AI-901
Deep Dive
Azure ML offers two tools for users without deep ML experience: AutoML automates algorithm selection and Designer provides a visual drag-and-drop canvas. The AI-901 tests when and how to use each one.
Contents
AutoML (Automated Machine Learning) automates the most technical parts of the ML process: trying multiple algorithms, tuning hyperparameters, and selecting the best model — all without the user writing ML code.
What AutoML automates
What AutoML does NOT automate
Critical exam trap
AutoML does not replace data scientists. It democratizes ML for business users and speeds up experimentation, but professionals are still needed to define the problem, prepare quality data, interpret results, and oversee models in production.
Prepare the dataset
Upload data to Azure ML as a tabular or time-series dataset. AutoML requires structured data (tables).
Configure the experiment
Select: task type (classification/regression/forecasting), target column (label), optimization metric, maximum training time.
Select compute
Choose a Compute Cluster for parallel training. AutoML trains multiple models in parallel if the cluster has several nodes.
Automatic run
AutoML tries combinations of algorithms + hyperparameters. Each combination is a "child run" within the parent experiment.
Review results
The UI shows a ranking of models with metrics. You can explore explainability (feature importance) for each model.
Deploy the best model
One click to register the model in the Model Registry and deploy it as an online or batch inference endpoint.
Classification
Predicts which category an observation belongs to.
Examples
Metrics
Accuracy, AUC, F1, Precision, Recall
Regression
Predicts a continuous numeric value.
Examples
Metrics
RMSE, MAE, R², MAPE
Forecasting (time series)
Predicts future values based on historical patterns with a time index.
Examples
Metrics
MAPE, RMSE, MAE with a time horizon
Computer vision and NLP with AutoML
AutoML also supports image tasks (image classification, object detection) and text tasks (text classification, NER) by training deep learning models. For the AI-901, the emphasis is on classic tabular tasks.
AutoML ranks candidate models by the chosen optimization metric. You must pick the right metric for the problem.
| Scenario | Recommended metric | Why |
|---|---|---|
| Fraud detection (imbalanced classes) | AUC_weighted | Robust with unequal classes; accuracy would be misleading |
| Classification with balanced classes | Accuracy | The most intuitive metric when classes are equal |
| Medical diagnosis (minimize false negatives) | Recall / F1 | More important not to miss positives (sick patients) |
| Spam moderation (minimize false positives) | Precision | More important not to block legitimate emails |
| Price prediction | RMSE or R² | RMSE penalizes large errors; R² measures explained variance |
| Demand forecasting (relative errors) | MAPE | Percentage error independent of scale |
Designer is Azure ML's visual drag-and-drop canvas. It lets you build ML pipelines by connecting modules graphically, with no code. Ideal for people who understand the ML process but don't want to (or can't) program.
Understand the business but not Python/R.
Familiar with data but not with ML code.
Visualizing the pipeline helps them understand the process.
Testing pipeline ideas before coding.
Data input
Data transformation
ML algorithms
Training and evaluation
Deployment
| Dimension | AutoML | Designer | SDK / Notebooks |
|---|---|---|---|
| Code required | No (UI) or minimal | No | Yes (Python/R) |
| Algorithm control | Automatic | Visual selection | Full |
| Ideal profile | Analysts, rapid exploration | No-code, visualization | Data scientists |
| Flexibility | Medium | Medium | Maximum |
| Production at scale | Good | Limited | Optimal |
| Reproducibility | High | Medium | Maximum (git) |
❓ Does AutoML automatically pick the correct optimization metric?
No. The user must choose the target metric based on the business problem. AutoML maximizes whichever metric you tell it to, but choosing the wrong metric (e.g., accuracy for imbalanced classes) will produce a model that is useless even if the metric is high.
❓ Does AutoML eliminate the need for data scientists?
No. AutoML automates algorithm and hyperparameter selection, but data scientists are still needed to define the problem, prepare data, interpret results, and manage models in production. AutoML democratizes ML, it does not replace it.
❓ Are Designer and AutoML the same thing?
No. Designer is a visual canvas where you build the pipeline step by step by connecting modules. AutoML automatically decides which algorithms to try and how to configure them. Designer gives visual control; AutoML gives full automation.
Did you get this topic?
Test what you just learned
A telecom company wants to predict which customers will cancel their subscription next month. They have a dataset of 500,000 customers where historically only 3% canceled. A business analyst (with no Python experience) will build the model. Which tool and metric are most appropriate?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.