Feature Engineering¶

Overview¶

Creating and managing features for ML models.

SageMaker Feature Store¶

Centralized repository for ML features.

Store Types¶

Type	Use Case	Latency
Online Store	Real-time inference	Milliseconds
Offline Store	Batch training	Minutes

Creating a Feature Group¶

from sagemaker.feature_store.feature_group import FeatureGroup

feature_group = FeatureGroup(
    name="customer-features",
    sagemaker_session=sagemaker_session,
    feature_definitions=[
        FeatureDefinition(feature_name="customer_id", feature_type=FeatureTypeEnum.STRING),
        FeatureDefinition(feature_name="age", feature_type=FeatureTypeEnum.INTEGRAL),
        FeatureDefinition(feature_name="total_purchases", feature_type=FeatureTypeEnum.FRACTIONAL),
    ]
)

feature_group.create(
    s3_uri="s3://bucket/feature-store/",
    record_identifier_name="customer_id",
    event_time_feature_name="event_time",
    role_arn=role,
    enable_online_store=True
)

Ingesting Features¶

# Ingest records
feature_group.ingest(data_frame=df, max_workers=3)

Retrieving Features¶

# Online store - real-time
record = feature_group.get_record(record_identifier_value_as_string="customer_123")

# Offline store - batch
query = feature_group.athena_query()
query.run(query_string="SELECT * FROM customer_features", output_location="s3://bucket/query-results/")
df = query.as_dataframe()

Common Feature Engineering Techniques¶

Numerical Features¶

Normalization (min-max scaling)
Standardization (z-score)
Log transformation
Binning/discretization

Categorical Features¶

One-hot encoding
Label encoding
Target encoding
Embedding (for high cardinality)

Text Features¶

TF-IDF
Word embeddings (Word2Vec, BERT)
N-grams

Time Series Features¶

Lag features
Rolling statistics
Seasonal decomposition

Exam Tips¶

!!! warning "Feature Store Key Points" - Online store for low-latency inference - Offline store for training data - Feature groups organize related features - Point-in-time queries prevent data leakage