What's changed: Deepened AI-900 Chapter 2 to the AZ-900 baseline (features/labels/generalization/overfitting, deep learning = multi-layer NN, metrics accuracy/precision/recall/confusion matrix, workspace, FAQ)
2.2Azure Machine Learning Services and Workflow
Understand the basic ML workflow (prepare data → train → evaluate → deploy → predict) and Azure Machine Learning’s no-code features (Automated ML and Designer).
A model is not "build once and done"—it grows through prepare → train → evaluate → deploy → predict. On Azure, Azure Machine Learning supports this end to end, from no-code options (Automated ML, Designer) to full Python development.
2.2.1The machine learning workflow
- Prepare data: collect, clean missing/outlier values, label if needed, and split into training and evaluation sets.
- Train: train the model on the training dataset.
- Evaluate: measure accuracy on data not used for training to judge quality.
- Deploy: publish the model as a service (endpoint), for real-time or batch.
- Predict (inference): return predictions for new data; retrain as needed.
2.2.2Features of Azure Machine Learning
- Automated ML: automatically tries many algorithms and preprocessing and picks the best model (no/low-code).
- Designer: a drag-and-drop, no-code way to build pipelines.
- Notebooks / Python SDK: full code-based development.
- Manage compute, datasets, registered models, and endpoints in one place (a workspace).
| Feature | Trait | For whom |
|---|---|---|
| Automated ML | Auto-selects the best model | ML beginners, quick trials |
| Designer | Drag-and-drop (no-code) | Visual pipeline building |
| Notebooks/SDK | Full control in Python | Data scientists |
2.2.3Evaluation metrics in a bit more detail
Metrics depend on the task type. For classification, accuracy (overall correct rate) is the basic one, but where missing positives is costly (e.g., disease detection) you weigh recall (of true positives, how many you caught), and where false alarms are costly you weigh precision (of predicted positives, how many were truly positive). These come from the confusion matrix and reveal weaknesses accuracy alone hides. For regression, you measure the size of the error between predicted and actual values, using MAE or RMSE. In all cases, always measure on data not used for training.
Metrics depend on the task: classification uses accuracy, precision/recall, etc.; regression uses error measures (MAE, RMSE). Always evaluate on data not used for training to catch overfitting (fitting the training data too closely).
Q. What is an Azure ML "workspace"? A container that centralizes compute, datasets, registered models, and endpoints, so experiments and artifacts live in one place for reproducibility and team sharing. Q. Real-time vs batch deployment? Real-time returns predictions one request at a time (e.g., online decisions); batch processes large datasets together (e.g., nightly bulk scoring).
Scenario: quickly try churn prediction. A team without data scientists feeds customer data to Automated ML to auto-find the best classifier → uses Designer to visualize/adjust preprocessing → deploys to an endpoint for existing systems to call. Evaluate on data not used for training.
Watch the mix-ups: (1) Automated ML (auto-select best model) vs Designer (no-code pipeline building)—different purposes. (2) Evaluate on data not used for training (same-data evaluation overstates and hides overfitting). (3) Even "managed," data quality and bias are the customer’s responsibility (Responsible AI, previous chapter).
Common mappings: Automated ML = auto-select the best model, Designer = drag-and-drop no-code, evaluate on held-out data, classification = accuracy / regression = error.
2.2.4Section summary
- Workflow: prepare (split) → train → evaluate → deploy → predict
- Azure ML: Automated ML (auto-select) / Designer (no-code) / notebooks & SDK
- Evaluate on data not used for training; metrics: classification = accuracy, regression = error
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which Azure Machine Learning feature automatically tries algorithms to find the best model?
Q2. Which feature builds ML pipelines via drag-and-drop without writing code?
Q3. To judge a model fairly, which data should be used for evaluation?
Q4. Which is a common evaluation metric for a classification model?
Q5. Which step publishes a trained model so it can be called to return predictions?