كنفرانس بين المللی مهندسی صنايع و سيستم ها

صفحه اصلی / دوازدهمین كنفرانس بين المللی مهندسی صنايع و سيستم ها

A Leakage-Aware Residual Learning and CRITIC-VIKOR Framework for Predictive Seller Prioritization in E-Commerce Marketplaces

نویسندگان :

Alireza Esmaili jobani¹

1- Istanbul Aydin University

کلمات کلیدی :

e-commerce،CRITIC،seller prioritization،VIKOR،residual learning،multi-criteria decision making،LightGBM

چکیده :

Seller evaluation on digital marketplaces is frequently retrospective, static, or dependent on subjective weights, whereas operational interventions require forecasts of future performance. This study proposes a leakage-aware framework that integrates residual gradient boosting with objective multi-criteria decision analysis for next-quarter seller prioritization. The Brazilian Olist e-commerce dataset was transformed into 2,195 consecutive seller-quarter transitions after excluding multi-seller orders whose order-level reviews could not be attributed unambiguously. Fiftytwo historical, lagged, trend, seasonal, and cumulative features forecast six criteria: order volume, revenue, review score, latedelivery rate, handling time, and freight ratio. A tuned residual LightGBM model estimates transformed quarter-to-quarter changes. CRITIC weights and normalization ranges are learned exclusively from historical outcomes, and VIKOR produces the ex-ante ranking. On a strict out-of-time test of 430 sellers in 2018Q3, the proposed method achieved Spearman rho=0.7448 and Kendall tau=0.5578, compared with 0.7175 and 0.5338 for persistence. Mean absolute rank error decreased from 68.63 to 65.87, while top-20 overlap increased from 55% to 65%. The results support future-oriented seller management while showing that gains are not uniform across all top-heavy metrics.