Large-Scale Multi-Platform Product Review Dataset for Sentiment Analysis: Construction, Quality Control, And Cross-Platform Benchmarking
This study develops and evaluates a large-scale multi-platform product review benchmark for sentiment analysis. Public review data were integrated through provenance auditing, unified schema mapping, quality control, conservative deduplication, sentiment harmonization, source-aware sampling, and stratified experimentation. Four platforms Amazon, Flipkart, Lazada, and Tokopedia were retained after quality control. The sentiment-ready master dataset contains 877,394 reviews, while the multi-platform benchmark contains 509,315 reviews. The benchmark was divided into 407,452 training, 50,931 validation, and 50,932 testing reviews using stratification by platform and sentiment. Two classifiers were evaluated: TF-IDF with Linear Support Vector Machine (SVM) and IndoBERT. TF-IDF+SVM achieved 93.21% accuracy and 83.11% Macro-F1, whereas IndoBERT achieved 93.79% accuracy and 83.46% Macro-F1. Although IndoBERT obtained a higher Macro-F1, the paired bootstrap 95% confidence interval for the difference included zero, so the aggregate improvement was not statistically conclusive. An exact McNemar test showed a significant difference in paired prediction behavior (p = 5.737 × 10⁻¹⁰). Performance was platform-dependent, with IndoBERT providing its clearest gains on Lazada and Tokopedia. Leave-One-Platform-Out evaluation produced substantially lower Macro-F1 values than random stratified evaluation, indicating platform/domain shift. The results demonstrate the importance of dataset quality and cross-platform evaluation for sentiment analysis on heterogeneous product reviews.
Authors
- Jati Sasongko Wibowo (ORCID: https://orcid.org/0000-0001-8887-819X)
- Mardi Siswo Utomo (ORCID: https://orcid.org/0009-0008-8972-908X)
- Eddy Nurraharjo (ORCID: https://orcid.org/0000-0003-2015-5993)
- Herny Februariyanti (ORCID: https://orcid.org/0000-0002-1031-1238)
- Sulastri .
Institutions
- Universitas Stikubank (ID)
Publication Details
- Journal
- Engineering and Technology Journal
- Published
- 2026-11-21
- DOI
- https://doi.org/10.47191/etj/v11i11.10
- Primary Topic
- Sentiment Analysis and Opinion Mining
- Type
- article
- Field-Weighted Citation Impact
- 0.00