Purpose

Retention Rate: 97.1% - A high proportion of the original dataset was preserved, indicating minimal data loss during cleaning
Rows Removed: 1,452 observations (2.9%) were excluded, suggesting moderate filtering for data quality issues
Final Dataset Size: 48,548 rows provided sufficient volume for training (38,839) and testing (9,709) with balanced class distribution (49.8% positive cases)

This section identifies which features contribute most to the model's decision-making through gain-based importance. Gain measures the information value each feature provides when splitting data in the boosting trees. Understanding feature importance reveals which transaction attributes are most predictive of high-value versus low-value classifications.

Key Findings

qty_capped dominance: 60.3% of total gain—overwhelmingly the strongest predictor of transaction value classification
log_unit_price secondary importance: 38% gain, the second-most influential feature with comparable coverage (0.37) and frequency (0.37)
Geographic features negligible: Country-based features (Cyprus, Netherlands, France, Germany, Spain, Portugal) contribute zero gain, indicating geographic location does not meaningfully distinguish transaction value
Temporal features minimal: hour_of_day and day_of_week show minimal gain (0.01 and 0), suggesting timing is not a strong classifier

Interpretation

The model relies almost exclusively on quantity and unit price to classify transactions. The extreme concentration in qty_capped (60.3%) indicates this single feature carries the majority of predictive power. The near-zero contributions from geographic and temporal features suggest the transaction value classification is fundamentally driven by product-level characteristics rather than when or where transactions occur. This aligns with the model's perfect performance (AUC