피처 엔지니어링

# 피처 엔지니어링 및 Feature Store 활용 예시
import pandas as pd
import numpy as np
from datetime import datetime, timedelta
from feast import FeatureStore, Entity, FeatureView, Field, FileSource
from feast.types import Float32, Int64, String

# 1. 기본 피처 엔지니어링 함수
def engineer_customer_features(df: pd.DataFrame) -> pd.DataFrame:
    """고객 데이터에서 ML 피처 생성"""

    # 날짜/시간 분해
    df['signup_dayofweek'] = df['signup_date'].dt.dayofweek
    df['signup_month'] = df['signup_date'].dt.month
    df['signup_is_weekend'] = df['signup_dayofweek'].isin([5, 6]).astype(int)
    df['days_since_signup'] = (datetime.now() - df['signup_date']).dt.days

    # 수치 변환 (정규분포화)
    df['log_total_spend'] = np.log1p(df['total_spend'])
    df['sqrt_order_count'] = np.sqrt(df['order_count'])

    # 비율 피처 (Ratio Features)
    df['avg_order_value'] = df['total_spend'] / (df['order_count'] + 1)
    df['order_frequency'] = df['order_count'] / (df['days_since_signup'] + 1)

    # 구간화 (Binning)
    df['spend_tier'] = pd.cut(
        df['total_spend'],
        bins=[0, 100, 500, 2000, np.inf],
        labels=['low', 'medium', 'high', 'vip']
    )

    # 그룹 대비 비율 (Relative Features)
    df['spend_vs_segment_avg'] = df['total_spend'] / df.groupby('segment')['total_spend'].transform('mean')

    return df

# 2. 시계열 집계 피처 (Rolling/Lag Features)
def create_time_features(df: pd.DataFrame, user_col: str, date_col: str) -> pd.DataFrame:
    """시계열 기반 피처 생성"""

    df = df.sort_values([user_col, date_col])

    # 이동 평균 (Rolling Mean)
    df['revenue_ma_7d'] = df.groupby(user_col)['revenue'].transform(
        lambda x: x.rolling(7, min_periods=1).mean()
    )
    df['revenue_ma_30d'] = df.groupby(user_col)['revenue'].transform(
        lambda x: x.rolling(30, min_periods=1).mean()
    )

    # Lag 피처 (이전 값)
    df['revenue_lag_1'] = df.groupby(user_col)['revenue'].shift(1)
    df['revenue_lag_7'] = df.groupby(user_col)['revenue'].shift(7)

    # 변화율
    df['revenue_pct_change'] = df.groupby(user_col)['revenue'].pct_change()

    # 누적 피처
    df['cumulative_orders'] = df.groupby(user_col).cumcount() + 1
    df['cumulative_revenue'] = df.groupby(user_col)['revenue'].cumsum()

    return df

# 3. Feast Feature Store 정의
# feature_repo/features.py

# 엔티티 정의 (Primary Key)
customer = Entity(
    name="customer_id",
    description="고객 고유 식별자"
)

# 오프라인 데이터 소스
customer_source = FileSource(
    path="s3://feature-store/customer_features.parquet",
    timestamp_field="event_timestamp",
)

# 피처 뷰 정의
customer_features = FeatureView(
    name="customer_features",
    entities=[customer],
    ttl=timedelta(days=1),
    schema=[
        Field(name="total_spend", dtype=Float32),
        Field(name="order_count", dtype=Int64),
        Field(name="avg_order_value", dtype=Float32),
        Field(name="days_since_signup", dtype=Int64),
        Field(name="spend_tier", dtype=String),
        Field(name="revenue_ma_7d", dtype=Float32),
        Field(name="revenue_ma_30d", dtype=Float32),
        Field(name="order_frequency", dtype=Float32),
    ],
    source=customer_source,
    online=True,  # 온라인 서빙 활성화
)

# 4. Feature Store 사용
fs = FeatureStore(repo_path="feature_repo/")

# 오프라인: 학습 데이터 생성 (Point-in-Time Join)
entity_df = pd.DataFrame({
    "customer_id": ["C001", "C002", "C003"],
    "event_timestamp": [datetime.now()] * 3,
})

training_df = fs.get_historical_features(
    entity_df=entity_df,
    features=[
        "customer_features:total_spend",
        "customer_features:avg_order_value",
        "customer_features:revenue_ma_7d",
        "customer_features:order_frequency"
    ]
).to_df()

print(f"학습 데이터: {training_df.shape}")

# 온라인: 실시간 추론용 피처 조회
online_features = fs.get_online_features(
    features=[
        "customer_features:total_spend",
        "customer_features:avg_order_value",
    ],
    entity_rows=[{"customer_id": "C001"}]
).to_dict()

print(f"실시간 피처: {online_features}")

# 5. 자동 피처 엔지니어링 (Featuretools)
import featuretools as ft

# EntitySet 생성
es = ft.EntitySet(id="customer_data")
es.add_dataframe(dataframe_name="customers", dataframe=customers_df, index="customer_id")
es.add_dataframe(dataframe_name="orders", dataframe=orders_df, index="order_id")
es.add_relationship("customers", "customer_id", "orders", "customer_id")

# 자동 피처 생성 (Deep Feature Synthesis)
feature_matrix, feature_defs = ft.dfs(
    entityset=es,
    target_dataframe_name="customers",
    agg_primitives=["mean", "sum", "count", "max", "min", "std", "mode"],
    trans_primitives=["month", "weekday", "is_weekend", "year"],
    max_depth=2
)

print(f"자동 생성된 피처 수: {len(feature_defs)}")