← Back to MLA-C01 Overview
Feature Engineering & Optimization Scenario-Based Selector Guide
Matrix 1 of 17This scenario-based selector guide maps common dataset patterns directly to the correct feature engineering and optimization techniques tested on the AWS exam.
| Dataset Condition | The Problem | **Correct Technique** | AWS Exam Keyword Triggers |
|---|---|---|---|
| Unordered text words *(Low count)* | Models only read numbers. | **One-Hot Encoding** | `Fabric type`, `City names`, `Device type`. |
| Ordered text words *(e.g., Low, Medium, High)* | Order holds mathematical meaning. | **Ordinal / Label Encoding** | `Education level`, `Survey ratings`, `T-shirt sizes`. |
| Too many unique words *(High Cardinality)* | One-hot creates too many columns. | **Target Encoding / Embeddings** | `Zip codes`, `User IDs`, `IP addresses`. |
| Dense combined strings | Information is mashed together. | **Feature Splitting** | `Timestamps`, `Full Addresses`, `Product codes`. |
| Dataset Condition | The Problem | **Correct Technique** | AWS Exam Keyword Triggers |
| :--- | :--- | :--- | :--- |
| Vastly differing ranges | Big numbers drown out small numbers. | **Standardization (Z-score)** | `Distance-based models`, `KNN`, `House size vs. Bedrooms`. |
| Bounded range required | Model requires strict 0-to-1 input boundaries. | **Min-Max Normalization** | `Image pixel arrays (0-255)`, `Neural network input layers`. |
| Heavily skewed distribution | Long tail of extreme outliers breaks math. | **Log Transformation** | `Income distribution`, `Company revenues`, `Population sizes`. |
| Continuous range variance | Fine-grained values reduce accuracy. | **Binning / Bucketing** | `Age into life-stages`, `Credit scores into tiers`. |
| Machine Learning Condition | The Problem | **Correct Technique** | AWS Exam Keyword Triggers |
| :--- | :--- | :--- | :--- |
| Tabular model overfitting | Model memorizes training noise. | **L1 Regularization (Lasso)** | `Drive weights to zero`, `Enforce feature selection`. |
| Neural network overfitting | Connected nodes become overly dependent. | **Dropout Layers** | `Randomly deactivate neurons`, `Force robust learning paths`. |
| Retraining loss drift | Performance degrades as epochs continue. | **Early Stopping** | `Monitor validation loss`, `Halt training automatically`. |
| Edge / Mobile deployment | Model file is too heavy for phone memory. | **Pruning & Quantization** | `SageMaker Neo`, `Convert 32-bit to 8-bit`, `Cut dead layers`. |
💡 Quick Elimination Rules for the Exam
- Is it an image pixel or audio stream? ➔ **Eliminate** One-Hot Encoding and PCA.
- Is it a classification problem (Yes/No)? ➔ **Eliminate** RMSE and MAE.
- Is it an unsupervised problem (No labels)? ➔ **Eliminate** Linear Learner and XGBoost.
- Is it an edge device (IoT/Mobile) problem? ➔ **Eliminate** standard EC2 endpoints. Look for **SageMaker Neo**.