Bài 2: Data Transformation & Feature Engineering

1. Data Transformation trong ML Pipeline

Trước khi train model, raw data phải qua nhiều bước transformation. Đây là nguồn gốc của câu nói nổi tiếng: "Garbage in, garbage out". Đề thi MLS-C01 thường hỏi kỹ thuật xử lý data và tools phù hợp.

2. SageMaker Processing Jobs

SageMaker Processing Jobs là managed service để chạy data processing scripts (Python, Spark) trên ephemeral compute clusters.

Processor Type	Framework	Use Case
ScriptProcessor	Custom Docker container	Any custom script
SKLearnProcessor	scikit-learn	Classic ML preprocessing
PySparkProcessor	Apache Spark	Large-scale distributed processing
FrameworkProcessor	TensorFlow/PyTorch	Deep learning data prep

SageMaker Processing Job Flow:

S3 (input data)
      ↓
┌─────────────────────┐
│  Processing Job     │
│  (compute cluster)  │
│                     │
│  - Preprocess data  │
│  - Feature engineer │
│  - Split train/test │
└─────────────────────┘
      ↓
S3 (output: train/, validation/, test/)

3. Xử lý Missing Values

Strategy	Method	When to Use
Deletion	Drop rows/columns	MCAR, ít missing (<5%)
Mean/Median Imputation	Điền giá trị trung bình	Numeric, MCAR/MAR
Mode Imputation	Điền giá trị phổ biến nhất	Categorical
KNN Imputation	Dùng K neighbors gần nhất	Patterns in data, không quá lớn
Model-based (MICE)	Multiple imputation	Complex missingness patterns
Indicator Feature	Thêm cột is_missing	Khi missingness chứa thông tin

Exam tip: Ba loại missing data: MCAR (Missing Completely At Random) — deletion an toàn; MAR (Missing At Random) — imputation phù hợp; MNAR (Missing Not At Random) — cần indicator feature hoặc domain knowledge.

4. Categorical Encoding

Encoding	Method	When to Use	Issues
One-Hot Encoding	Binary columns mỗi category	Nominal (no order), ít categories	High cardinality → curse of dimensionality
Label Encoding	0, 1, 2, 3...	Ordinal (có thứ tự)	Implies false order for nominal
Target Encoding	Mean of target per category	High cardinality nominal	Data leakage risk nếu không cẩn thận
Embeddings	Dense vector representation	Text, high cardinality	Cần đủ data để learn

5. Normalization & Scaling

Technique	Formula	Output Range	Best For
Min-Max Normalization	(x - min) / (max - min)	[0, 1]	Neural networks, distance-based
Standardization (Z-score)	(x - mean) / std	Mean=0, SD=1	Linear models, SVM, PCA
Robust Scaler	(x - median) / IQR	Centered	Outliers present
Log Transform	log(x)	Compressed	Skewed distributions

6. Xử lý Imbalanced Data

Class imbalance (e.g., fraud detection: 99% normal, 1% fraud) khiến model bias về majority class.

Technique	Method	Direction
Oversampling	Duplicate minority class samples	↑ minority
SMOTE	Synthetic Minority Oversampling Technique — generate synthetic samples	↑ minority
Undersampling	Remove majority class samples	↓ majority
Class Weights	Penalize misclassification of minority more	No data change
Ensemble Methods	BalancedBagging, EasyEnsemble	Algorithm-level

Exam tip: Metric phù hợp cho imbalanced data: F1 Score, AUC-ROC, Precision-Recall — KHÔNG dùng Accuracy (misleading). AWS SageMaker Clarify có thể detect class imbalance.

7. SageMaker Feature Store

SageMaker Feature Store là centralized repository để store, share và reuse ML features.

Feature Store Architecture:

          Feature Groups
         ┌──────────────────────────────┐
         │  user_features               │
         │  ┌──────┬────────┬────────┐  │
         │  │ id   │ age    │ recency│  │
         │  └──────┴────────┴────────┘  │
         └──────────────────────────────┘
               ↓ writes              ↑ reads
    ┌──────────────────┐   ┌──────────────────┐
    │  Offline Store   │   │  Online Store    │
    │  (S3 - training) │   │  (DynamoDB -     │
    │  batch reads     │   │  low-latency     │
    │                  │   │  inference)      │
    └──────────────────┘   └──────────────────┘

8. Cheat Sheet — Feature Engineering

Problem	Solution
High cardinality categorical	Target encoding hoặc embeddings
Missing values (numeric)	Median imputation + indicator feature
Skewed distribution	Log transform hoặc Box-Cox
Outliers	Robust Scaler hoặc clip/winsorize
Imbalanced classes	SMOTE + class weights + AUC metric
Reuse features across teams	SageMaker Feature Store

9. Practice Questions

Q1: A dataset for fraud detection has 98% negative (non-fraud) and 2% positive (fraud) examples. Which metric is MOST appropriate to evaluate the model?

• A) Accuracy
• B) R-squared
• C) AUC-ROC ✓
• D) Mean Absolute Error

Explanation: Accuracy is misleading for imbalanced data (predicting all negative gives 98% accuracy). AUC-ROC measures the model's ability to distinguish classes across all thresholds, making it ideal for imbalanced classification.

Q2: Which technique generates SYNTHETIC samples to address class imbalance?

• A) Random undersampling
• B) SMOTE (Synthetic Minority Oversampling Technique) ✓
• C) Class weighting
• D) Feature scaling

Explanation: SMOTE creates new synthetic samples for the minority class by interpolating between existing minority class examples, rather than just duplicating them.

Q3: A company wants to share engineered features between their training pipeline and real-time inference service. Which SageMaker feature addresses this?

• A) SageMaker Processing Jobs
• B) SageMaker Experiments
• C) SageMaker Feature Store ✓
• D) SageMaker Data Wrangler

Explanation: SageMaker Feature Store provides both an offline store (S3, for batch training) and online store (DynamoDB-backed, for low-latency real-time inference), ensuring feature consistency between training and serving.