AWS Certified Machine Learning - Specialty Quick Facts (2025)

Exam Domain Breakdown

Domain 1: Data Engineering (20% of the exam)

1.1 Create data repositories for ML.

• Identify data sources (for example, content and location, primary sources such as user data).
• Determine storage mediums (for example, databases, Amazon S3, Amazon Elastic File System [Amazon EFS], Amazon Elastic Block Store [Amazon EBS]).

1.1 summary:
In this section, you will learn the foundational skills needed to identify and collect the right data for ML workflows. Being able to differentiate between structured, semi-structured, and unstructured sources ensures you know how to properly bring data into an ML-ready environment. Examples include customer transaction logs, images, or IoT sensor streams, each with varying requirements for access and storage.

You will also focus on selecting the correct storage solution within AWS. Knowledge of when to use S3 object storage versus block storage like EBS or a file system such as EFS plays a big role in keeping ML workflows both efficient and cost-effective. The emphasis is on aligning storage mediums with data accessibility, durability, and the specific processing needs of your ML use case.

1.2 Identify and implement a data ingestion solution.

• Identify data job styles and job types (for example, batch load, streaming).
• Orchestrate data ingestion pipelines for batch and streaming ML workloads (using Kinesis, Data Firehose, EMR, Glue, Amazon Managed Service for Apache Flink).
• Schedule jobs.

1.2 summary:
This section focuses on how data enters your system and becomes available for machine learning use cases. You will learn about different ingestion styles, whether continuous streams or scheduled batch uploads, and how to manage them effectively. Services like Amazon Kinesis and Glue make it easier to capture, prepare, and move data reliably at scale.

You’ll also gain the knowledge to design architectures that automate ingestion with scheduling, ensuring consistent delivery into repositories. Understanding when to use managed services like Firehose or distributed compute with EMR provides you flexibility to support both real-time model training or periodic batch-oriented ML workloads with confidence.

1.3 Identify and implement a data transformation solution.

• Transform data in transit (ETL, AWS Glue, Amazon EMR, AWS Batch).
• Handle ML-specific data by using MapReduce (for example, Apache Hadoop, Apache Spark, Apache Hive).

1.3 summary:
This section teaches you how to prepare raw, ingested data for actual ML modeling. Real-world datasets typically need cleaning, restructuring, or reformatting, which is achieved through ETL (extract, transform, load) processes. You will learn how AWS Glue automates schema discovery and transformation and how EMR clusters process large-scale data.

Special attention is placed on handling complex ML-specific workloads using Spark or MapReduce frameworks. These tools let you scale your preprocessing tasks across large, high-volume datasets. Establishing a structured transformation pipeline ensures models receive data that is consistent, normalized, and ready for advanced machine learning workflows.

Domain 2: Exploratory Data Analysis (24% of the exam)

2.1 Sanitize and prepare data for modeling.

• Identify and handle missing data, corrupt data, and stop words.
• Format, normalize, augment, and scale data.
• Determine whether there is sufficient labeled data and apply mitigation strategies.
• Use data labeling tools (for example, Amazon Mechanical Turk).

2.1 summary:
This section covers the practical steps of cleaning and enhancing data for use in modeling. You’ll explore the value of dealing with missing values, correcting corrupted data, and removing unnecessary stop words in text to produce cleaner datasets. Learning how to format, scale, and normalize datasets ensures models train effectively without bias caused by discrepancies in input features.

You will also evaluate if the available dataset has enough labels to support accurate model building. Mitigation strategies include augmenting data with synthetic examples or utilizing labeling services like Amazon Mechanical Turk. These skills ensure you can maximize the training potential of your datasets, improving quality and fairness of ML outputs.

2.2 Perform feature engineering.

• Identify and extract features from datasets (including text, speech, images, and public datasets).
• Analyze and evaluate feature engineering concepts such as binning, tokenization, outliers, synthetic features, one-hot encoding, and dimensionality reduction.

2.2 summary:
Here, you focus on the art of extracting and refining features to maximize model performance. From creating derived variables to tokenizing text or applying one-hot encoding to categorical values, feature engineering ensures that the most informative inputs are passed to models. You’ll also learn how to manage outliers and apply synthetic feature creation to improve predictive power.

The section highlights dimensionality reduction techniques such as PCA, which streamline complex datasets into more efficient representations. Emphasis is placed on identifying which features offer the highest predictive value while minimizing noise, a skill critical in building models that are both accurate and computationally efficient.

2.3 Analyze and visualize data for ML.

• Create graphs (for example, scatter plots, time series, histograms, box plots).
• Interpret descriptive statistics (for example, correlation, summary statistics, p-value).
• Perform cluster analysis (for example, hierarchical, diagnosis, elbow plot, cluster size).

2.3 summary:
This section emphasizes developing insights through visualization and statistical exploration. You will create plots and charts that reveal patterns, trends, and distributions in your dataset. Understanding how variability and relationships appear visually helps uncover the signals needed for stronger models.

You will also analyze quantitative summaries and employ cluster analysis methods to detect natural groupings within the data. This combination of statistics and visual exploration enables you to deeply understand your dataset, choose better algorithms, and guide future modeling choices.

Domain 3: Modeling (36% of the exam)

3.1 Frame business problems as ML problems.

• Determine when to use and when not to use ML.
• Differentiate supervised vs. unsupervised learning.
• Select among classification, regression, forecasting, clustering, recommendation, and foundation models.

3.1 summary:
This section teaches you to bridge practical business needs with machine learning solutions. You will identify scenarios where ML is the optimal approach and when simpler rule-based systems might suffice. Recognizing whether a problem is best suited for supervised or unsupervised methods is a core outcome.

You’ll also explore mapping specific business use cases to the right ML categories, such as recommendation engines for e-commerce, regression for demand forecasting, and clustering for customer segmentation. This alignment ensures solutions are not only technically sound but also deliver genuine business improvements.

3.2 Select the appropriate model(s) for a given ML problem.

• Common models include XGBoost, logistic regression, k-means, linear regression, decision trees, random forests, RNN, CNN, ensembles, transfer learning, and large language models (LLMs).
• Express the intuition behind models.

3.2 summary:
This section focuses on model selection. You’ll gain a strong overview of common ML models and the problems each one handles best. From traditional approaches such as decision trees and regressions to deep learning architectures like CNNs and RNNs, you will learn how to choose wisely based on input type and outcome.

Just as important, you’ll develop intuition to explain how models work, not just what they output. Being able to express why XGBoost might outperform linear regression or why transfer learning accelerates training showcases mastery in communicative and practical ML reasoning.

3.3 Train ML models.

• Split data into training and validation sets using techniques such as cross-validation.
• Understand optimization techniques for ML training like gradient descent, loss functions, convergence.
• Choose appropriate compute resources (CPU or GPU, distributed or local compute like Spark).
• Update and retrain models in batch or real time.

3.3 summary:
This section covers the core process of creating usable predictive models. Splitting data correctly into train, test, and validation sets ensures robust evaluation. You’ll study how optimization methods like gradient descent achieve convergence and how loss metrics guide training progress.

Additionally, you’ll learn how to select the right compute environment, whether GPUs for deep learning or CPU clusters for distributed processing. Retraining strategies help ensure that models evolve as data evolves, preparing you to manage ML in both dynamic real-time and scheduled batch contexts.

3.4 Perform hyperparameter optimization.

• Apply regularization strategies like dropout and L1/L2.
• Conduct cross-validation and model initialization.
• Optimize neural networks (layers, nodes, learning rate, activation).
• Optimize tree-based models (tree depth, number of trees).
• Optimize linear models (learning rate).

3.4 summary:
This section addresses fine-tuning model performance. You’ll learn to systematically test and adjust hyperparameters such as learning rates, dropout rates, or the number of hidden layers. Regularization techniques like L1, L2, and dropout help reduce overfitting while maintaining predictive accuracy.

By experimenting with parameters that govern neural networks, tree models, or linear regressions, you’ll discover how to balance flexibility and precision. Through cross-validation, adjustments are validated objectively, allowing you to maximize generalization and confidently deploy optimized ML models.

3.5 Evaluate ML models.

• Avoid overfitting or underfitting by detecting bias and variance issues.
• Evaluate metrics like accuracy, precision, recall, RMSE, AUC-ROC, and F1 score.
• Interpret confusion matrices and conduct offline and online evaluation (A/B testing).
• Use cross-validation to compare models and interpret criteria like training efficiency and engineering costs.

3.5 summary:
This section teaches you how to determine if models perform well. Metrics like accuracy, recall, and F1 score allow you to quantify predictive strength. Confusion matrices help break down classification performance across true positives and negatives, giving a clearer picture than raw accuracy alone.

Advanced strategies such as A/B testing and ROC curve analysis ensure models are validated both in isolation and in production-like environments. By balancing model quality with cost and time efficiency, you’ll know how to make smart choices that support sustainable ML deployment.

Domain 4: Machine Learning Implementation and Operations (20% of the exam)

4.1 Build ML solutions for performance, availability, scalability, resiliency, and fault tolerance.

• Log and monitor AWS environments using CloudTrail and CloudWatch.
• Build error monitoring solutions.
• Deploy across multiple Regions and Availability Zones.
• Create AMIs, golden images, and Docker containers.
• Implement scaling, load balancing, and rightsizing.

4.1 summary:
This section emphasizes how to prepare ML solutions to run effectively in production. You’ll study monitoring practices using native AWS services and learn how to deploy across multiple locations for high availability. Concepts such as autoscaling groups, load balancing, and right-sizing create robust systems adapted to variable workloads.

Containerization and automation also play a significant role. Using Docker or AMIs makes deployments repeatable and portable, a necessity for enterprise-scale ML applications. This knowledge ensures your ML environments are resilient, cost-efficient, and secure.

4.2 Recommend and implement the appropriate ML services and features for a given problem.

• Leverage AWS ML application services like Amazon Polly, Lex, Transcribe, and Q.
• Decide when to build custom models or use SageMaker built-in algorithms.
• Understand cost considerations and infrastructure choices such as Spot Instances.

4.2 summary:
This section equips you with the judgment to know when to use prebuilt solutions versus creating custom models. While managed AI services such as Amazon Lex or Polly speed time to market, other use cases still require custom deep learning via SageMaker. Knowing this trade-off makes you more flexible as an ML engineer.

You also gain insight into infrastructure and cost alignment. For example, training models on Spot Instances significantly reduces compute costs, while SageMaker built-ins save engineering effort. This combination helps you evaluate both functional and cost considerations wisely.

4.3 Apply basic AWS security practices to ML solutions.

• Implement IAM permissions and S3 bucket policies.
• Use security groups and VPCs.
• Apply encryption and apply anonymization.

4.3 summary:
This section highlights how to protect ML pipelines through AWS security best practices. You’ll learn how IAM roles and bucket policies enforce proper access controls. VPC-level defenses using security groups further safeguard model endpoints and storage.

Encryption and anonymization protect sensitive data processed in ML workflows, which is especially important for compliance in industries like healthcare or finance. These skills ensure ML does not just perform effectively but also respects data integrity and privacy standards.

4.4 Deploy and operationalize ML solutions.

• Expose endpoints and use them for interaction.
• Conduct A/B testing on deployed ML models.
• Retrain pipelines, detect performance drops, and troubleshoot issues.

4.4 summary:
This section covers operational excellence once models are in use. You’ll learn about endpoint management and how to deploy ML models so they can serve traffic effectively. A/B testing methods validate changes, ensuring improvements before large-scale rollouts.

Troubleshooting skills are also crucial. Building monitoring workflows to catch performance degradations early allows you to respond proactively by retraining or adjusting pipeline parameters. By mastering these techniques, you ensure ML models remain effective and contribute long-term business value.