Review-Grounded LLM Decision Support for E-Commerce: Faithful Recommendation, Customer Segmentation, and Selective Personalization
Stout in Business and Management · 15.01.2026 · 2 (1) · 22-46
DOI: 10.61424/c20v0j71 ↗Kurz gefasst
Fachartikel in der Zeitschrift Stout in Business and Management, online erschienen am 15.01.2026. Verfasst von Fiona Huang.
Abstract (Originalsprache)
E-commerce decision support requires accurate ranking, review-grounded explanations, and a principled way to limit personalization when confidence is weak. This study evaluates those capabilities on 701,528 Amazon Reviews’23 All Beauty records. Four- and five-star reviews defined positive preference, an iterative interaction core supported temporal leave-one-out evaluation, and every method ranked the exact 1,091-item warm catalog. Popularity, ItemKNN, PureSVD, BPR-MF, LightGCN, a training-review TextLSA ranker, and a validation-weighted hybrid were compared for 2,374 test users. The hybrid attained NDCG@10 of 0.09186 and Recall@20 of 0.19966, improving on LightGCN by 15.80% and 17.62%, respectively. Its paired NDCG@10 gain was 0.01253 (95% bootstrap interval [0.00595, 0.01958]; paired randomization p = .00030). Review fusion also reduced expected calibration error from 0.29112 to 0.25319 and area under the risk–coverage curve from 0.67798 to 0.63844, although absolute confidence remained overestimated. Six review-behavior segments showed substantial ranking heterogeneity, with Recall@20 ranging from 0.02000 to 0.25000. A local Qwen2.5-0.5B assessment on four balanced evidence packets achieved 0.75 selective-action accuracy but only 0.25 citation accuracy, exact-quote fidelity, and joint constraint compliance. The findings support review-language fusion for ranking and confidence ordering while showing that customer-facing explanations require deterministic validation and abstention safeguards.
Zitieren
APA
Huang, F. (2026). Review-Grounded LLM Decision Support for E-Commerce: Faithful Recommendation, Customer Segmentation, and Selective Personalization. Stout in Business and Management, 2(1), 22-46. https://doi.org/10.61424/c20v0j71
Vancouver
Huang F. Review-Grounded LLM Decision Support for E-Commerce: Faithful Recommendation, Customer Segmentation, and Selective Personalization. Stout in Business and Management. 2026;2(1):22-46. doi:10.61424/c20v0j71
BibTeX
@article{Huang2026,
author = {Huang, Fiona},
title = {Review-Grounded LLM Decision Support for E-Commerce: Faithful Recommendation, Customer Segmentation, and Selective Personalization},
journal = {Stout in Business and Management},
year = {2026},
volume = {2},
number = {1},
pages = {22--46},
doi = {10.61424/c20v0j71}
}
EUVIMED