Retail/Fashion EDA Framework

Comprehensive EDA approach for retail/fashion/tabular data, validated through Walmart, H&M, and 7+ fashion-lifecycle-pricing competitions.

Problem

Most ML projects skip EDA or do it superficially (just df.describe() + a histogram). This leads to:

• Missing data issues discovered too late
• Train/test distribution shift not caught
• High-cardinality features used as raw categoricals (overfit)
• Text features not extracted from product descriptions
• Time-based leakage in train/test split
• Hierarchical structure ignored (e.g., 5-level category hierarchy in H&M)

Real case (validated 2026-06-02 on H&M articles.csv):

• 105,542 rows × 25 cols
• Only 416 missing values (0.39% in detail_desc)
• 5-level hierarchical structure: index_group → section → department → product_type → product_group
• 45,875 unique prod_name (high cardinality)
• 43,404 unique detail_desc (text, 142 chars avg)

Without proper EDA, none of these are caught before feature engineering.

The 5-Stage EDA Pipeline

Stage 1: Data Quality Audit (FIRST)

Tools: ydata-profiling (13.5k★), missingno (4.2k★)

# Quick data quality report
import ydata_profiling
profile = ydata_profiling.ProfileReport(df, title="Data Quality Report")
profile.to_file("eda/data_quality.html")

# Missing data visualization
import missingno as msno
msno.matrix(df)        # Bar chart of missing per column
msno.heatmap(df)       # Correlation of missingness between columns
msno.dendrogram(df)    # Hierarchical clustering of missingness

Look for:

• Missing value patterns (random vs systematic)
• High-cardinality categoricals (will overfit tree models)
• Skewed numerical features (need log transform)
• Constant/quasi-constant features (drop immediately)
• Duplicate rows
• Outliers (use IQR or z-score, not just visual)

Stage 2: Statistical Profiling (Train vs Test)

Tools: sweetviz (3.1k★)

# Compare train vs test
import sweetviz as sv
report = sv.compare([train_df, "Train"], [test_df, "Test"], target_feat="target")
report.show_html("eda/train_vs_test.html")

Look for:

• Train/test distribution shift
• Per-segment behavior differences
• Data leakage (test features in train)
• Time-based trends (if applicable)

Stage 3: Domain-Specific EDA

For retail/fashion, focus on:

Stage 4: Time-Series EDA (if applicable)

Tools: tslumen (72★, HSBC-maintained), statsmodels

# For sales forecasting competitions
import tslumen
tslumen.from_ts(df.set_index('date')['sales']).plot()
# Decomposition, ACF/PACF, stationarity tests

Look for:

• Trend / seasonality / residual decomposition
• Stationarity (ADF test)
• Autocorrelation structure (ACF/PACF)
• Holiday/special-event effects
• Data frequency consistency (missing weeks, etc.)

Stage 5: EDA Summary

Auto-generated eda/EDA_SUMMARY.md with:

• Files generated checklist
• Manual review items
• Next-step recommendations

Top EDA Library Recommendations (2026-06-02)

Common EDA Mistakes (Anti-Patterns)

1. ❌ Skipping EDA to save time → always leads to feature engineering errors
2. ❌ Only looking at summary statistics → miss distribution shape
3. ❌ Ignoring train/test distribution shift → catastrophic in production
4. ❌ Not checking missingness correlation → indicates systematic missingness
5. ❌ Assuming all categorical = independent → categories may be related
6. ❌ Forgetting to check time-based leakage → future data in training
7. ❌ Not visualizing outliers → they may be valid (e.g., luxury items)
8. ❌ Treating all features equally → ID-like features need special handling

Real Case: H&M Personalized Fashion Recommendations

Data: articles.csv (105,542 rows × 25 cols) Date: 2026-06-02

EDA Findings

• Data Quality: Only 416 missing (0.39% in detail_desc) — extremely clean
• Hierarchical structure: 5 levels (index_group → section → department → product_type → product_group)
  • index_group_name (5 unique): Ladieswear, Menswear, Baby/Children, Sport, Divided
  • department_name (250 unique)
  • product_type_name (131 unique)
• Top product types: Trousers (11K), Dress (10K), Sweater (9K)
• Top colors: Black (22K), Blue (18K), White (12K) — top 5 = 67% of articles
• Text features: prod_name 45,875 unique, detail_desc 43,404 unique (mean 142 chars)
• Data quality issues:
  • Some numerical cols have -1 values (treat as NaN)
  • article_id is large numeric (use as index only, not as feature)

Recommendations for Feature Engineering

1. Hierarchy-based target encoding — 5 levels perfect for tree models
2. Text features from detail_desc — 142 chars avg, ideal for embeddings
3. Color groupings — reduce 50 colors to 10-15 effective groups
4. Product variant detection — product_code shared across variants
5. Time-aware features — article age, recent sales velocity, seasonality

How to Apply

Option 1: Use the existing pipeline (recommended)

The ml-agent-code-template includes a pre-built EDA pipeline:

# Install dependencies
pip install ydata-profiling missingno sweetviz

# Run the pipeline
bash ml-agent-code-template/.claude/hooks/eda_pipeline.sh <data.csv>

# With optional flags
bash ml-agent-code-template/.claude/hooks/eda_pipeline.sh <data.csv> \
  --segment=customer_type --time-series=date

Option 2: Manual invocation

import ydata_profiling
import missingno as msno
import sweetviz as sv

# Stage 1
profile = ydata_profiling.ProfileReport(df, title="Data Quality Report", minimal=True)
profile.to_file("eda/data_quality.html")

# Stage 2
report = sv.compare([train, "Train"], [test, "Test"], target_feat="target")
report.show_html("eda/train_vs_test.html")

Related Skills

• adversarial-validation-kaggle — for distribution shift detection
• kaggle-data-format-first — for understanding schema
• kaggle-top-performer-replication — for 1st place feature analysis

Empirical Evidence

MLE-Bench experiments:

• Spaceship Titanic Gold: 0.8506 — EDA revealed Cabin split into deck/side/num was key
• Jigsaw Toxic Gold: 0.98829 — EDA on per-label distribution identified imbalance
• TPS May Silver: 0.99754 — EDA on f_27 string structure drove feature breakthrough

fashion-lifecycle-pricing:

• H&M R27 (best, Priv 0.02314): 27 experiments preceded by proper EDA
• Walmart R08 (best, LB=2720): EDA on MarkDown missing patterns + seasonality