Loading article...
Loading article...
Stream/Area:
-
Subject:
-
Paper ID:
IJSREM68996
Volume:
10 |
11 |
Nov 2026
Pages:
1-5
Publication Frequency:
Monthly (12 issues per year)
License:International Journal of Scientific Research in Engineering and Management follows the Creative Commons Attribution 4.0 (CC BY 4.0) License, allowing free use and distribution with proper attribution.
Department of Computer Science and Engineering
Don Bosco Institute of Technology
Department of Computer Science and Engineering
Don Bosco Institute of Technology
Department of Computer Science and Engineering
Don Bosco Institute of Technology
Department of Computer Science and Engineering
Don Bosco Institute of Technology
Department of Computer Science and Engineering
Don Bosco Institute of Technology
Online reviews strongly affect purchase decisions, which is making them a target for fake and tricky reviews that are misleading consumers and warp product ratings, to a significant degree; the trained model is delivered through a website and a browser extension that are checking reviews as people are reading them, broadly speaking. Published detection studies use diverse datasets, attributes and evaluation protocols, so their reported precisions are hard to compare; this paper benchmarks four classic classifiers (Naïve Bayes, Logistic Regression, Linear SVM and Random Forest) on TF-IDF bigram traits, a Bidirectional LSTM (Bi-LSTM), Graph Convolutional Networks (GCN) and a fine-tuned RoBERTa-base model under one controlled protocol on two corpora, like the Cornell hotel tricky opinion spam corpus (1,600 reviews) and a balanced Amazon fake reviews dataset (about 40,432 reviews). On the Cornell corpus, TF-IDF with a Linear SVM is reaching 87.50% accuracy, close to the 87.81% reported for a BERT-embedding SVM [1]; on the Amazon dataset, Linear SVM is reaching 91.34%, Bi-LSTM 91.52% and a 2-layer GCN 91.21%. On Amazon the Bi-LSTM and GCN differ from the best standard model by less than 0.2 percentage points, so a tuned linear model remains a strong baseline there; indeed, before, fine-tuning RoBERTa-base with 5-fold stratified cross-validation is reaching 93.06% ± 0.72%.
KEYWORDS
This paper benchmarked standard, sequential, and graph-based models for fake review detection on a hotel corpus and a larger Amazon dataset, and delivered the trained model, through a website and a browser extension, in many respects; On Amazon, Bi-LSTM (91.52%) and GCN (91.21%) perform at the same level, within 0.2 points, in real-world scenarios. A TF-IDF bigram Linear SVM is a fast and exact baseline (87.50% on Cornell, 91.34% on Amazon), and the broad spread in published Naïve Bayes results is suggesting that comparisons need a typical protocol. Fine-tuned RoBERTa-base is typically reaching 93.06% ± 0.72% on Cornell, the best finding there. Future work includes checking RoBERTa-base on the Amazon dataset, cross-dataset check, significance testing, listed false-positive assessment, and cost and latency measurements for the rolled out service, broadly speaking.
[1] A. Q. Mir, F. Y. Khan, and M. A. Chishti, “Online fake review detection using supervised machine learning and BERT model,” arXiv:2301.03225, 2023. [2] D. Dilsha and V. P. Lijo, “Opinion spam detection in online reviews using machine learning techniques,” Int. J. Eng Res Technol. (IJERT), vol. 9, no. 6, 2020. [3] R. Mohawesh, S. Xu, S. N. Tran, R. Ollington, M. Springer, and Y. Jararweh, “Fake reviews detection: A survey,” IEEE Access, vol. 9, pp. 65753-65769, 2021. [4] K. L. Tan, C. P. Lee, K. S. M. Anbananthen, and K. M. Lim, “RoBERTa-LSTM: A hybrid model for sentiment analysis with transformer and recurrent neural network,” IEEE Access, vol. 10, pp. 21517-21525, 2022. [5 ] Y. Liu et al., “RoBERTa: A robustly optimized BERT pretraining approach,” arXiv:1907.11692, 2019.
Publication Date
11/14/2026
Peer Review Type
Double-Blind Peer Review
Review Rounds: 2
Documents
Article Views
0
Total views since publication
PDF Downloads
2
Total downloads since publication
Geographic Reach
5
Digital Preservation: This article is permanently archived on Zenodo and complies with all data preservation standards across major indexing platforms.
Plagiarism Screening
This manuscript has been checked for plagiarism using TurnitinPlagiarismChecker software.
Conflict of Interest
The authors declare no conflict of interest related to this research work.
Funding Information
No funding information available for this research.
Related research in the same subject area
Loading similar articles...