Full course summary: SageMaker algorithms, AWS AI services, evaluation metrics, important formulas, and common exam traps.
📝 Exam Prep — Lesson 11
Lesson 11: Comprehensive Cheat Sheet
AWS Certified Machine Learning - Specialty Exam Prep
Part 4: Review & Exam Strategy
xdev.asia
1. SageMaker Built-in Algorithms Master Table
Algorithm Type Best For Input
XGBoost Supervised (C/R) Tabular data, competitions CSV, LibSVM Linear Learner Supervised (C/R) High-dimensional, fast CSV, RecordIO Random Cut Forest Unsupervised Anomaly detection CSV, RecordIO K-Means Unsupervised Customer segmentation CSV, RecordIO PCA Dimensionality Reduction Feature reduction CSV, RecordIO Factorization Machines Supervised (C/R) Sparse data, recommendations RecordIO only DeepAR+ Supervised Time series forecasting JSON Lines BlazingText Supervised / Unsupervised Text classification, word2vec Text files Object2Vec Supervised Semantic similarity JSON Lines Image Classification Supervised Image labels RecordIO, raw images Object Detection Supervised Bounding boxes RecordIO, JSON Semantic Segmentation Supervised Pixel-level classification Images + masks LDA Unsupervised Topic modeling CSV, RecordIO NTM Unsupervised Neural topic modeling CSV, RecordIO IP Insights Unsupervised IP address anomaly CSV
2. AWS AI Services — No-Code ML
Service Purpose Output
Amazon Rekognition Image/video analysis Labels, faces, text, moderation Amazon Textract Document extraction Text, forms, tables Amazon Comprehend NLP, text analytics Entities, sentiment, topics Amazon Translate Machine translation Translated text Amazon Transcribe Speech to text Transcription, subtitles Amazon Polly Text to speech Audio Amazon Lex Conversational AI (chatbot) Intent, slots Amazon Kendra Intelligent search Answers, documents Amazon Personalize Recommendations Item rankings, user recs Amazon Forecast Time series forecasting Predictions + confidence Amazon Lookout for Vision Visual anomaly (manufacturing) Pass/Fail, anomaly map Amazon Lookout for Equipment Equipment anomaly (IoT) Anomaly scores
3. Evaluation Metrics Quick Reference
Metric Formula Use When
Accuracy (TP+TN)/(TP+TN+FP+FN) Balanced classes Precision TP/(TP+FP) FP is costly (spam filter) Recall (Sensitivity) TP/(TP+FN) FN is costly (cancer diagnosis) F1 Score 2×(P×R)/(P+R) Imbalanced classes AUC-ROC Area under TPR vs FPR curve Overall classifier quality RMSE √(Σ(yᵢ-ŷᵢ)²/n) Regression, penalizes outliers MAE Σ|yᵢ-ŷᵢ|/n Regression, robust to outliers MAPE Σ|yᵢ-ŷᵢ|/|yᵢ| × 100% Forecasting, interpretable %
4. Common Exam Traps
Trap What the Exam Says Correct Answer
Imbalanced data + accuracy "Model has 99% accuracy" (fraud) Use Precision/Recall/F1, not accuracy FM input format Factorization Machines Requires RecordIO ONLY (not CSV) Managed vs custom "quickest to implement" Prefer managed (Personalize, Forecast) Overfitting fix Training accuracy high, validation low Regularization (L1/L2) or more data SageMaker + Internet "secure environment, no internet" VPC + Network Isolation + VPC Endpoints Ground Truth labeling "reduce labeling cost" Auto-labeling (active learning) in GT Bias in model "identify bias before deployment" SageMaker Clarify Multiple models behind one endpoint "save cost, single endpoint" Multi-Model Endpoint (MME)
Scenario Best Choice
Large tabular training data S3 + CSV or Parquet; RecordIO for SageMaker Real-time streaming data ingestion Kinesis Data Streams → Firehose → S3 Ad-hoc SQL queries on S3 data Amazon Athena ETL transformation → Feature Store AWS Glue (Spark ETL) → SageMaker Feature Store Business intelligence / dashboards Amazon QuickSight Data warehouse for ML features Amazon Redshift → ML (Redshift ML)
Exam tip: Remember: when the question says "fastest / easiest / no code" → AWS Managed AI Service. When it says "custom model / flexibility / bring your own" → SageMaker. This is the most important decision boundary.