inference, and how the model is actually scored
  • Exam-pattern recognition: feature/label stems and the train-vs-validation traps

Training dataset vs validation dataset

AspectTraining subsetValidation (held-back) subset
PurposeFit the function f to the dataTest how well the fitted model predicts
Seen during trainingYes, the algorithm learns from itNo, held back so it is unseen
Labels known?Yes, used to derive fYes, but used only to compare against predictions
ProducesThe trained model (function f)The evaluation metric (e.g. RMSE, accuracy)
Risk if used aloneLooks accurate by memorizing examplesReveals whether the model generalizes
Cheat sheet

Cheat sheet

  • Features (x) are the inputs; the label (y) is what you predict
  • A label column exists only in labelled (supervised) data
  • Split labelled data into a training subset and a held-back validation subset
  • Evaluate on the held-back data, never on the training data
  • Overfitting = great on training data, poor on new data
  • Training fits the function y = f(x) from labelled data
  • Inferencing applies the finished model to new inputs
  • Predicted label is ŷ (y-hat); true label is y
  • Evaluation aggregates the ŷ-vs-y differences into one metric
  • Model development is iterative: train, evaluate, refine, repeat
  • x and y are the parameters; hyperparameters are tunable algorithm settings
  • Numeric label → regression metrics; categorical label → classification metrics
  • Regression: lower MAE/MSE/RMSE is better; higher R² is better
  • Accuracy misleads on imbalanced data; precision and recall expose it
  • Confusion matrix: TP, FP, FN, TN defined by predicted vs actual
  • The confusion-matrix diagonal holds the correct predictions per class
  • Precision is of-predicted-positive; recall is of-actual-positive
  • Feature engineering, normalization, and one-hot encoding prep data for training
  • K-Means: you set k, the number of centroids representing the clusters

Unlock with Premium — includes all practice exams and the complete study guide.

Also tested in

References

  1. Identify features and labels in a dataset for machine learning (AI-900 study guide)
  2. Describe how training and validation datasets are used in machine learning