LEARN. PRACTICE. GROW.Concept and mechanism
The task depends on the target. Regression predicts a numeric value, such as batch duration. Classification assigns a category learned from labeled examples, such as network or application. Clustering explores similarities without those prior classes. Features are inputs; the label is the known target in supervised examples. Deep learning uses multilayer neural networks to learn representations. Transformers use attention to relate context elements; they are not a truth guarantee. Azure Machine Learning supplies data, compute, experiments, model management, and deployment capabilities. Automated ML can compare configurations, but the team still defines the problem and acceptance.
Guided application
In a fictional exercise, predict closing-delay risk at 09:00. Use only inputs available at that time. A field calculated after closing introduces target leakage even if testing looks excellent. Separate learning and evaluation to measure behavior on data not used for fitting. A large gap between training and validation results warrants investigation of overfitting and representativeness. With rare incidents, overall accuracy can reward a model that never identifies them. Inspect the confusion matrix, critical-class recall, and false-positive cost. Retain data and model versions to compare releases and reproduce the acceptance decision.
980 normal and 20 critical tickets: predicting normal for all gives 98% accuracy and 0% critical recall.
Common pitfalls
Future information as a feature; training as testing; accuracy as sufficient; AutoML as automatic approval.
Related topics: Workloads and operational responsibility · Vision and documents with validation · Language, speech, and operational meaning
Evaluate the target and errors that matter at the actual decision time.
Reference: Automated machine learning · AI-900 historical skills measured 2025-05-02; exam retired 2026-06-30