AIF-C01 Domain 1 Deep Dive: AI/ML Fundamentals and AWS AI Services (20%)
Domain 1 of the AWS AI Practitioner exam tests your grasp of classical machine learning concepts — supervised vs. unsupervised learning, the ML lifecycle, model evaluation metrics — and your ability to select the right AWS pre-built AI service for a given scenario. This deep dive covers every subtopic with exam-focused explanations and the service selection patterns that appear most on AIF-C01.
Domain 1 of the AWS Certified AI Practitioner (AIF-C01) accounts for 20% of the exam — roughly 13 questions out of 65. That makes it the third-largest domain, behind Applications of Foundation Models (28%) and Fundamentals of Generative AI (24%). Many candidates underestimate Domain 1 because they assume it is just "basic ML theory." In reality, this domain has two distinct components that together require precise knowledge: the conceptual layer (what is supervised learning, what is overfitting, how does cross-validation work) and the practical AWS layer (which managed AI service solves which business problem without custom model training). Getting both right is how you pass this domain. This deep dive covers everything you need.
- ML Paradigms: Supervised, Unsupervised, Reinforcement
- The ML Development Lifecycle
- Key Algorithms and When They Apply
- Model Evaluation Metrics
- Overfitting, Underfitting, and Regularization
- AWS Pre-Built AI Services Map
- Amazon SageMaker: The Custom ML Platform
- Service Selection Patterns
- 5 Practice Questions with Explanations
ML Paradigms: Supervised, Unsupervised, Reinforcement
The exam consistently tests whether you can identify the correct ML paradigm for a given problem. Here is the framework the exam uses:
Training data includes labeled examples (input → known output). The model learns a mapping function. Used when you have historical data with correct answers.
Examples: Email spam classification, loan default prediction, image classification, sentiment analysis.
Training data has no labels. The model finds patterns, groupings, or structure on its own. Used when you do not know the categories in advance.
Examples: Customer segmentation, anomaly detection, topic modeling, recommendation systems.
An agent learns by interacting with an environment, receiving rewards for good actions and penalties for bad ones. Learns optimal behavior through trial and error.
Examples: Game playing (AlphaGo), robotic control, trading strategies, ad bidding optimization.
AIF-C01 sometimes mentions semi-supervised learning (small labeled dataset + large unlabeled dataset) and self-supervised learning (model creates its own labels from the data structure — used in training foundation models). Know the difference: self-supervised is how LLMs are pretrained (predicting the next token), while semi-supervised is used when labeling is expensive but you have abundant raw data.
The ML Development Lifecycle
The exam uses a standard lifecycle model that you should know cold. Questions often ask what step should happen before or after another, or which AWS service supports a specific lifecycle phase.
- Business problem definition — Translate business need into an ML problem type (classification, regression, clustering, etc.)
- Data collection — Gather training data from relevant sources. AWS: S3 for storage, Glue for cataloging, Kinesis for streaming data.
- Data preprocessing / feature engineering — Handle missing values, normalize features, encode categoricals, create derived features. AWS: SageMaker Data Wrangler, Glue DataBrew.
- Model training — Select algorithm, train on prepared dataset. AWS: SageMaker Training Jobs, SageMaker JumpStart (pre-trained models).
- Model evaluation — Measure accuracy, precision, recall, F1, AUC-ROC on a held-out test set. AWS: SageMaker Clarify (also for bias evaluation).
- Model deployment — Serve predictions in production. AWS: SageMaker Endpoints (real-time), SageMaker Batch Transform (batch).
- Monitoring and retraining — Detect model drift, trigger retraining pipelines. AWS: SageMaker Model Monitor.
Key Algorithms and When They Apply
You do not need to implement these algorithms. You need to know which algorithm type solves which category of problem and recognize the exam's terminology for each.
| Algorithm | Type | Best For |
|---|---|---|
| Linear / Logistic Regression | Supervised | Regression (continuous output) / Binary classification |
| Decision Trees / Random Forest | Supervised | Classification with interpretable rules; handles mixed data types well |
| XGBoost | Supervised (ensemble) | Structured/tabular data; frequently referenced in SageMaker built-in algorithms |
| K-Means Clustering | Unsupervised | Customer segmentation; grouping without predefined labels |
| Principal Component Analysis (PCA) | Unsupervised | Dimensionality reduction; reducing features while preserving variance |
| Convolutional Neural Networks (CNN) | Supervised (deep learning) | Image and video analysis; spatial pattern recognition |
| Recurrent Neural Networks (RNN) / LSTM | Supervised (deep learning) | Sequential data: time series, text sequences (pre-transformer era) |
| Isolation Forest | Unsupervised | Anomaly detection in unlabeled datasets |
Model Evaluation Metrics
The exam tests whether you can identify the appropriate metric for a given scenario — not calculate it by hand. Here are the critical metrics and when to use each:
Classification Metrics
Accuracy = (True Positives + True Negatives) / Total. Use when classes are balanced. Misleading on imbalanced datasets (e.g., 99% accuracy detecting fraud when only 1% of transactions are fraudulent — useless model).
Precision = TP / (TP + FP). Out of everything predicted positive, how many were actually positive? Optimize when false positives are costly (e.g., marking legitimate emails as spam).
Recall (Sensitivity) = TP / (TP + FN). Out of all actual positives, how many did we catch? Optimize when false negatives are costly (e.g., missing a cancer diagnosis).
F1 Score = Harmonic mean of Precision and Recall. Useful when you need balance between precision and recall, especially with imbalanced classes.
AUC-ROC (Area Under the Receiver Operating Characteristic Curve). Measures overall model discrimination ability across all thresholds. AUC of 1.0 = perfect, 0.5 = random. Use when you need threshold-independent evaluation.
Regression Metrics
RMSE (Root Mean Square Error) — Penalizes large errors more heavily. MAE (Mean Absolute Error) — More robust to outliers. R² (R-squared) — Proportion of variance explained; 1.0 = perfect fit.
When a scenario involves medical screening, fraud detection, or security threats — anything where missing a true case (false negative) has catastrophic consequences — the answer is maximize Recall. When the scenario involves spam filters, recommendation systems, or ad targeting — where bombarding users with false positives is the problem — the answer is maximize Precision.
Overfitting, Underfitting, and Regularization
These three concepts appear repeatedly in Domain 1 questions. Here is the exam-focused summary:
| Problem | Symptom | Fix |
|---|---|---|
| Overfitting | High training accuracy, low test accuracy ("memorized the training set") | More training data, regularization (L1/L2), dropout, reduce model complexity, early stopping |
| Underfitting | Low training accuracy and low test accuracy ("too simple") | More complex model, more features, less regularization, train longer |
| Well-fitted | Good training accuracy and comparable test accuracy | Monitor for drift over time |
Cross-validation (especially k-fold cross-validation) is the standard technique for estimating how well a model will generalize to unseen data. The dataset is split into k subsets; the model trains on k-1 folds and validates on the remaining fold, rotating until every fold has been used for validation. This gives a more reliable performance estimate than a single train/test split.
AWS Pre-Built AI Services Map
This is the most exam-critical section of Domain 1. AWS provides a set of pre-built AI services that require no ML expertise and no model training — you simply call an API with your data and receive predictions. The exam tests whether you can match a business requirement to the correct service.
Vision Services
- Image and video analysis: object detection, scene detection, facial analysis, facial comparison
- Content moderation (detect inappropriate content in user-uploaded images/video)
- Text detection in images (OCR-like for printed text in images)
- Celebrity recognition, personal protective equipment (PPE) detection
- Key distinction: Rekognition uses AWS-provided models — you cannot train custom models for general vision tasks here (use SageMaker for that)
- Rekognition Custom Labels: the exception — fine-tune Rekognition on your own labeled images for specific object categories
Natural Language Processing Services
- Sentiment analysis (positive / negative / neutral / mixed)
- Entity extraction (people, places, organizations, dates, quantities)
- Key phrase extraction, language detection
- PII (Personally Identifiable Information) detection and redaction
- Comprehend Medical: specialized NLP for medical records (extract diagnoses, medications, dosages)
- Comprehend Custom: train custom classifiers or entity recognizers on your own labeled text data
Speech Services
- Automatic speech recognition (ASR): audio/video → text
- Speaker diarization (identify who is speaking when)
- Custom vocabulary for domain-specific terms (medical, legal, technical)
- Transcribe Medical: HIPAA-eligible version for medical transcription
- Text-to-speech: convert text into lifelike speech audio
- Supports multiple languages and voices (neural vs. standard)
- SSML (Speech Synthesis Markup Language) support for pronunciation control
- Common scenario: accessibility features, voice interfaces, podcast generation from blog posts
Translation and Language Services
- Neural machine translation: text from one language to another
- Supports 75+ language pairs
- Custom terminology: ensure brand names and technical terms are translated consistently
- Common scenario: localizing e-commerce product descriptions, customer support multilingual content
Conversational AI
- Build conversational interfaces (chatbots) using the same technology as Alexa
- Supports voice and text channels
- Intent recognition + slot filling (extracting entities from user utterances)
- Key distinction vs. Bedrock: Lex is for rule-based/intent-driven dialogue flows; Bedrock is for open-ended generative AI conversations
Document Intelligence
- Extract text and structured data from scanned documents and forms
- Understands form structure (key-value pairs), tables, and layout
- HIPAA-eligible for healthcare documents
- vs. Rekognition text detection: Textract is for document structure extraction; Rekognition text is for detecting text in natural images (signs, labels)
Forecasting and Recommendations
- Time-series forecasting: sales, inventory, energy demand, traffic
- Automatically selects best algorithm (DeepAR, ARIMA, ETS, etc.)
- Integrates related time series and item metadata to improve accuracy
- Real-time personalized recommendations using your own data
- Product recommendations, content recommendations, user segmentation
- Uses the same technology as Amazon.com's recommendation engine
- Scenario: "An e-commerce company wants to show each customer personalized product recommendations based on their browsing history" → Amazon Personalize
Fraud and Anomaly Detection
- Build fraud detection models with no ML experience
- Pre-trained models for online fraud, account takeover, payment fraud
- Uses your historical fraud data to customize the model
Amazon SageMaker: The Custom ML Platform
SageMaker sits above the pre-built AI services — it is for teams that need to train custom models on their own data. Domain 1 tests SageMaker at a high level. You do not need to know SageMaker internals deeply (that belongs to the ML Engineer Associate), but you must understand the component roles:
| Component | Role |
|---|---|
| SageMaker Studio | Web-based IDE for ML development (notebooks, experiments, pipelines) |
| SageMaker JumpStart | Pre-trained model hub + solution templates; fine-tune foundation models (including Llama, Falcon) on your data |
| SageMaker Autopilot | AutoML: automatically tries multiple algorithms and hyperparameters |
| SageMaker Clarify | Bias detection, model explainability (SHAP values) |
| SageMaker Model Monitor | Detect data drift and model quality drift in production |
| SageMaker Pipelines | MLOps: define, orchestrate, and reuse ML workflows |
Service Selection Patterns
Domain 1 service selection questions follow predictable patterns. Memorize these decision rules:
Does the problem fit a known category (image analysis, NLP, speech, translation)? → Use a pre-built service (Rekognition, Comprehend, Transcribe, Translate). Only move to SageMaker when you need domain-specific custom behavior the pre-built services cannot achieve.
Extracting structured data from a form/invoice/contract? → Textract. Detecting text (a sign, a label, printed text in a photo)? → Rekognition text detection.
Structured conversational flow with defined intents (customer service bot, booking system)? → Amazon Lex. Open-ended generative responses, summarization, content generation? → Amazon Bedrock.
Online transactions, account takeover, payment fraud with labeled fraud history? → Amazon Fraud Detector. General anomaly detection without labeled data (IT operations, manufacturing defects)? → SageMaker with Isolation Forest or Random Cut Forest.
5 Practice Questions with Explanations
Q1. A retail company wants to automatically detect whether customer-uploaded product photos contain inappropriate content before publishing them on their marketplace. Which AWS service should they use?
A) Amazon Comprehend — content moderation
B) Amazon Rekognition — content moderation feature
C) Amazon SageMaker — custom image classifier
D) Amazon Textract — document analysis
Q2. A healthcare company needs to extract medication names, dosages, and diagnoses from unstructured clinical notes at scale. Which AWS service is MOST appropriate?
A) Amazon Transcribe Medical
B) Amazon Textract
C) Amazon Comprehend Medical
D) Amazon Rekognition
Q3. A data science team is evaluating a fraud detection model. The business requirement states that missing a fraudulent transaction is far more costly than incorrectly flagging a legitimate one. Which metric should they prioritize when evaluating the model?
A) Precision
B) Accuracy
C) Recall
D) R²
Q4. A training accuracy of 98% and a validation accuracy of 61% on the same model indicates which issue?
A) Underfitting — the model needs more complexity
B) Overfitting — the model has memorized the training data
C) Data leakage — training and test sets are contaminated
D) Class imbalance — the dataset needs resampling
Q5. An e-commerce company wants to automatically suggest personalized product recommendations to each customer on their homepage based on browsing and purchase history. They want to use a managed AWS service that requires no custom model development. Which service is MOST appropriate?
A) Amazon SageMaker Autopilot
B) Amazon Personalize
C) Amazon Forecast
D) Amazon Bedrock with Claude
CertLand's AIF-C01 question bank has 95 questions covering Domain 1. Use domain-filtered practice mode to focus exclusively on AI/ML fundamentals until your score is consistently above 75% before moving on to Domains 2 and 3.
Comments
No comments yet. Be the first!
Comments are reviewed before publication.