基于聚类微调的在线商品虚假评论识别

doi:10.7523/j.ucas.2025.014

摘要/Abstract

摘要：

虚假评论影响在线消费者的购买决策，如何高效识别虚假评论是当前电子商务发展中亟待解决的问题。传统虚假评论识别方法容易受到评论文本风格、句法词法和上下文差异的影响，准确性偏低；大语言模型（LLM）虽然可以解决这一问题，但训练过程往往耗费机时过长。为此，提出一种新的基于聚类微调的在线商品虚假评论识别方法（CF-DRI）。该方法通过筛选聚类评论样本来微调LLM的预训练知识，以显著提升LLM识别虚假评论的训练效率。相较于传统方法，CF-DRI方法在微调样本数量较少的情况下即可表现出优异的识别性能。基于Yelp.com数据集的实验结果表明：CF-DRI方法仅利用20%的聚类样本，即可实现92.29%的虚假评论识别精确率和90.03%的识别召回率。

关键词: 虚假评论识别, 大语言模型, 微调, 聚类

Abstract:

Fake reviews affect online consumers’purchasing decisions. Efficiently identifying fake reviews is a pressing issue in the current development of e-commerce. Traditional methods for detecting fake reviews are often influenced by variations in review text style， syntax， and context， resulting in lower accuracy. Although large language models （LLMs） can address this accuracy issue， their training process is typically time-consuming. To tackle this problem， we propose a novel method called CF-DRI （cluster-based fine-tuning for deceptive review identification）. This method fine-tunes the pre-trained knowledge of LLMs by selecting clustered review samples， significantly enhancing the training efficiency for fake review identification. Compared to traditional methods， CF-DRI demonstrates superior performance with fewer fine-tuning samples. Experimental results on the Yelp.com dataset show that CF-DRI achieves a precision of 92.29% and a recall of 90.03% in fake review identification using only 20% of the clustered samples. This research provides new perspectives and solutions for managing fake reviews on e-commerce platforms， potentially promoting healthy industry development.

Key words: fake review identification, large language models, fine-tuning, clustering

中图分类号:

TP39

刘津浩, 权沛, 张文. 基于聚类微调的在线商品虚假评论识别[J]. 中国科学院大学学报, 2026, 43(2): 230-239.

Jinhao LIU, Pei QUAN, Wen ZHANG. Fake review identification for online products based on clustering fine-tuning[J]. Journal of University of Chinese Academy of Sciences, 2026, 43(2): 230-239.

图/表 6

图1 CF-DRI：基于聚类微调的在线商品虚假评论识别方法

Fig.1 CF-DRI：cluster-based fine-tuning for online product fake review identification

图2 评论星级与真假数量分布图

Fig.2 Distribution of review ratings and fake/real counts

表1 基于深度学习的实验结果 (%)

Table 1 Experimental results based on deep learning

模型	精确率	召回率	F₁
LSTM	82.21	83.77	82.91
CNN	81.56	80.87	81.19
LSTM-CNN	84.20	82.97	83.58

表2 随机采样微调的实验结果 (%)

Table 2 Experimental results of random sampling fine-tuning

	精确率	召回率	F₁
随机5%	85.13	83.67	84.39
随机10%	88.23	85.85	87.03
随机20%	90.26	88.97	89.61
所有数据	92.91	90.39	91.38

表3 基于轮廓系数采样策略的微调实验结果 (%)

Table 3 Experimental results of fine-tuning based on silhouette coefficient sampling

抽样	精确率	召回率	F₁
轮廓系数最高5%	87.05	84.03	85.51
轮廓系数最高10%	90.78	89.02	89.89
轮廓系数最高20%	91.27	89.42	90.34
轮廓系数大小共5%	89.22	86.68	87.94
轮廓系数大小共10%	91.91	89.30	90.59
轮廓系数大小共20%	92.29	90.03	91.15

表4 基于欧氏距离采样策略的微调的实验结果 (%)

Table 4 Experimental results of fine-tuning based on Euclidean distance sampling

抽样	精确率	召回率	F₁
距离最近5%	86.71	85.12	85.92
距离最近10%	90.62	87.98	89.31
距离最近20%	91.28	89.01	90.14
距离远近共5%	88.11	85.32	86.71
距离远近共10%	91.31	88.85	90.06
距离远近共20%	91.89	89.80	90.83

参考文献 31

[1]	Paul H， Nikolaev A. Fake review detection on online E-commerce platforms： a systematic literature review［J］. Data Mining and Knowledge Discovery， 2021， 35（5）： 1830-1881. DOI： 10.1007/s10618-021-00772-6 .
[2]	Gupta P， Gandhi S， Chakravarthi B R. Leveraging transfer learning techniques- BERT， RoBERTa， ALBERT and DistilBERT for fake review detection［C］//Proceedings of the 13th Annual Meeting of the Forum for Information Retrieval Evaluation. December 13 - 17， 2021， Virtual Event， India. ACM， 2021： 75-82. DOI： 10.1145/3503162.3503169 .
[3]	Mohawesh R， Xu S X， Tran S N， et al. Fake reviews detection： a survey［J］. IEEE Access， 2021， 9： 65771-65802. DOI： 10.1109/ACCESS.2021.3075573 .
[4]	Refaeli D， Hajek P. Detecting fake online reviews using fine-tuned BERT［C］//Proceedings of the 2021 5th International Conference on E-Business and Internet. October 15 - 17， 2021， Singapore， Singapore. ACM， 2021： 76-80. DOI： 10.1145/3497701.3497714 .
[5]	Hu B Z， Sheng Q， Cao J， et al. Bad actor， good advisor： exploring the role of large language models in fake news detection［J］. Proceedings of the AAAI Conference on Artificial Intelligence， 2024， 38（20）： 22105-22113. DOI： 10.1609/aaai.v38i20.30214 .
[6]	Huang Y， Sun L， FakeGPT： fake news generation， explanation and detection of large language models［EB/OL］.arXiv 2023：2310.05046.（2023-10-08）［2024-08-23］..
[7]	Ke J， Xu Z， Xu T，et al.An implicit semantic enhanced fine-grained fake news detection method based on large language models［J］. Journal of Computer Research and Development， 2024， 61（5）： 1250-1260.DOI： 10.7544/issn1000-1239.202330967 .
[8]	Devlin J， Chang M， Lee K， et al. BERT： pre-training of deep bidirectional transformers for language understanding［C］//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics： Human Language Technologies， Vol 1（Long and Short papers）： 4171-4186.
[9]	Oliver M， Wang G. Crafting efficient fine-tuning strategies for large language models［EB/OL］.arXiv 2023：2407.13906.（2023-07-18）［2024-08-23］..
[10]	Singh A， Pandey N. A study of optimizations for fine-tuning large language models［EB/OL］.arXiv 2024：2406.02290.（2024-06-06）［2024-08-23］..
[11]	Duma R A， Niu Z D， Nyamawe A S， et al. Fake review detection techniques， issues， and future research directions： a literature review［J］. Knowledge and Information Systems， 2024， 66（9）： 5071-5112. DOI： 10.1007/s10115-024-02118-2 .
[12]	Zhang D S， Zhou L N， Kehoe J L， et al. What online reviewer behaviors really matter？ effects of verbal and nonverbal behaviors on detection of fake online reviews［J］. Journal of Management Information Systems， 2016， 33（2）： 456-481. DOI：10.1080/07421222.2016.1205907 .
[13]	Casillo M， Colace F， Gupta B B， et al. Fake news detection using LDA topic modelling and K-nearest neighbor classifier［C］// Computational Data and Social Networks. Cham： Springer International Publishing， 2021： 330-339. DOI： 10.1007/978-3-030-91434-9_29 .
[14]	Kumar N， Venugopal D， Qiu L F， et al. Detecting review manipulation on online platforms with hierarchical supervised learning［J］. Journal of Management Information Systems， 2018， 35（1）： 350-380. DOI：10.1080/07421222.2018.1440758 .
[15]	Zhang W， Zhang X， Chen J D， et al. Stacking GA²M for inherently interpretable fraudulent reviewer identification by fusing target and non-target features［J］.International Journal of General Systems，2024，54（3）：298-333. DOI：10.1080/03081079.2024.2384404 .
[16]	Du Q， Zong C， Zhang J. MoDS： model-oriented data selection for instruction tuning［EB/OL］.arXiv 2023：2311.15653.（2023-11-27）［2024-08-22］..
[17]	Li M， Zhang Y， Li Z T， et al. From quantity to quality： boosting LLM performance with self-guided data selection for instruction tuning［C］//Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics： Human Language Technologies，2024，Vol 1： 7602-7635.DOI： 10.18653/v1/2024.naacl-long.421
[18]	Mehta V， Agarwal M， Kaliyar R K. A comprehensive and analytical review of text clustering techniques［J］. International Journal of Data Science and Analytics， 2024， 18（3）： 239-258. DOI： 10.1007/s41060-024-00540-x .
[19]	Wang H， Zhou C D， Li L X. Design and application of a text clustering algorithm based on parallelized K-means clustering［J］. Revue D’ Intelligence Artificielle， 2019， 33（6）： 453-460. DOI： 10.18280/ria.330608 .
[20]	Petukhova A， Matos-Carvalho J P， Fachada N. Text clustering with large language model embeddings［J］. International Journal of Cognitive Computing in Engineering， 2025， 6： 100-108. DOI： 10.1016/j.ijcce.2024.11.004 .
[21]	Ding N， Qin Y J， Yang G， et al. Parameter-efficient fine-tuning of large-scale pre-trained language models［J］. Nature Machine Intelligence， 2023， 5（3）： 220-235. DOI： 10.1038/s42256-023-00626-4 .
[22]	Hu E J， Shen Y， Wallis P， et al. LoRA： low-rank adaptation of large language models［C］. ICLR 2022 Conference，2022，1912. .
[23]	Pennington J， Socher R， Manning C. Glove： global vectors for word representation［C］//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing （EMNLP）. Doha， Qatar. Stroudsburg， PA， USA： ACL， 2014： 1532-1543. DOI： 10.3115/v1/d14-1162 .
[24]	Hartigan J A， Wong M A. Algorithm AS 136： a K-means clustering algorithm［J］. Applied Statistics， 1979， 28（1）： 100. DOI： 10.2307/2346830 .
[25]	Oyewole G J， Thopil G A. Data clustering： application and trends［J］. Artificial Intelligence Review， 2023， 56（7）： 6439-6475. DOI： 10.1007/s10462-022-10325-y .
[26]	Upadhye A. A survey of text clustering techniques： algorithms， applications， and challenges［J］. International Journal of Science and Research （IJSR）， 2021， 10（9）： 1749-1752. DOI： 10.21275/sr24304163737 .
[27]	Rayana S， Akoglu L. Collective opinion Spam detection： bridging review networks and metadata［C］//Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Sydney NSW Australia. ACM， 2015： 985-994. DOI： 10.1145/2783258.2783370 .
[28]	Loshchilov I， Hutter F. Decoupled weight decay regularization［EB/OL］. arXiv 2017：1711.05101. （2017-11-14）［2024-08-22］. .
[29]	Liu Y， Ott M， Goyal N， et al. RoBERTa： a robustly optimized BERT pretraining approach［EB/OL］.arXiv 2019：1907.11692. （2019-07-26）［2024-08-23］..
[30]	Chawla N V， Bowyer K W， Hall L O， et al. SMOTE： synthetic minority over-sampling technique［J］. Journal of Artificial Intelligence Research， 2002， 16（1）： 321-357. DOI： 10.5555/1622407.1622416 .
[31]	Parsaeimehr E， Fartash M， Akbari Torkestani J. Improving feature extraction using a hybrid of CNN and LSTM for entity identification［J］. Neural Processing Letters， 2023， 55（5）： 5979-5994. DOI： 10.1007/s11063-022-11122-y .