Built-in algorithms, script mode, or bring your own
Lesson 2 of 5 in Choosing a Modeling Approach.
When the problem genuinely needs your model trained on your data, SageMaker AI offers three rungs of control. Built-in algorithms: AWS wrote the model code; you bring data channels and hyperparameters. Script mode: you bring a training script; AWS brings the framework container and the managed fleet. Bring your own container (BYOC): you ship the whole image — full control, full responsibility. MLA-C02 trimmed BYOC depth from the exam; know that the rung exists and what forces you onto it.
| Option | You bring | AWS provides | Pick when |
|---|---|---|---|
Built-in algorithm | Data in S3 + hyperparameters | The algorithm, its container, tuned implementations | A built-in matches the problem shape (tabular, clustering, forecasting, anomaly) |
Script mode | Your training script (and requirements) | Framework container (PyTorch, TensorFlow, Hugging Face, XGBoost, scikit-learn) + managed infra | You need custom architecture or loss, but standard frameworks suffice |
Your own container | The full Docker image | Orchestration: instances, channels, artifact handling | Exotic runtimes or dependencies no framework container covers |
XGBoost
Gradient-boosted trees for tabular classification and regression — the default serious answer for structured data. Exam cue: tabular features, labeled outcomes, strong accuracy without deep learning. Also available as a framework (script mode) when you want your own training loop around it.
Linear Learner
Linear and logistic models at scale: fast, cheap, and interpretable — coefficients tell you what drives predictions. Exam cue: simple baseline, explainability requirement, or very high-dimensional sparse data.
K-Means
Unsupervised clustering into k groups. Exam cue: "segment customers," "group similar items," no labels anywhere in the stem.
Random Cut Forest
Unsupervised anomaly detection that scores how far each point falls from learned structure — works on streams too. Exam cue: "detect unusual," "flag abnormal," no labeled failures to learn from.
DeepAR
Probabilistic forecasting that trains one model across many related time series and returns prediction intervals, not just point estimates. Exam cue: thousands of SKUs/households/servers, each with its own history.
Interactive sorting exercise: Route each problem to the built-in that owns it — or off the built-in rung entirely.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.