DECOUPLING COMMERCIAL SCALE AND SERVICE QUALITY: A SCALABLE PCA– TOPSIS PIPELINE FOR SELLER EVALUATION IN E-COMMERCE MARKETPLACES

Abstract

Marketplace operators rank sellers on many operational indicators, several of which measure the same underlying property. When such indicators enter a multi-criteria decision-making (MCDM) model directly, the shared property is implicitly weighted more than intended. This article presents a reproducible seller-evaluation pipeline built on the public Olist dataset (96,478 delivered orders, 2,113 sellers with at least three orders). The pipeline derives six indicators, extracts latent factors with principal component analysis (PCA) and varimax rotation, and ranks sellers with TOPSIS. Two factors, Commercial Scale and Service Quality, explain 68.8% of the variance and are stable across bootstrap resamples, activity thresholds and years. Adding two redundant indicators changes an equal-weight TOPSIS ranking substantially (Kendall τ = 0.77), whereas the factor-based ranking with explicit factor weights is more stable (τ = 0.91); weighting factors by explained variance re-introduces the redundancy bias (τ = 0.70). No single composite score predicts both next-year revenue and next-year customer satisfaction, while the two factors do so separately. Sellers with high scale but low service quality generate 47% of marketplace revenue. The pipeline is also implemented in Spark SQL with HiveQL-compatible statements and processes 32 times the original data volume in 26 seconds on a single 4-core node.