Haris Saefuloh
Universitas Logistik dan Bisnis Internasional

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

Text-Based Recommender Systems for User Cold-Start Problems in E-Commerce Product Recommendation: A Systematic Literature Review Muhammad Yusril Helmi Setyawan; Haris Saefuloh
SISTEMASI Vol 15, No 9 (2026): Sistemasi: Jurnal Sistem Informasi
Publisher : Universitas Islam Indragiri

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.32520/stmsi.v15i9.6817

Abstract

The user cold-start problem remains a major challenge in personalized e-commerce product recommendation because new users have little or no interaction history, limiting the effectiveness of conventional collaborative filtering and rating-based models. Textual information, such as user reviews, item reviews, product descriptions, metadata, and semantic product attributes, can provide valuable initial signals for modeling user preferences and product characteristics. This study aims to identify textual data sources, AI/NLP methods, and model families suitable for addressing the user cold-start problem in e-commerce product recommendation. A systematic literature review was conducted following the PRISMA 2020 framework, covering 55 selected articles from the Scopus database that were screened and extracted with the assistance of Watase. The studies were categorized into 31 core studies and 24 supporting studies and assessed using five quality criteria. Unlike previous cold-start reviews that broadly map mitigation strategies, this review narrows its scope to text-based evidence and links each model family to the data conditions required for its application. The results show that review text is the dominant textual data source, appearing in 33 of the 55 studies, followed by product descriptions and metadata. The studies were mapped into seven model families, and 43 studies (78.2%) explicitly reported evaluation metrics. Prominent model directions include hybrid deep sentiment-aware recommenders, review-based transformer or attention models, SBERT/BERT-based recommendation, graph-based recommendation, and ontology-based semantic recommendation. Based on the reported evidence, text-semantic hybrid recommenders represent a promising direction, although differences in datasets, evaluation protocols, and metrics across studies prevent direct performance comparisons.