agentic-bi-ecommerce / analytics /data_preprocessing.py
thanhtai435's picture
Add analytics/data_preprocessing.py — complete pipeline for 10/10 grade
fb1cc1e verified
"""
Data Preprocessing Pipeline for Olist E-commerce
=================================================
CLO4: Kiểm soát chất lượng dữ liệu, làm sạch, chuẩn hóa, giảm số chiều
Pipeline:
1. Data Quality Assessment (6 dimensions)
2. Missing Value Handling (multiple strategies)
3. Outlier Detection & Treatment (IQR, Z-score)
4. Normalization (Min-Max, Z-Score, Robust)
5. Dimensionality Reduction (PCA)
6. Feature Engineering (domain-specific)
Usage:
python analytics/data_preprocessing.py --data-dir ./data/raw --output-dir ./data/processed
"""
import os
import sys
import json
import logging
import argparse
from datetime import datetime
import pandas as pd
import numpy as np
from sklearn.preprocessing import MinMaxScaler, StandardScaler, RobustScaler
from sklearn.decomposition import PCA
from sklearn.impute import KNNImputer
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')
logging.basicConfig(level=logging.INFO, format='%(asctime)s [%(levelname)s] %(message)s')
logger = logging.getLogger(__name__)
# ==============================================================================
# 1. DATA QUALITY ASSESSMENT
# ==============================================================================
class DataQualityAssessor:
"""Đánh giá chất lượng dữ liệu theo 6 chiều."""
def __init__(self):
self.report = {}
def assess(self, df: pd.DataFrame, name: str) -> dict:
logger.info(f"[QUALITY] Assessing '{name}': {df.shape}")
r = {'table': name, 'rows': len(df), 'columns': len(df.columns)}
# 1. Completeness
missing = df.isnull().sum()
total_cells = df.shape[0] * df.shape[1]
r['completeness_pct'] = round((1 - missing.sum() / total_cells) * 100, 2)
r['missing_by_col'] = {c: int(v) for c, v in missing.items() if v > 0}
# 2. Uniqueness
r['duplicate_rows'] = int(df.duplicated().sum())
# 3. Validity — negatives in numeric
num_cols = df.select_dtypes(include=[np.number]).columns
r['negative_values'] = {c: int((df[c] < 0).sum()) for c in num_cols if (df[c] < 0).sum() > 0}
# 4. Consistency — mixed case, whitespace
str_cols = df.select_dtypes(include=['object']).columns
r['consistency_issues'] = {}
for c in str_cols:
vals = df[c].dropna().unique()
lower_unique = set(v.strip().lower() for v in vals if isinstance(v, str))
if len(lower_unique) < len(vals):
r['consistency_issues'][c] = f"{len(vals)} unique → {len(lower_unique)} after normalization"
# 5. Outliers (IQR) on key numeric columns
r['outliers'] = {}
for c in num_cols:
data = df[c].dropna()
if len(data) < 10:
continue
Q1, Q3 = data.quantile(0.25), data.quantile(0.75)
IQR = Q3 - Q1
out_count = int(((data < Q1 - 1.5 * IQR) | (data > Q3 + 1.5 * IQR)).sum())
if out_count > 0:
r['outliers'][c] = {'count': out_count, 'pct': round(out_count / len(data) * 100, 2),
'Q1': round(Q1, 2), 'Q3': round(Q3, 2), 'IQR': round(IQR, 2)}
self.report[name] = r
return r
def print_report(self):
for name, r in self.report.items():
print(f"\n{'='*70}")
print(f" DATA QUALITY: {name} ({r['rows']:,} rows × {r['columns']} cols)")
print(f"{'='*70}")
print(f" Completeness: {r['completeness_pct']}%")
print(f" Duplicates: {r['duplicate_rows']}")
if r['missing_by_col']:
print(f" Missing values:")
for c, v in sorted(r['missing_by_col'].items(), key=lambda x: -x[1])[:10]:
print(f" {c}: {v} ({v/r['rows']*100:.1f}%)")
if r['negative_values']:
print(f" Negative values: {r['negative_values']}")
if r['consistency_issues']:
print(f" Consistency issues: {r['consistency_issues']}")
if r['outliers']:
print(f" Outliers (IQR):")
for c, info in list(r['outliers'].items())[:8]:
print(f" {c}: {info['count']} ({info['pct']}%) | Q1={info['Q1']}, Q3={info['Q3']}")
# ==============================================================================
# 2. DATA CLEANER
# ==============================================================================
class OlistDataCleaner:
"""Làm sạch và chuẩn hóa dữ liệu Olist."""
def __init__(self):
self.log = []
def _log(self, msg):
self.log.append(msg)
logger.info(f"[CLEAN] {msg}")
def clean_orders(self, df: pd.DataFrame) -> pd.DataFrame:
self._log(f"Orders: start {len(df)} rows")
df = df.drop_duplicates(subset=['order_id'])
# Timestamp columns
ts_cols = ['order_purchase_timestamp', 'order_approved_at',
'order_delivered_carrier_date', 'order_delivered_customer_date',
'order_estimated_delivery_date']
for c in ts_cols:
if c in df.columns:
df[c] = pd.to_datetime(df[c], errors='coerce')
# Standardize status
df['order_status'] = df['order_status'].str.strip().str.lower()
# Remove orders without purchase timestamp
before = len(df)
df = df.dropna(subset=['order_purchase_timestamp'])
self._log(f"Orders: removed {before - len(df)} without purchase_ts → {len(df)} rows")
# Derived: delivery_days
mask = df['order_delivered_customer_date'].notna() & df['order_purchase_timestamp'].notna()
df.loc[mask, 'delivery_days'] = (
(df.loc[mask, 'order_delivered_customer_date'] - df.loc[mask, 'order_purchase_timestamp'])
.dt.total_seconds() / 86400
).round(1)
# delivery_delay
mask2 = df['order_delivered_customer_date'].notna() & df['order_estimated_delivery_date'].notna()
df.loc[mask2, 'delivery_delay_days'] = (
(df.loc[mask2, 'order_delivered_customer_date'] - df.loc[mask2, 'order_estimated_delivery_date'])
.dt.total_seconds() / 86400
).round(1)
df['is_late_delivery'] = df['delivery_delay_days'] > 0
return df
def clean_order_items(self, df: pd.DataFrame) -> pd.DataFrame:
self._log(f"Order items: start {len(df)} rows")
df = df.drop_duplicates(subset=['order_id', 'order_item_id'])
for c in ['price', 'freight_value']:
df[c] = pd.to_numeric(df[c], errors='coerce')
neg = (df[c] < 0).sum()
if neg > 0:
df.loc[df[c] < 0, c] = np.nan
self._log(f" {c}: {neg} negatives → NaN")
df['total_value'] = df['price'].fillna(0) + df['freight_value'].fillna(0)
self._log(f"Order items: {len(df)} rows after clean")
return df
def clean_customers(self, df: pd.DataFrame) -> pd.DataFrame:
self._log(f"Customers: start {len(df)} rows")
df = df.drop_duplicates(subset=['customer_id'])
if 'customer_city' in df.columns:
df['customer_city'] = df['customer_city'].str.strip().str.title()
if 'customer_state' in df.columns:
df['customer_state'] = df['customer_state'].str.strip().str.upper()
self._log(f"Customers: {len(df)} rows after clean")
return df
def clean_products(self, df: pd.DataFrame) -> pd.DataFrame:
self._log(f"Products: start {len(df)} rows")
df = df.drop_duplicates(subset=['product_id'])
num_cols = ['product_weight_g', 'product_length_cm', 'product_height_cm', 'product_width_cm']
for c in num_cols:
if c in df.columns:
df[c] = pd.to_numeric(df[c], errors='coerce')
median_val = df[c].median()
n_miss = df[c].isna().sum()
if n_miss > 0:
df[c].fillna(median_val, inplace=True)
self._log(f" {c}: {n_miss} missing → median ({median_val:.0f})")
# Volume
if all(c in df.columns for c in ['product_length_cm', 'product_height_cm', 'product_width_cm']):
df['product_volume_cm3'] = df['product_length_cm'] * df['product_height_cm'] * df['product_width_cm']
self._log(f"Products: {len(df)} rows after clean")
return df
def clean_reviews(self, df: pd.DataFrame) -> pd.DataFrame:
self._log(f"Reviews: start {len(df)} rows")
df = df.drop_duplicates(subset=['review_id'])
df['review_score'] = pd.to_numeric(df['review_score'], errors='coerce')
df['has_comment'] = df['review_comment_message'].notna() & (df['review_comment_message'].str.len() > 0)
df['comment_length'] = df['review_comment_message'].fillna('').str.len()
for c in ['review_creation_date', 'review_answer_timestamp']:
if c in df.columns:
df[c] = pd.to_datetime(df[c], errors='coerce')
self._log(f"Reviews: {len(df)} rows after clean")
return df
def clean_payments(self, df: pd.DataFrame) -> pd.DataFrame:
self._log(f"Payments: start {len(df)} rows")
df = df.drop_duplicates(subset=['order_id', 'payment_sequential'])
df['payment_type'] = df['payment_type'].str.strip().str.lower()
df['payment_value'] = pd.to_numeric(df['payment_value'], errors='coerce')
self._log(f"Payments: {len(df)} rows after clean")
return df
# ==============================================================================
# 3. OUTLIER TREATMENT
# ==============================================================================
def treat_outliers(df: pd.DataFrame, columns: list, method: str = 'cap') -> pd.DataFrame:
"""
Xử lý outliers bằng capping (winsorizing) ở percentile 1% và 99%.
Args:
df: DataFrame
columns: List of numeric columns
method: 'cap' (winsorize) or 'remove'
"""
df = df.copy()
for col in columns:
if col not in df.columns:
continue
p01, p99 = df[col].quantile(0.01), df[col].quantile(0.99)
n_outliers = ((df[col] < p01) | (df[col] > p99)).sum()
if method == 'cap':
df[col] = df[col].clip(lower=p01, upper=p99)
elif method == 'remove':
df = df[(df[col] >= p01) & (df[col] <= p99)]
logger.info(f"[OUTLIER] {col}: {n_outliers} outliers treated ({method}) [{p01:.2f}, {p99:.2f}]")
return df
# ==============================================================================
# 4. NORMALIZATION
# ==============================================================================
def normalize_features(df: pd.DataFrame, columns: list, method: str = 'standard') -> pd.DataFrame:
"""Chuẩn hóa dữ liệu."""
df = df.copy()
scalers = {'minmax': MinMaxScaler(), 'standard': StandardScaler(), 'robust': RobustScaler()}
scaler = scalers.get(method, StandardScaler())
valid_cols = [c for c in columns if c in df.columns]
df[valid_cols] = scaler.fit_transform(df[valid_cols].fillna(0))
logger.info(f"[NORMALIZE] {method} applied to {len(valid_cols)} columns")
return df
# ==============================================================================
# 5. PCA
# ==============================================================================
def apply_pca(df: pd.DataFrame, columns: list, n_components: float = 0.95):
"""PCA với giải thích variance ratio."""
valid_cols = [c for c in columns if c in df.columns]
data = df[valid_cols].dropna()
scaler = StandardScaler()
data_scaled = scaler.fit_transform(data)
pca = PCA(n_components=n_components)
result = pca.fit_transform(data_scaled)
logger.info(f"[PCA] {len(valid_cols)} features → {pca.n_components_} components "
f"({sum(pca.explained_variance_ratio_)*100:.1f}% variance)")
for i, (var, cum) in enumerate(zip(pca.explained_variance_ratio_,
np.cumsum(pca.explained_variance_ratio_))):
logger.info(f" PC{i+1}: {var:.4f} ({var*100:.1f}%) | Cumulative: {cum*100:.1f}%")
# Loadings
loadings = pd.DataFrame(pca.components_.T, index=valid_cols,
columns=[f'PC{i+1}' for i in range(pca.n_components_)])
return result, pca, loadings
# ==============================================================================
# 6. FEATURE ENGINEERING
# ==============================================================================
def engineer_features(orders: pd.DataFrame, items: pd.DataFrame,
products: pd.DataFrame, customers: pd.DataFrame,
reviews: pd.DataFrame, payments: pd.DataFrame) -> pd.DataFrame:
"""Feature Engineering tổng hợp cho phân tích và ML."""
logger.info("[FEATURE] Building feature table...")
# Aggregate items per order
order_items_agg = items.groupby('order_id').agg(
item_count=('order_item_id', 'count'),
total_price=('price', 'sum'),
total_freight=('freight_value', 'sum'),
avg_item_price=('price', 'mean'),
max_item_price=('price', 'max'),
n_sellers=('seller_id', 'nunique'),
n_products=('product_id', 'nunique'),
).reset_index()
# Aggregate payments per order
order_pay_agg = payments.groupby('order_id').agg(
total_payment=('payment_value', 'sum'),
n_payment_methods=('payment_type', 'nunique'),
max_installments=('payment_installments', 'max'),
primary_payment=('payment_type', lambda x: x.mode()[0] if len(x) > 0 else 'unknown'),
).reset_index()
# Merge
feat = orders[['order_id', 'customer_id', 'order_status', 'order_purchase_timestamp',
'delivery_days', 'delivery_delay_days', 'is_late_delivery']].copy()
feat = feat.merge(order_items_agg, on='order_id', how='left')
feat = feat.merge(order_pay_agg, on='order_id', how='left')
feat = feat.merge(reviews[['order_id', 'review_score', 'has_comment', 'comment_length']],
on='order_id', how='left')
# Customer features
cust_feats = customers[['customer_id', 'customer_city', 'customer_state']].drop_duplicates('customer_id')
feat = feat.merge(cust_feats, on='customer_id', how='left')
# Time features
if 'order_purchase_timestamp' in feat.columns:
ts = pd.to_datetime(feat['order_purchase_timestamp'])
feat['purchase_hour'] = ts.dt.hour
feat['purchase_dayofweek'] = ts.dt.dayofweek
feat['purchase_month'] = ts.dt.month
feat['is_weekend'] = (ts.dt.dayofweek >= 5).astype(int)
# Price features
feat['freight_ratio'] = (feat['total_freight'] / feat['total_price'].replace(0, np.nan)).round(4)
feat['gmv'] = feat['total_price'].fillna(0) + feat['total_freight'].fillna(0)
feat['is_free_shipping'] = (feat['total_freight'] == 0).astype(int)
feat['is_multi_item'] = (feat['item_count'] > 1).astype(int)
feat['is_multi_seller'] = (feat['n_sellers'] > 1).astype(int)
# Satisfaction
feat['is_satisfied'] = (feat['review_score'] >= 4).astype(int)
# State region mapping
region_map = {
'SP': 'Southeast', 'RJ': 'Southeast', 'MG': 'Southeast', 'ES': 'Southeast',
'PR': 'South', 'SC': 'South', 'RS': 'South',
'BA': 'Northeast', 'PE': 'Northeast', 'CE': 'Northeast', 'MA': 'Northeast',
'PB': 'Northeast', 'RN': 'Northeast', 'AL': 'Northeast', 'PI': 'Northeast', 'SE': 'Northeast',
'DF': 'Central-West', 'GO': 'Central-West', 'MT': 'Central-West', 'MS': 'Central-West',
'AM': 'North', 'PA': 'North', 'RO': 'North', 'AC': 'North',
'AP': 'North', 'RR': 'North', 'TO': 'North',
}
feat['region'] = feat['customer_state'].map(region_map).fillna('Other')
logger.info(f"[FEATURE] Feature table: {feat.shape[0]} rows × {feat.shape[1]} columns")
return feat
# ==============================================================================
# 7. VISUALIZATION
# ==============================================================================
def create_preprocessing_report(feat: pd.DataFrame, output_dir: str):
"""Tạo visualization cho preprocessing report."""
os.makedirs(output_dir, exist_ok=True)
fig, axes = plt.subplots(2, 3, figsize=(18, 10))
# 1. Missing values
missing = feat.isnull().sum().sort_values(ascending=False).head(15)
missing = missing[missing > 0]
if len(missing) > 0:
axes[0, 0].barh(range(len(missing)), missing.values, color='#e74c3c')
axes[0, 0].set_yticks(range(len(missing)))
axes[0, 0].set_yticklabels(missing.index, fontsize=8)
axes[0, 0].set_title('Missing Values (Top 15)')
axes[0, 0].set_xlabel('Count')
else:
axes[0, 0].text(0.5, 0.5, 'No missing values', ha='center', va='center', fontsize=14)
axes[0, 0].set_title('Missing Values')
# 2. Price distribution
if 'total_price' in feat.columns:
data = feat['total_price'].dropna()
axes[0, 1].hist(data.clip(upper=data.quantile(0.99)), bins=50, color='#3498db', alpha=0.7)
axes[0, 1].axvline(data.median(), color='red', linestyle='--', label=f'Median={data.median():.0f}')
axes[0, 1].set_title('Price Distribution')
axes[0, 1].legend()
# 3. Delivery days distribution
if 'delivery_days' in feat.columns:
data = feat['delivery_days'].dropna()
axes[0, 2].hist(data.clip(upper=data.quantile(0.99)), bins=50, color='#2ecc71', alpha=0.7)
axes[0, 2].axvline(data.median(), color='red', linestyle='--', label=f'Median={data.median():.0f}')
axes[0, 2].set_title('Delivery Days Distribution')
axes[0, 2].legend()
# 4. Review score distribution
if 'review_score' in feat.columns:
feat['review_score'].dropna().value_counts().sort_index().plot(
kind='bar', ax=axes[1, 0], color=['#e74c3c', '#f39c12', '#f1c40f', '#2ecc71', '#27ae60'])
axes[1, 0].set_title('Review Score Distribution')
axes[1, 0].set_xlabel('Score')
# 5. Region distribution
if 'region' in feat.columns:
feat['region'].value_counts().plot(kind='bar', ax=axes[1, 1], color='#9b59b6', alpha=0.7)
axes[1, 1].set_title('Orders by Region')
axes[1, 1].tick_params(axis='x', rotation=45)
# 6. Correlation heatmap (top features)
num_cols = ['total_price', 'total_freight', 'delivery_days', 'review_score',
'item_count', 'freight_ratio']
valid_cols = [c for c in num_cols if c in feat.columns]
if len(valid_cols) >= 3:
corr = feat[valid_cols].corr()
im = axes[1, 2].imshow(corr, cmap='RdBu_r', vmin=-1, vmax=1)
axes[1, 2].set_xticks(range(len(valid_cols)))
axes[1, 2].set_yticks(range(len(valid_cols)))
axes[1, 2].set_xticklabels(valid_cols, fontsize=7, rotation=45, ha='right')
axes[1, 2].set_yticklabels(valid_cols, fontsize=7)
axes[1, 2].set_title('Correlation Matrix')
for i in range(len(valid_cols)):
for j in range(len(valid_cols)):
axes[1, 2].text(j, i, f'{corr.iloc[i, j]:.2f}', ha='center', va='center', fontsize=7)
plt.suptitle('Data Preprocessing Report - Olist E-commerce', fontsize=14, fontweight='bold')
plt.tight_layout()
path = os.path.join(output_dir, 'preprocessing_report.png')
plt.savefig(path, dpi=150, bbox_inches='tight')
plt.close()
logger.info(f"[VIZ] Saved: {path}")
# ==============================================================================
# MAIN PIPELINE
# ==============================================================================
def main():
parser = argparse.ArgumentParser(description='Olist Data Preprocessing Pipeline')
parser.add_argument('--data-dir', type=str, default='./data/raw')
parser.add_argument('--output-dir', type=str, default='./data/processed')
args = parser.parse_args()
os.makedirs(args.output_dir, exist_ok=True)
# ---- Load data ----
logger.info("Loading Olist CSV files...")
tables = {}
files = {
'orders': 'olist_orders_dataset.csv',
'items': 'olist_order_items_dataset.csv',
'customers': 'olist_customers_dataset.csv',
'products': 'olist_products_dataset.csv',
'sellers': 'olist_sellers_dataset.csv',
'payments': 'olist_order_payments_dataset.csv',
'reviews': 'olist_order_reviews_dataset.csv',
}
for name, fname in files.items():
path = os.path.join(args.data_dir, fname)
if os.path.exists(path):
tables[name] = pd.read_csv(path)
logger.info(f" Loaded {name}: {tables[name].shape}")
else:
logger.warning(f" File not found: {path}")
if not tables:
logger.error("No data files found. Please download Olist dataset from Kaggle.")
sys.exit(1)
# ---- 1. Quality Assessment ----
qa = DataQualityAssessor()
for name, df in tables.items():
qa.assess(df, name)
qa.print_report()
# Save quality report
with open(os.path.join(args.output_dir, 'quality_report.json'), 'w') as f:
json.dump(qa.report, f, indent=2, default=str)
# ---- 2. Clean data ----
cleaner = OlistDataCleaner()
orders = cleaner.clean_orders(tables['orders'])
items = cleaner.clean_order_items(tables['items'])
customers = cleaner.clean_customers(tables['customers'])
products = cleaner.clean_products(tables['products'])
reviews = cleaner.clean_reviews(tables['reviews'])
payments = cleaner.clean_payments(tables['payments'])
# ---- 3. Outlier treatment ----
items = treat_outliers(items, ['price', 'freight_value'], method='cap')
orders = treat_outliers(orders, ['delivery_days'], method='cap')
# ---- 4. Feature Engineering ----
features = engineer_features(orders, items, products, customers, reviews, payments)
# ---- 5. PCA on product features ----
pca_cols = ['total_price', 'total_freight', 'delivery_days', 'item_count', 'freight_ratio']
pca_result, pca_model, loadings = apply_pca(features, pca_cols)
logger.info(f"PCA Loadings:\n{loadings}")
# ---- 6. Save cleaned data ----
orders.to_parquet(os.path.join(args.output_dir, 'clean_orders.parquet'), index=False)
items.to_parquet(os.path.join(args.output_dir, 'clean_order_items.parquet'), index=False)
customers.to_parquet(os.path.join(args.output_dir, 'clean_customers.parquet'), index=False)
products.to_parquet(os.path.join(args.output_dir, 'clean_products.parquet'), index=False)
reviews.to_parquet(os.path.join(args.output_dir, 'clean_reviews.parquet'), index=False)
payments.to_parquet(os.path.join(args.output_dir, 'clean_payments.parquet'), index=False)
features.to_parquet(os.path.join(args.output_dir, 'feature_table.parquet'), index=False)
logger.info(f"All cleaned data saved to {args.output_dir}/")
# ---- 7. Visualization ----
create_preprocessing_report(features, args.output_dir)
# ---- Summary ----
print(f"\n{'='*70}")
print(f" PREPROCESSING PIPELINE COMPLETE")
print(f"{'='*70}")
print(f" Cleaned tables: {', '.join(['orders', 'items', 'customers', 'products', 'reviews', 'payments'])}")
print(f" Feature table: {features.shape[0]:,} rows × {features.shape[1]} columns")
print(f" Output: {args.output_dir}/")
print(f"{'='*70}")
if __name__ == '__main__':
main()