2024 International Conference on Expert Clouds and Applications (ICOECA)

Predicting the Fake Review to Amazon Product

Review Dataset using Fuzzy Optimized
Convolution Neural Network
1st Aswaththaman A 2nd Lokesh N 3rd , Parthiban D
Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
2024 International Conference on Expert Clouds and Applications (ICOECA) | 979-8-3503-8579-3/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICOECA62351.2024.00125

Karpagam Academy of Higher Karpagam Academy of Higher Karpagam Academy of Higher

Education Education Education
Coimbatore Coimbatore Coimbatore

4th Gokilavani
Department of Computer Science and
Engineering
Karpagam Academy of Higher
Education
Coimbatore
[email protected]*

Abstract— E-commerce product reviews have become I. INTRODUCTION

crucial for online users in making purchase decisions,
particularly on platforms like Amazon and Flipkart, where user
Sentiment analysis is useful in predicting and classifying the
interactions and feedback are highly emphasized. Recently, user opinion [1]. Sentiment analysis is particularly significant
sentiment analysis is employed to assess user opinions on in predicting the changes in user opinion [2]. E-Commerce
products by extracting valuable insights regarding various user data helps to decide the purchase decision [3]. User
features and benefits. These reviews are widely distributed to intention on many content is extensively employed to extract
leverage and disseminate user perspectives, providing valuable and predict the user opinion [4]. Those data are primarily
information on user sentiments in comment sections to extracted from user sentiment analysis and opinion mining,
differentiate positive and negative feedback. With the which reveals the normal and fake review of users [5]. As
proliferation of Artificial Intelligence (AI) technologies, Deep Learning (DL) and Artificial Intelligence (AI)
detecting fake reviews has become imperative to ensure the
authenticity of user opinions. To address this, a fuzzy-based
technologies is advancing, it becomes imperative to predict
Convolutional Neural Network (CNN) is proposed to the evolution of user opinion in assessing the long-term
differentiate genuine and fake product reviews based on user reliance of users over numerous time frames. In order to give
opinions and their review across similar products. The proposed useful forecasting, a novel fuzzy optimized convolution
architecture is designed to be resilient to temporal changes in neural network is constructed.
user behavior. Initially, data preprocessing is performed to The proposed model is capable of handling the
obtain user profile and product characteristics and parse temporal and time-changing features. In this process. stop
statements. The preprocessed data is then subjected to Principal word removal and stemming is carried out as a preprocessing
Component Analysis (PCA) to extract features representing step. Normalized opinion is employed to perform POS
user intentions. These features are given as input to the fuzzy-
based CNN to analyze user behavior across various aspects. The
tagging for work parsing. Parse data is managed as a feature.
obtained optimal behavioral data is utilized to train the fuzzy- That feature is profiled as user features and product features.
optimized CNN, which provide valuable predictions on changes Profiled features are employed in fuzzy optimized CNN
in user opinion towards different products. The proposed architecture to predict and classify the review using various
architecture employs convolutional layers, pooling layers, and layers of the network. Especially convolution and max
fully connected networks for review classification and fake pooling layer generates the feature map and fully connected
review prediction. Performance evaluation on an Amazon layer composing Sofmax function, activation function and
product review dataset demonstrates superior performance in loss function processes the feature map as fuzzy set [7] to
terms of precision, recall, accuracy, and F-measure compared to classify it as fake review and normal review using fuzzy rules
conventional architectures.
in softmax function [8].
Keywords— Product Review, Online Social Networks, Fuzzy The remaining paper is segmented as following,
Optimized Convolution Neural Network, Principle Component section 2 mentions the review of traditional work related to
Analysis, Sentiment Analysis sentiment analysis and opinion mining, section 3 provides the
design of the present architecture as fuzzy optimized
Convolution Neural Network (CNN) to predict and classify
the fake review. Experimental analysis of the proposed model

979-8-3503-8579-3/24/$31.00 ©2024 IEEE 689

DOI 10.1109/ICOECA62351.2024.00125
Authorized licensed use limited to: Panimalar Engineering College - Chennai. Downloaded on November 09,2024 at 07:32:03 UTC from IEEE Xplore. Restrictions apply.
 is Network Parameter Values predicted and classified on processing the user opinion to the
product on its implicit and explicit attributes. Table 1
Batch size 108 illustrates the hyper parameter of fuzzy optimized
convolution neural network.
Learning Rate 0.06 Table 1: Hyper Parameter Tuning of Fuzzy Optimized Convolution Neural
Network
Epoch 45

Activation Function ReLu

Error function Cross entropy

presented and evaluated using performance measures in

section 4 through benchmark amazon dataset. Finally, the
article is summarized in section 5.
x Convolution layer

II. RELATED WORKS In this layer, highly weighted product feature

This section mentions a detailed study of the sentiment opinion of the user are extracted and feature map is generated
analysis technique based on the extraction of user feature and in this model. It utilizes the activation function to process the
product features to determine the product sentiment on basis review on the product feature mentioned as implicit and
of different classes. Sentiment analysis is carried out using explicit features in the vector form.
machine learning and deep learning architectures.
x Max pooling Layer
a. Tripartite Graph for Opinion Evolution prediction
In max pooling layer, feature vectors of the user
explicit and implicit features are extracted and represented in
Specified approach employs a tripartite graph to
tree structure for efficient evaluation of the user opinion
highlight the evolution of user opinions while applying
evolution. Further, the user state mentions the opinion of the
sentiment analysis to user review on multiple product
user extracted. The max pooling layer of the proposed
context. Sentiment analysis is incorporated to analysis the
architecture employs the ReLU function as its activation
dependency structures of the user feature to product features.
function to calculate the state changes of the user to the
Opinion features is classified using Naive Bayes or C 4.5
specified product features. Figure 1 mentions the architecture
method as opinion class [10].
of the current paradigm.

III. PROPOSED MODEL

A comprehensive design of the fuzzy optimized
Convolution Neural Network is implemented towards
prediction and classification of fake review on processing the
user product reviews on the product features on the multiple
time frame. Hyper Parametric tuning of the CNN layers has
been carried on basis of activation function, softmax function
and error function is to generate the effective prediction
architecture to the fake review prediction on processing the
user reviews.
a. Data Pre-processing
Amazon dataset is pre-processed to remove the stop word
and normalize the words using stemming process.
Normalized words is parsed using POS tagging to generate
the feature vector. Generated feature vector containing user
information and product information is profile as user
profiling and product profiling

User profiling
It extracts the explicit and implicit attributes of the Figure 1: Prediction Architecture of fake review
user to specified product feature on opinion. It is given as
Up = {UI1, UE1UI2, UE2.... UI2, UE2} The implicit and explicit information of the user review
to product feature represented as feature map in the
b. 3.2. Fuzzy optimized Convolution Neural Network Convolution layer is used for sentiment analysis.
Fuzzy optimized Convolution Neural Network is a
type of convolution Neural Network to pedict and classify the x Softmax Function
user opinion and its evolution. In this work, fake review is

690

Authorized licensed use limited to: Panimalar Engineering College - Chennai. Downloaded on November 09,2024 at 07:32:03 UTC from IEEE Xplore. Restrictions apply.
 Softmax function uses fuzzy rules to process the a. Dataset Description
feature map. Fuzzy rule applied to fuzzy set classify the Fake The detailed analysis of the model is carried out using
review. Further fuzzy set mentioned to the inference model amazon dataset [19], which contains nearly 2lakh product
with fuzzified input. The Standard illustration of fuzzy review composed of original and fake review. It extracted
inference model are Mamdani and Sugeno. using data crawler approaches and it is applied to experiment
analysis to compute the user review evolution to the multiple
Algorithm 1: Prediction of fake review on user opinion product aspect and to compute the opinion evolution of the
evolution to product review user on the various time frame using python framework.

Input: Amazon Dataset b. Evaluation

Output: fake review prediction Performance of current architecture towards
Process prediction of user opinion evolution and fake review to
Data Pre- Process () product is evaluated against user opinion prediction and
User Profiling () sentiment prediction model to compute the fake review using
Extraction Intention and Preference conventional deep learning approaches. Further current
Apply GCNN Learning () approach produces enhanced prediction accuracy to review
Convolution Layer() classification on 5-fold validation [20] and cross entropy
Computes user intensions to various features function is applied as loss function of the model.
Feature mapping of user intention
Max pooling Layer () c. Precision
Extract the user preference to high ranked It is a ratio of relevant user intention to the extracted
features() user intention to the multiple product context. It computes the
Compute the high level features intention as true positive and false positive value using
Feature mapping of user intention confusion matrix.
Fully Connected Layer ()
Activation Function ()
Employ ReLU Function Tanh()
Cross Entropy Layer ()
Minimize Reconstruction Error ()
Output Layer ()
Fuzzy Rule for classification and prediction Opinion
Class= {Normal, Fake )
Output: Fake Review Prediction

IV. EXPERIMENTAL RESULTS

Experimental analysis of Fuzzy optimized Convolution
Neural Network approach is modified using parametric
tuning of the network to amazon dataset is processed for
predicting fake review to product aspects on different time
Figure 2: Performance evaluation of review classification architectures on
period. Amazon dataset related various electronics product is basis of Precision Measure
extracted and transformed using data transformation and .
normalization methods. Normalized information is Precision of the proposed approach produces 99%
partitioned into two segment representing as Training data respectively on compared with traditional architectures.
and Testing data. Figure 2 mentions the performance assessment of the user
Training data is applied to train the model and test opinion prediction approach on basis of the precision.
data is to predict the efficiency of the model. Test data is ୘୰୳ୣ୮୭ୱ୧୲୧୴ୣ
Precision =
partitioned into cross fold validation. In this work, 5-fold ்௥௨௘௣௢௦௜௧௜௩௘ାி௔௟௦௘௉௢௦௜௧௜௩௘

validation is done using confusion matrix to compute the

performance of the prediction model on basis of precision, d. Recall
recall and f-measure metrics. The training parameter of the Recall is ratio of irrelevant user intention to the
optimized CNN model is mentioned in table 2 extracted user intention to the various product features. It is
compute true positive, false negative and false positive value
Table 2: Hyper parameters for Optimized model of various context using the confusion matrix. Recall of the
Network Parameter Value present approach produces 97% respectively on compared
Learning rate 0.01 with traditional technique. Figure 3 mentions performance
assessment of the prediction architectures on basis of the
Loss Function cross entropy
recall.
Activation function ReLU ୘୰୳ୣ୮୭ୱ୧୲୧୴ୣ
Epoch 40 Recall =
்௥௨௘௣௢௦௜௧௜௩௘ାி௔௟௦௘௡௘௚௔௧௜௩௘

691

Authorized licensed use limited to: Panimalar Engineering College - Chennai. Downloaded on November 09,2024 at 07:32:03 UTC from IEEE Xplore. Restrictions apply.
 performance of the prediction architecture based on the f
measure against conventional approaches to calculate the
fake review rate on user opinions. Table 3 mentions the
performance analysis of different sentiment analysis
architecture to calculate the user review and its evolution.

Table 3: Performance Evaluation

Architectu Precision Recall F-Measure
re
Trai Test Trai Test Trai Test
n n n
Fuzzy 99.2 98.4 97.5 96.4 98.7 99.5
optimized 8 5 4 8 9 8
Convolutio
Figure 3: Performance evaluation of review classification architectures on
basis of Recall Measure
n Neural
Network-
Prediction quality of the user review classification Proposed
model is to process the user intension on various product Convolutio 90.2 90.7 88.9 88.5 96.7 96.5
context using the activation function of the model as it n Neural 6 8 8 8 8 8
simulates the state changes of user feature. Network-
Existing

V. CONCLUSION

This study has designed and implemented a fuzzy

Convolutional Neural Network (CNN) for performing
sentiment analysis of user opinions in Amazon product
reviews. The approach utilized activation functions to
enhance discrimination and predict user evolution and
adaptation rates over various time frames for capturing
sentiment associated with user intentions across different
contexts. Additionally, the architecture has integrated fuzzy
logic algorithms into the softmax function to analyze user
intention behavior at discrete time intervals and identify fake
reviews. Hyperparameter tuning included activation
Figure 4: Performance evaluation of review classification functions and loss functions to optimize prediction and
architectures on F Measure classification accuracy. Performance analysis demonstrates
that the proposed architecture outperforms traditional
Further convolution layer and max pooling layer approaches in terms of both accuracy and efficiency.
generate the maps the state changes of the user. Especially
fully connected layer computes the predict and classifies the REFERENCES
user reviews as fake review and original reviews on [1] Min E, Guo X, Qiang “A survey of clustering with deep learning: from
processing the feature map and feature vector representing the perspective of network architecture” in IEEE Access, Vol. 6,
issue.39, pp: 501–14, 2018.
the user intensions to the various product features.
Furthermore, F measure is an accuracy metric used to [2] Chowdary NS, Prasanna DS, Sudhakar P. “Evaluating and analyzing
compute the quality of the prediction model. clusters in data mining using different algorithms”. International
Journal of Computer Science and Mobile Computing, Vol.3, PP: 86–
e. F measure 99, 2014.
It is mentioned as number of correct user [3] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
predictions class to the fake user prediction classes among the review and new perspectives,” IEEE transactions on pattern analysis
total number user participated on the specified product and machine intelligence, vol. 35, no. 8, pp. 1798–1828, 2013
review. F measure represents accuracy and it is mentioned as
[4] Dizaji KG, Herandi A, Cheng,” Deep clustering via joint convolutional
follows
autoencoder embedding and relative entropy minimization. In: 2017
IEEE International Conference on Computer Vision (ICCV). Venice,
Accuracy is given by Italy: IEEE, 2017, 5747–56.

”—‡’‘•‹–‹˜‡ ൅ ”—‡‡‰ƒ–‹˜‡ [5] Xie J, Girshick R, Farhadi A ”Unsupervised deep embedding for
ܶ‫ ݁ݒ݅ݐ݅ݏ݋݌݁ݑݎ‬൅ ܶ‫ ݁ݒ݅ݐܽ݃݁ܰ݁ݑݎ‬൅ ݂݈ܽ‫ ݁ݒ݅ݐ݅ݏ݋݌݁ݏ‬൅ ‫݁ݒ݅ݐܽ݃݁݊݁ݏ݈ܽܨ‬ clustering analysis. In: International Conference on Machine
Learning. New York City, NY, USA: ICMLR, 2016, 478–87.
The different user intentions are calculated based on
various user adaption rate. Figure 4 mentions the

692

Authorized licensed use limited to: Panimalar Engineering College - Chennai. Downloaded on November 09,2024 at 07:32:03 UTC from IEEE Xplore. Restrictions apply.
 [6] S.Praveen & R.Priya " A Deep Conceptual Incremental learning Based
High Dimensional Data Clustering model- A Deep Learning
Approach" in Turkish Journal of Computer and Mathematics, 2021

[7] H. Liu, M. Shao, S. Li, and Y. Fu, “Infinite ensemble for image
clustering,” in Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM, , pp.
1745–1754, 2016

[8] J. L. Elman, “Distributed representations, simple recurrent networks,

and grammatical structure,” Machine learning, vol. 7, no. 2-3, pp.
195–225,1991.122, 2015.

[9] K. G. Dizaji, A. Herandi, C. Deng, W. Cai, and H. Huang, “Deep

clustering via joint convolutional autoencoder embedding and relative
entropy minimization,” in 2017 IEEE International Conference on
Computer Vision (ICCV). IEEE, , pp. 5747–5756, 2017.

[10] P. Xiao, G. K. Venayagamoorthy, and K. A. Corzine, “Combined

training of recurrent neural networks with particle swarm optimization
and backpropagation algorithms for impedance identification,” in
Proceedings of the IEEE Swarm Intelligence Symposium (SIS ’07),
pp. 9–15, April 2007.

[11] P. Huang,Y. Huang,W.Wang, and L.Wang, ``Deep embedding

network for clustering,'' in Proc. 22nd International. Conference.
Pattern Recognition. (ICPR), Aug. 2014, pp. 1532-1537.

[12] S. Zhang, Y. Zhan, and D. N. Metaxas, “Deformable segmentation

via sparse representation and dictionary learning,” Med. Image Anal.,
vol. 16, no. 7, pp. 1385–1396, Oct. 2012.

[13] W. Harchaoui, P. A. Mattei, and C. Bouveyron, ``Deep adversarial

Gaussian mixture auto-encoder for clustering,'' in Proc. ICLR, 2017,
pp. 1-5.

[14] N. Dilokthanakul et al. (2016). ``Deep unsupervised clustering with

Gaussian mixture variational autoencoders.'' [Online]. Available:
https:// arxiv.org/abs/1611.02648

[15] G. Chen. (2015). ``Deep learning with nonparametric clustering.''

[Online]. Available: https://arxiv.org/abs/1501.03084

[16] S.Praveen & R.Priya "A unified deep learning framework for text data
mining using deep adaptive fuzzy clustering" in Solid state
technology, Vol.63, issue.6, 2020.

[17] J. Chung, Ç. Gülçehre, K. Cho and Y. Bengio, "Empirical Evaluation

of Gated Recurrent Neural Networks on Sequence Modeling," 2014.
[Online]. Available: https://arxiv.org/abs/1412.3555. [Accessed 31 1
2019].

[18] T. Pham, T. Tran, D. Phung and S. Venkatesh, "Predicting healthcare

trajectories from medical records: A deep learning approach," J.
Biomed. Inform., vol. 69, pp. 218-229, 2017.

[19] K. Zhou, H. Zha and L. Song, "Learning triggering kernels for

multidimensional hawkes processes," in Proc. 30th Int. Conf. Machine
learning, Atlanta, GA, 2013, pp. III-1301-III-1309.

[20] T. Ma, C. Xiao and F. Wang, "Health-ATM: A deep architecture for

multifaceted patient health record representation and risk prediction,"
in Proc. 2018 SIAM Int. Conf. Data Mining, San Diego, CA, pp. 261-
269

693

Authorized licensed use limited to: Panimalar Engineering College - Chennai. Downloaded on November 09,2024 at 07:32:03 UTC from IEEE Xplore. Restrictions apply.