JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE) ›› 2019, Vol. 54 ›› Issue (7): 57-67.doi: 10.6040/j.issn.1671-9352.1.2018.077

Previous Articles     Next Articles

Review spam detection based on the two-level stacking classification model

Xiang-wen LIAO1,2,3,*(),Yang XU1,2,3,Jing-jing WEI4,Ding-da YANG1,2,3,Guo-long CHEN1,2,3   

  1. 1. College of Mathematics and Computer Science, Fuzhou University, Fuzhou 350116, Fujian, China
    2. Fujian Provincial Key Laboratory of Network Computing and Intelligent Information Processing, Fuzhou University, Fuzhou 350116, Fujian, China
    3. Digital Fujian Institute of Financial Big Data, Fuzhou 350116, Fujian, China
    4. College of Electronics and Information Science, Fujian Jiangxia University, Fuzhou 350108, Fujian, China
  • Received:2018-10-17 Online:2019-07-20 Published:2019-06-27
  • Contact: Xiang-wen LIAO E-mail:liaoxw@fzu.edu.cn
  • Supported by:
    国家自然科学基金资助项目(61772135);国家自然科学基金资助项目(U1605251);福建省自然科学基金资助项目(2017J01755);中国科学院网络数据科学与技术重点实验室开放基金课题(CASNDST201708);中国科学院网络数据科学与技术重点实验室开放基金课题(CASNDST201606);北邮可信分布式计算与服务教育部重点实验室主任基金资助(2017KF01)

Abstract:

For the issue of review spam detection, on the one hand, the time and space complexity of existing methods is high when extracting user behavior relationships and training neural network. On the other hand, the non-standard writing format of E-commercial reviews leads to the indistinct contextual features and most experiment did not consider the effect of the imbalance of data. Therefore, we propose a method for review spam detection based on a two-level stacking classification model. In the method, the relationship between users and products is represented by a triplet. In order to characterize user's behavior and reduce complexity, low-dimensional feature representations are obtained by the principal component analysis. Then, the extracted paragraphs vector representation, information entropy and text similarity is represented as discrete feature to avoid indistinct of contextual features. Finally, the three connections are taken as the overall features combining text and behavioral features. These features are regarded as the input of the two-level stacking classification model in order to improve performance in unbalanced dataset. We conducted experiments in the Yelp 2013 dataset. Experimental results show the F1 value of our proposed method is 1.7%—5.2% better than the state-of-the-art method. What's more, the classification performance is significantly improved in the unbalanced dataset.

Key words: review detection, feature fusion, ensemble learning, principal component analysis

CLC Number: 

  • TP391

Fig.1

Review spam detection based on two-layer stacking model"

Table 1

Discrete feature"

类别特征
用户行为离散特征用户粉丝数(FC)
用户一天内最大发布评论数(MN)
用户极端评论比例(包括1星或者5星)(ER)
用户评论等级分布信息熵(RE)
评论文本离散特征该评论与其他同商品评论的平均等级差(AR)
是否是极端评论(1星或5星)(IE)
该评论与同商品评论最大余弦相似度(MC)
该评论被点赞次数(AN)

Fig.2

Construction of classification model based on ensemble learning"

Table 2

Review spam detection algorithm based on two-layer stacking classification model"

基于双层堆叠分类模型的水军评论检测算法
输入:评论数据集合X{x1, x2, …, xn}、预设参数
输出:评论检测结果集合Y{y1, y2, …, yn}
1:初始化模型参数;
2:为获取低维用户关系向量T,利用公式(3)进行主成分分析和SVD奇异值分解;
3:利用Doc2vec方法得到评论上下文的段落向量表示Di;
4:计算用户交互行为获取离散特征集Li;
5:拼接特征F=concatenate{T, D, L}得到总特征表示;
6:对基模型进行模型选择并构造双层堆叠分类器;
7:利用公式(9)进行交叉验证预测,其结果作为新特征表示;
8:利用公式(15)学习训练融合模型并对新数据分类;
9:输出结果Y{y1, y2, …, yn}

Table 3

Dataset statistics"

数据项目酒店饭店
水军评论数8028 368
非水军评论数4 87650 149
水军评论占比14.1%14.3%
总评论数5 67858 517
总评论者数5 12435 593

Table 4

Experimental parameter setting"

参数类型可调参数参数值
PCA首次降维目标维度150
第二次降维目标维度(M+N)/2
Doc2vec段落向量维度80
词向量训练模型c-bow
堆叠分类器交叉验证次数5
XGBoost分类器学习率0.005
最大深度5

Table 5

Models evaluation results"

方法数据分布酒店数据集饭店数据集
PrecisionRecallF1AccuracyPrecisionRecallF1Accuracy
M_BF+BIGRAM50:5082.886.984.885.182.888.585.683.3
ND46.582.559.484.948.287.962.378.5
HAAT50:5061.354.757.864.469.459.063.866.5
ND32.753.140.856.435.978.948.168.3
SPEAGLE50:5075.783.079.181.080.583.281.882.5
ND26.556.036.080.448.270.558.682.0
Rescal+BIGRAM50:5084.289.987.086.586.891.889.289.9
ND48.285.061.585.958.290.370.887.8
Our_Method50:5087.390.788.988.888.793.290.990.7
ND52.090.065.986.664.692.476.088.3

Table 6

Comparison of different feature extraction methods"

方法数据分布酒店数据集饭店数据集
PrecisionRecallF1AccuracyPrecisionRecallF1Accuracy
Rescal50:5083.487.084.584.182.888.585.683.3
ND44.387.758.883.653.988.366.983.4
Our_Method(SVM)50:5085.486.085.785.786.893.189.889.5
ND48.686.362.185.061.392.473.787.3

Table 7

Two types of reviews typical discrete feature"

特征离散型交互特征数据
FCMNERREARIEMCAN
A19.03.00.140.791.80.00.025.0
B0.01.01.00.02.61.00.090.0

Table 8

Comparison of discrete feature effects"

方法评价指标酒店数据集
(50:50)
酒店数据集
(ND)
饭店数据集
(50:50)
饭店数据集
(ND)
不考虑离散交互特征F187.258.890.373.1
A86.280.189.286.6
考虑离散交互特征(全特征)F188.965.990.976.0
A88.886.690.788.3

Fig.3

Comparison of F1 values of different classifier models"

Fig.4

Impact of different distribution datasets on F1 values"

1 OTT M, CHOI Y, CARDIE C, et al. Finding deceptive opinion spam by any stretch of the imagination[C]// Proceedings of the Meeting of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg: ACM, 2011: 309-319.
2 KIM S, CHANG H, LEE S, et al. Deep semantic frame-based deceptive opinion spam analysis[C]// Proceedings of the 24th ACM International on Conference on Information and Knowledge Management. New York: ACM, 2015: 1131-1140.
3 KO M C, CHEN H H. Analysis of cyber army's behaviours on web forum for elect campaign[C]// Proceedings of the Asia Information Retrieval Symposium. Switzerland: Springer, Cham, 2015: 394-399.
4 LI Huayi, FEI Geli, SHAO Weixiang, et al. Bimodal distribution and co-bursting in review spam detection[C]// Proceedings of the International Conference on World Wide Web. Republic and Canton of Geneva, Switzerland: International World Wide Web Conferences Steering Committee, 2017: 1063-1072.
5 REN Yafeng, ZHANG Yue. Deceptive opinion spam detection using neural network[C]// Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. Osaka: The COLING 2016 Organizing Committee, 2016: 140-150.
6 WANG Xuepeng, LIU Kang, ZHAO Jun. Handling cold-start problem in review spam detection by jointly embedding texts and behaviors[C]// Proceedings of the Meeting of the Association for Computational Linguistics. Vancouver: ACM, 2017: 366-376.
7 KIM Y. Convolutional neural networks for sentence classification[C]// Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. Doha: EMNLP, 2014: 1746-1751.
8 SANTOSH K C, MAITY S K, MUKHERJEE A. ENWalk: learning network features for spam detection in twitter[C]// Proceedings of the International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation. Switzerland; Springer, Cham, 2017: 90-101.
9 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. New York: ACM, 2015: 985-994.
10 WANG Xuepeng, LIU Kang, HE Shizhu, et al. Learning to represent review with tensor decomposition for spam detection[C]// Proceedings of the Conference on Empirical Methods in Natural Language Processing. Austin: EMNLP, 2016: 866-875.
11 WANG Yalin , SUN Kenan , YUAN Xiaofeng , et al. A novel sliding window PCA-IPF based steady-state detection framework and its industrial application[J]. IEEE Access, 2018, 6: 20995- 21004.
doi: 10.1109/ACCESS.2018.2825451
12 LE Q, MIKOLOV T.Distributed representations of sentences and documents[C]// Proceedings of the International Conference on Machine Learning. Beijing: JMLR, 2014: 1188-1196.
13 CHEN Yijun, MAN Leungwong.Optimizing stacking ensemble by an ant colony optimization approach[C]// Proceedings of the 13th Annual Conference Companion on Genetic and Evolutionary Computation. New York: ACM, 2011: 7-8.
14 SANTOSH K C, ARJUN Mukherjee. On the temporal dynamics of opinion spamming: case studies on yelp[C]// Proceedings of the 25th International Conference on World Wide Web. Republic and Canton of Geneva, Switzerland: WWW, 2016: 369-379.
15 MUKHERJEE A, VENKATARAMAN V, LIU B, et al. What yelp fake review filter might be doing[C]// Proceedings of the International AAAI Conference on Web and Social Media. Menlo Park: AAAI, 2013: 409-418.
16 HAI Zeng, ZHAO Peilin, CHENG Peng, et al. Deceptive review spam detection via exploiting task relatedness and unlabeled data[C]// Proceedings of the Conference on Empirical Methods in Natural Language Processing. Austin: EMNLP, 2016: 1817-1826.
17 FAKHRAEI S, SHASHANKA M. Collective spammer detection in evolving multi-relational social networks[C]// Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2015: 1769-1778.
[1] LI Run-chuan, ZAN Hong-ying, SHEN Sheng-ya, BI Yin-long, ZHANG Zhong-jun. Spam messages identification based on multi-feature fusion [J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2017, 52(7): 73-79.
[2] SHAO Wei1, ZHU Li-ping2, LIU Fu-Guo2, WANG Qiu-Ping2. Sparse principal component analysis for symmetric matrix and  application in sufficient dimension reduction [J]. J4, 2012, 47(4): 116-120.
[3] ZHOU Juan1, WANG Ren-qing2, GUO Wei-hua2*, WANG Qiang2, WANG Wei2, . Soil geochemical elements in the Yutai high quality rice base [J]. J4, 2012, 47(3): 5-9.
[4] WANG De-liang, LI Ke, LU Li-ling. Analysis of habitat characteristics of Tanichthys albonubes in Shimen National Forest Park [J]. J4, 2012, 47(3): 1-4.
[5] ZHU Shi-wei,SAI Ying . The prediction model of financial distress of Chinese listed corporations based on a hybrid RPR model [J]. J4, 2008, 43(11): 48-53 .
[6] YANG Shao-hua,LIN Pan,PAN Chen . Performance improvement of face recognition based on kernel principal component analysis using wavelet transform [J]. J4, 2007, 42(9): 96-100 .
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
[1] ZHANG Jing-you, ZHANG Pei-ai, ZHONG Hai-ping. The application of evolutionary graph theory in the design of knowledge-based enterprises’ organization strucure[J]. J4, 2013, 48(1): 107 -110 .
[2] GUO Lan-lan1,2, GENG Jie1, SHI Shuo1,3, YUAN Fei1, LEI Li1, DU Guang-sheng1*. Computing research of the water hammer pressure in the process of #br# the variable speed closure of valve based on UDF method[J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2014, 49(03): 27 -30 .
[3] SHI Kai-quan. P-information law intelligent fusion and soft information #br# image intelligent generation[J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2014, 49(04): 1 -17 .
[4] TANG Xiao-hong1, HU Wen-xiao2*, WEI Yan-feng2, JIANG Xi-long2, ZHANG Jing-ying2, SHAO Xue-dong3. Screening and biological characteristics studies of wide wine-making yeasts[J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2014, 49(03): 12 -17 .
[5] ZENG Weng-fu1, HUANG Tian-qiang1,2, LI Kai1, YU YANG-qiang1, GUO Gong-de1,2. A local linear emedding agorithm based on harmonicmean geodesic kernel[J]. J4, 2010, 45(7): 55 -59 .
[6] GUO Wen-juan, YANG Gong-ping*, DONG Jin-li. A review of fingerprint image segmentation methods[J]. J4, 2010, 45(7): 94 -101 .
[7] MENG Xiang-bo1, ZHANG Li-dong1, DU Zi-ping2. Investment and reinsurance strategy for insurers under #br# mean-variance criterion with jumps#br#[J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2014, 49(05): 36 -40 .
[8] PENG Zhen-hua, XU Yi-hong*, TU Xiang-qiu. Optimality conditions for weakly efficient elements of nearly preinvex set-valued optimizaton#br#[J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2014, 49(05): 41 -44 .
[9] HU Ming-Di, SHE Yan-Hong, WANG Min. Topological properties of  three-valued   logic  metric space[J]. J4, 2010, 45(6): 86 -90 .
[10] HE Hai-lun, CHEN Xiu-lan* . Circular dichroism detection of the effects of denaturants and buffers on the conformation of cold-adapted protease MCP-01 and  mesophilic protease BP01[J]. JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE), 2013, 48(1): 23 -29 .