AWS Machine Learning Specialty (AWS) study plan

Map each major SageMaker built-in algorithm family to the problem type it solves (regression, classification, clustering, dimensionality reduction, NLP, computer vision, forecasting)

Describe the SageMaker training job lifecycle end to end: estimator, container image, input channels, output artifacts, and distributed training modes

Compute accuracy, precision, recall, F1, ROC-AUC, and RMSE by hand from a small confusion matrix and state when each one misleads

Profile a real dataset with Data Wrangler and load it into S3 and Feature Store

Train three different built-in algorithms on the same dataset and compare validation results on a common metric

Run an automatic model tuning job and interpret the objective metric, ranges, and best-training-job output

Apply regularization and class-imbalance techniques (weighting, resampling, threshold shifting) and measure the effect

Detect and remediate skewed features, missing values, and outliers using Data Wrangler and SageMaker Processing

Read a correlation matrix and a PCA projection to justify dropping or keeping specific features

Package a Bring-Your-Own-Script or framework estimator and run it through a complete training job

Configure a Bayesian hyperparameter search with sensible static and categorical ranges and a defensible objective metric

Generate a SageMaker Clarify explainability and bias report and interpret the feature-attribution output

Establish a per-run record of parameters, metrics, and artifacts so experiments are reproducible

Compare Parquet, ORC, Avro, and CSV on query cost and compression, and select a partition scheme for a stated query pattern

Query S3 data with Athena and cut bytes scanned using partitioning plus columnar compression

Convert raw semi-structured logs into partitioned columnar data with Glue ETL or DataBrew recipes

Answer one stated business question from the curated data with a chart or dashboard

Stream records through Kinesis Data Streams or Firehose into S3 and query the landed data with Athena

Orchestrate a multi-step pipeline with Glue workflows or Step Functions and configure retry and failure handling

Apply imputation, encoding, and scaling at scale with Spark on Glue or EMR Serverless

Distinguish Feature Store online versus offline stores and explain which serves training versus low-latency inference

Deploy one trained model to a real-time endpoint and to a batch transform job, then compare latency, cost, and operational overhead

Configure endpoint autoscaling, instance selection, and a multi-model endpoint for a stated traffic pattern

Lock down the pipeline with IAM execution roles, VPC endpoints, KMS encryption, and least-privilege S3 bucket policies

Automate training, registration, and deployment with SageMaker Pipelines and the Model Registry

Configure model monitor for data drift and bias drift and interpret what a triggered violation means

Diagnose a failing or degraded endpoint using CloudWatch metrics, logs, and Debugger output

Choose between A/B testing, shadow deployment, and blue/green rollout for a given risk profile and justify the choice

Close two self-identified weak modeling areas (for example time-series forecasting, recommendation, or text algorithms)

Complete two full-length timed practice exams and log every missed question by domain

Re-derive each missed answer from primary AWS documentation and record the exact doc section

Explain the main SageMaker cost drivers (training hours, endpoint uptime, storage, and data processing) and identify where to cut spend

Rehearse selection rules: Glue versus EMR versus Athena, batch versus real time, built-in versus custom container, single versus multi-model endpoint