Integrated feature engineering based deep learning model for predicting customer’s review helpfulness

Abstract

In the current market scenario, online customer reviews had a significant impact on boosting the sale of online products. Recently, there has been exponential growth in e-commerce industry owning to the online customer reviews. Over the years, researchers has observed the importance of online consumer reviews for purchasing online products. Hence, in this study, authors made an attempt to develop an efficient convolutional neural network (CNN) based classification model that aims to predict the usefulness of product reviews with higher accuracy on two different types of data sets (i.e., search product and experienced product). In our proposed study, to determine the usefulness of a review in terms of structural, linguistic, sentimental, lexical, and voting feature sets, we build a deep learning model to predict the review helpfulness as a binary classification problem. The performance of the proposed method is evaluated in terms of accuracy, precision, F1 score etc. and had been compared against the various leading machine learning (ML) state of art models viz., K-nearest neighbor (KNN), Linear regression (LR), Gaussian Naive Bays (GNB), Linear Discriminant Analysis (LDA) etc. The results demonstrate that CNN achieved better classification performance in comparison to other state of art models, with highest accuracy of 99.26% and 98.97%, precision of 99% and 99.01%, F1 score of 99% and 99.89%, AUC of 0.9999 and 0.9998, Average Precision (AP) of 0.9999 and 0.9997 and recall of 100% and 100% for two different amazon product datasets.

Keywords

CNN reviews helpfulness online reviews machine learning binary classification reviews feature set

1 Introduction

These days, number of social networking sites are working as independent organizations and helping customers to find better rated products online. Google is one such example. On the basis of large number of reviews and ratings obtained from multiple websites, including pbtech, eBay, Samsung, and others, Google tend to produce a cumulative rating of a product that had remarkable impact on boosting the online sale of product. One such rating for an electronic gadget can be seen in the Fig. 1.

Fig. 1

Example of star rating by Google.

Nowadays, online Products are usually evaluated in term of quality by the online customers before buying it. This is done by reading the online reviews to get assistance in selecting the best product. In current scenario, e-commerce websites try to gauge the value of a product by taking online feedback from the customers whether their product was helpful to them or not. It has been general observation of Consumer psychologists that a helpful product review might significantly affects the consumers view point and decision-making in the online shopping [1, 2]. Current trends in shopping from online stores like Amazon, Flipkart, eBay etc., says that users most likely buy recommended products as compared to the non-recommended ones. It has direct impact on the finances of the company and plays a very vital role in overall business growth. Customer reviews are basically expression of sentiments of a customer after buying any product or service. This can easily be seen in the customer reviews as illustrated the Fig. 2.

Fig. 2

Individual customer review (expression of sentiments) [10].

Millions of customer reviews on online review sites significantly influenced market trends as well as many potential consumers’ shopping decisions [3]. According to [4], it is reported that 86% of customers are in the practice of reading online reviews and around 91% out of them believe in those reviews. In general, customers have found helpful vote’s truly effective while selecting a particular product [7]. Despite of variations in review sites, similar concept of the helpfulness related to online reviews is observed everywhere in the current market scenario [8]. Challenges in front of customers and businesses are to how to control or get the benefit from such reviews which are voluminous, inconsistent and of varied nature. Online reviews, which are available as unstructured big data might have possibly both positive as well as adverse impacts on consumers. Helpfulness is preferred as key feature when several aspects of reviews are taken into account. Review helpfulness is computed as the ratio of the number of helpful votes and the total number of votes [10]. Thereafter, the reviews receiving the highest ratings are reorganized to the top of the web page so that customers can easily check them. Leading online retailers such as Amazon and Trip Advisor— also use this method to measure review helpfulness. Figure 3 demonstrate the procedure of how Amazon.com collected reviews helpful votes from their users [11].

Fig. 3

An illustration of typical Amazon review, Source: Amazon, 2021.

Apart from prediction of helpfulness of user generated review, machine learning (ML) and deep learning (DL) algorithms had played a vital in different domains such as to detect the types of cancer [6 , 28], intrusion detection [42, 43], classification and identification of unstructured road [5], fault detection in insulators [29].Besides these, ML and DL has also been utilized widely to solve the problem of the movement of a robot using path planning algorithms [9] in 2-dimensional search space.

The work flow of the proposed methodology was initially resumed with collection of datasets. To carry out the experiments, we initially collected two set of datasets from website https://jmcauley.ucsd.edu/data/amazon/. The procured datasets were based on experienced and search based product. The data initially being in the raw form underwent the rigorous preprocessing process to make it suitable for extraction of desired features and subsequently to be fed as an input to CNN and other ML algorithms. The subsequent step involves manual extraction of the characteristics from the dataset and categorizing them into five feature sets for reviews viz., structural, lexical, sentimental, linguistic, and voting. In next step, the dataset is divided into two halves that is utilized for training purposes (80%) and validation purposes (20%). 10 cross validation was used to assess the system performance. The training dataset was used to create five ML prediction models, including, K-nearest neighbor (KNN), Linear regression, second (LR), Gaussian-Naive Bays (GNB), linear discriminant analysis (LDA) and Convolutional neural networks (CNN). Finally, the model performance is evaluated based on accuracy, precision, recall, F1, AUC (area under curve), and AP (average precision).

The rest of the article is organized as follows: Section 2 describes Literature review, Section 3 presents the proposed methodology that is used in this paper for helpfulness prediction of online reviews. The section 4 describe the architecture of the CNN based classification model, while in section 5, we demonstrated and discussed experimental results. Section 6 discuss the conclusion and future scope of work.

2 Literature review

Online product reviews attract interest from several academic disciplines, marketing is the main amongst them. Online product reviews are usually considered as one of the most valuable and promising tool to promote products. In addition, it also supports the collection of consumer feedback and boosting of product’s sale [12 –14]. For example, online movie and book reviews might have a major impact on box office revenues [15], as well as on book sales [16]. However, online reviews gets impacted significantly by category, location, and other factors [17, 18] and hence heavily relies on these factors.

To forecast the significance of product reviews, authors in [19] developed a regression model. The authors used part-of-speech (POS)-based syntactic terms and lexical similarity as characteristics in these experiments [20]. Besides this, lexical subjectivity was also used as feature.

Review helpfulness assist in consumers’ information search process, and also perform other task like converting readers to buyers to increases the business [21, 22]. Over the past 20 years, the usefulness of reviews has been assessed using the star rating, reviewer’s credibility, price and categories of the products. [17, 23]. In this regard, numerous ANN based multilayer perceptron model has been developed by authors [24, 25] using product, review information, and review characteristics as common attributes. The prime objective of their work was to improve helpfulness prediction by employing neural network models instead of linear regression approaches [26].

Ref. [28] proposed a methodology for predicting the usefulness of travel-related websites. For the purpose of predicting helpfulness, they have combined reviewer and review features. More specifically, the authors developed a text regression model employing features including reviewer identity, reputation, competence, review valence, and readability in order to assess how helpful a review will be. They have used the reviews’ writing style, timeliness, and reviewer experience as features for the prediction the helpfulness of customer generated reviews. In [32], a new approach to calculating how useful reviews are, is provided. Instead of using a conventional review helpfulness problem, the authors calculate helpfulness using the confidence interval and the helpfulness distribution data. The key conclusions of many cutting-edge approaches about the worth of reviews are compiled in Table 1.

Table 1
A summary of the recent works provided in the literature

References Data source Problem statement Contribution

[2] Amazon.com Examined the usefulness of product reviews with their source and content-depended review features. Product reviews generated by customer were found more helpful in comparison to those done by experts. A customer-written product based review with a small volume of content abstractness produced the highest score for review usefulness.

[17] Amazon.com Utilized a linear regression model based on the information economics paradigm to forecast how useful experienced products will be. The perceived review helpfulness is often influenced by their depth, extremity, and product category. Review depth has a positive impact on how useful a review is, however for search or experience products, the type of product alters the review effect on the helpfulness.

[20] Amazon.com Customer generated reviews helpful prediction. CNN deep learning model as a powerful predictive tool for the classification of helpful-unhelpful reviews based on manually extracted features set on reviews text.

[22] CNET Increasing the number of helpfulness votes is the major goal of this work. The number of helpfulness votes reviewers shall receive depends more strongly on semantic characteristics compared to other aspects.

[26] Amazon.com Developed a prediction model to investigate the factors affecting the impact of online reviews. Verb, adjective, and other language category elements are thought to have a greater influence on experienced and search items.

[35] Amazon.com Reviews helpfulness prediction using classification with random forests algorithm. Reviews having blended objective and highly subjective sentences were more helpful in rating.

[51] Yelp Examined the features of high importance for review helpfulness prediction using regression and classification models. SNS features are highly correlated with reviewer popularity for predicting the online product reviews “Helpfulness”.

[55] TripAdvisor Developed a conceptual model for helpfulness using quantitative and qualitative factors. Types and concepts related to review have a varying degree of impact on review helpfulness. Review length has a negative impact on the helpfulness. Too-long reviews are less helpful and fails to attract the reader’s observation.

[56] Amazon.com The study examined different factors affecting online review helpfulness. In case of online product reviews, organizational structure, language style as well as content features influenced the reviews helpfulness.

[57] Amazon.com Building a conceptual model based on central cues and peripheral cues for reviews helpfulness. Rating, integrity as well as content of reviews affect the reviews helpfulness.

References	Data source	Problem statement	Contribution
[2]	Amazon.com	Examined the usefulness of product reviews with their source and content-depended review features.	Product reviews generated by customer were found more helpful in comparison to those done by experts. A customer-written product based review with a small volume of content abstractness produced the highest score for review usefulness.
[17]	Amazon.com	Utilized a linear regression model based on the information economics paradigm to forecast how useful experienced products will be.	The perceived review helpfulness is often influenced by their depth, extremity, and product category. Review depth has a positive impact on how useful a review is, however for search or experience products, the type of product alters the review effect on the helpfulness.
[20]	Amazon.com	Customer generated reviews helpful prediction.	CNN deep learning model as a powerful predictive tool for the classification of helpful-unhelpful reviews based on manually extracted features set on reviews text.
[22]	CNET	Increasing the number of helpfulness votes is the major goal of this work.	The number of helpfulness votes reviewers shall receive depends more strongly on semantic characteristics compared to other aspects.
[26]	Amazon.com	Developed a prediction model to investigate the factors affecting the impact of online reviews.	Verb, adjective, and other language category elements are thought to have a greater influence on experienced and search items.
[35]	Amazon.com	Reviews helpfulness prediction using classification with random forests algorithm.	Reviews having blended objective and highly subjective sentences were more helpful in rating.
[51]	Yelp	Examined the features of high importance for review helpfulness prediction using regression and classification models.	SNS features are highly correlated with reviewer popularity for predicting the online product reviews “Helpfulness”.
[55]	TripAdvisor	Developed a conceptual model for helpfulness using quantitative and qualitative factors.	Types and concepts related to review have a varying degree of impact on review helpfulness. Review length has a negative impact on the helpfulness. Too-long reviews are less helpful and fails to attract the reader’s observation.
[56]	Amazon.com	The study examined different factors affecting online review helpfulness.	In case of online product reviews, organizational structure, language style as well as content features influenced the reviews helpfulness.
[57]	Amazon.com	Building a conceptual model based on central cues and peripheral cues for reviews helpfulness.	Rating, integrity as well as content of reviews affect the reviews helpfulness.

In Ref. [30] authors suggested a regression model that uses several variables based on semantic, structural, and meta-data to predict how useful online reviews will be. Along with other variables like review expertise, writing style, and timeliness, aspects like review feelings (valence), length, and unigrams are also recognized as significant predictors [33].

A unique suggestion about reviewer engagement (RFM) features was made in Ref. [34] in order to improve the effectiveness of the helpfulness prediction system. The authors stated that a hybrid model was made up of RFM and textual elements produced more accurate predictions. Subjectivity, readability, and meta-data aspects [30 , 35], are also some important features that were excluded despite of being superior than other predictors. In Ref. [36], authors asserted a connection between valence consistency and review usefulness to track their impact on the outcome. As per study [37], in order to perform the review helpfulness, product reviews were divided into low, medium, high, and spam categories. Thereafter, a machine learning model was created utilizing the numerous parameters like readability, subjectivity, and information In order to categories poor and high quality reviews, Ref. [38] explored regression model to understand the significance of textual and non-textual elements in prediction of reviews helpfulness. As per the study, items were divided into two groups: experience-based products and search-based products. Extreme ratings of experiencing products have been found to be less helpful than moderate one.

In contrast to experience products, many evaluations of search products were more beneficial than moderate reviews. The review length is another important feature that has high impact on the prediction of reviews helpfulness, however, its influence rely upon the product category. For instance, in case of search product, review length affected the prediction result more positively as compared to experience products. The authors came to the conclusion that other aspects, such as star rating and review length, also have a higher influence on product categories based on the aforementioned findings.

In Ref., (Iwendi et al., 2022) [31] an item recommendation system based on pointer-based scheme was developed in the study as a modern machine learning attention model. All user reviews and comments are taken into consideration when recommending an item. With a Mean Absolute Error of 21%, 79% Precision, 80% Recall, and 79% F1-Score, the model was able to achieve 79% Accuracy on the Yelp dataset.

Consumer sentiments, emotion features and review helpfulness: Emotion feature is referred to as an assessment of a change in an individual’s feelings [39]. There are different kinds of emotion that can be classified as intrapersonal or interpersonal. The intrapersonal dimension of emotions include an individual’s internal affective experience; whereas the interpersonal dimension point out an external identifiable expression of emotions during communication between peoples [40].

A recurrent convolutional neural network (RCNN) model was used by [45] to estimate the reviews’ helpfulness, and they came to the conclusion that semi-supervised learning performed better than supervised learning. In Ref. [46], Chau et al. proposed a model for retrieving useful reviews using the elastic net regularization and multiple linear regression algorithm based on reviews ranked lists. Furthermore, [47] Describes the development of a machine learning model to predict the usefulness of user evaluations based on the polarity, subjectivity, entropy, and readability textual properties of reviews. When compared to other machine learning-oriented methodologies, the Valence Aware Dictionary and Sentiment Reasoner (VADER) model [48] performs better for individual human raters in terms of F1 score and accuracy. Based on user-submitted evaluations, authors of study [49] developed a feature extraction approach to quantify and assess the helpfulness for each product. A methodology called Helpful Quality-related Review Mining (HQRM) was proposed by Jiang et al. in their study [50] to aid the multi-class classification problem pertaining to prediction of useful quality-related reviews. A descriptive and comparative analysis-based methodology was proposed in research [51] as a way to look into and report on changes in user preferences, corporate reputation, and rating behavior. Based on a prior study, authors of [52] proposed that the kind of product modifies the impact of emotions on perceived review usefulness.

In Ref. [11] authors used Yelp dataset’s review word count as a structural feature category to predict customer reviews’ helpfulness using B-GBT (Bagging Gradient-Boosted Trees) architecture, and they achieved accuracy of 95.3% with the highest area under curve (AUC) of 0.988. Based on the review rating and word count in structural features categories, authors in Ref. [12] achieved the confidence interval of 95% using SVR on Amazon dataset. Using Random Forest algorithm, authors [26] obtained the satisfactory performance with an accuracy of 81.33 percent based on the linguistic features viz., adjective and verb. In [38], authors employed DNN model on the Amazon dataset and achieved the maximum accuracy of 84.54 percent, on the search product by utilizing a hybrid features set (i.e. readability feature, sentimental feature, and linguistic features). For the classification of the review helpfulness prediction problem, Ref. [46] used a multiple linear regression model with the elastic net regularization method on the Amazon dataset to predict the outcome. The developed model achieved an accuracy of 83 percent using review length, character count, polarity feature, helpfulness vote feature, and readability features. In Ref. [47] the Gradient Boosting approach was developed based on ensemble learning by using sentimental feature (Polarity), linguistic feature (Noun, Adjective, Verb as), lexical feature (Flesch reading edge, Dale Chall RE), structural feature (Review length, Sentence Length, one letter words, two letter words, longer letter words, review rating), and voting feature (Helpful Ratio) to forecast the best mean square error (MSE) when the number of ensemble trees is 100 on the Amazon dataset. In Ref. [49], authors combined lexical (Flesch reading ease), structural (length of review, number of sentences, and character count), and semantic features to create hybrid features that were applied to the Amazon dataset and achieved an accuracy of 77.6 percent by decision tree classifier for classification of review helpfulness. A summary of current studies on how helpful reviews might be classified is shown in Table 2.

Table 2

An overview of recent studies on the classification of reviews’ helpfulness

Reference	Data sources	Helpfulness	Feature Prob. Type					ML algorithms	Predicted result
			RC	P	R	C	REG
[15]	TripAdvisor	ratio	√	×	√	√	√	LGR, DT, RandF, SVM	Accuracy, F1 score, AUC
[20]	Amazon.com	ratio	√	×	×	√	×	KNN, LR, GNB, CNN	Accuracy, Precision, Recall, F1 score, AUC
[25]	Amazon.com	ratio	√	√	√	×	√	LNR, MLP	MSE
[26]	Amazon.com		×	×		√	×	NB, SVM, RandF	Accuracy, F1 score
[35]	Amazon.com	ratio	√	√	√	√	×	RandF, SVM	Accuracy, AUC
[38]	Amazon.com	ratio	√	×	×	√	×	NB, RDA SVM, RandF, Pruned C4.5, DNN	Accuracy, F1 score
[46]	Amazon.com	vote	√	×	×	√	×	NB, SVM, RandF	Accuracy, NDCG
[47]	Amazon.com	ratio	√	×	×	×	√	GBT	MSE
[51]	Yelp	ratio	√	√	√	√	√	DT, RandF, GBT, N.Net	RMSE, MAE, RAE, RRSE, MSE, CC, R²
[52]	Amazon.com	ratio	√	×	×	×	√	LNR, SVR, M5P, RandF	MAE
[53]	Amazon.com	ratio	√	×	√	×	√	CART, MLP, MARS,GLM, Ensamble	RMSE, MSE, RAE, RRSE, MAE
[54]	Amazon.com	ratio	√	√	√	×	√	MLP-BP, MARS, GLM, CART	MSE, RAE, RMSE, RRSE, MAE

RC (Review Content), P (Product), R (Reviewer), ML (Machine Learning),C (Classification), REG (Regression), GBT((Gradient Boosted Tree), LNR (Linear Regression), RandF (Random Forest), SVR (Support Vector Regressor), M5P (M5 Model Trees), CART (Classification and Regression Tree), MARS (Multivariate Adaptive Regression Splines), GLM (Generalized Linear Model), GBT (Gradient-Boosted Trees), N.Net (Neutral Network), MLP (Multilayer Perceptron), MLP-BP(Multilayer Perceptron with back propagation), DNN (Deep Neural Network), NDCG (Normalized Discounted Cumulative Gain), LGR (Graphical model for Local Gaussian Regression), MAE (Mean Absolute Error), RMSE (Root Mean Square Error), MSE (Mean Squared Error), RAE (Relative Absolute Error), RSE (Relative Squared Error), RRSE (Root Relative Squared Error), R² (R-squared), CC (Correlation Coefficient), KNN (K-nearest neighbor), LR (Linear Regression), GNB (Gaussian Naïve Bays), CNN(Convolution Neural Network).

In our study, we integrated various types of features to form five feature vectors. These features are taken from different studies [17 , 49–52] based on their performance. In this study, we have selected those features only, which have been proven effective by the different authors based on the outcomes obtained in their previous studies. As of now, no study has been carried out on these features collectively as per the author’s knowledge, which motivated us to integrate and investigate these features (i.e., lexical [38], structural [46], linguistic [47], sentimental [17], and voting [26]) to build an effective classier using CNN for prediction of useful and unhelpful reviews.

The performance of suggested CNN model is compared against different current state-of-the-art models, including LR, KNN, GNB, etc. In the developed model, integrated features inherited the inherent quality of individual features like structural [10 , 52], lexical [38 , 49], linguistic [26 , 50], sentimental [38 , 51], and voting [11 , 52], thereby enhancing the overall performance of the suggested model.

In this study, authors proposed a DNN model that was customized for an effective prediction of customer reviews based on the characteristics of customer evaluations. The major contributions of the proposed study are as follows:

In this study, two different amazon datasets (i.e., one is based on search goods and other is based on experienced goods) were utilized to extract distinct features related to goods reviews and divided into five feature sets viz., voting, structural, lexical, sentimental, and linguistic features.

The features extracted in step 1 were utilized as an input to train the DNN model, where the underlying model was customized and tuned to build an efficient and reliable prediction model for predicting the usefulness of customer generated review of a particular product.

In the end, our suggested model’s performance was compared to other current-generation models in terms of sensitivity, precision, recall, F1 score, and area under curve (AUC).

3 Research methodology

In this section, we primarily focus on formulation and development of classification model that aims to predict review helpfulness of online products. The review helpfulness prediction model was built using five feature sets, wherein each feature set had varying number of parameters possessing different characteristics as illustrated in Fig. 5. In our study, we have utilized the most recent baseline attributes [5 , 49–52] to create an effective CNN based prediction model.

Fig. 5

Sample example of Review dataset (source amazon.com).

The framework proposed in this study comprises of phases, which is illustrated in Fig. 4.

Fig. 4

Framework of research methodology.

In the first phase, appropriate datasets were collected, followed by implantation of preprocessing step. The data initially being in the raw form underwent the rigorous preprocessing process to make it suitable for extraction of desired features and subsequently to be fed as an input to CNN and other ML algorithms. The subsequent step involves manual extraction of the characteristics from the dataset and categorizing them into five feature sets for reviews viz., structural, lexical, sentimental, linguistic, and voting. In next step, the dataset is divided into two halves that is utilized for training purposes (80%) and validation purposes (20%). 10 cross validation was used to assess the system performance. The training dataset was used to create five ML prediction models, including, K-nearest neighbor (KNN), Linear regression, second (LR), Gaussian-Naive Bays (GNB), linear discriminant analysis (LDA) and Convolutional neural networks (CNN). Finally, the model performance was evaluated based on accuracy, precision, recall, F1, AUC (area under curve), and AP (average precision).

The review information that has been utilized in our experiment has been taken from Amazon.com, which is freely available data source. This data source is made publically available especially to perform the experiment by researchers and scientific community. This dataset has three primary components such as reviews corpus, social information about those reviews, and reviewer data.

3.1 Data collection and preprocessing

To conduct the experiments, we procured the data from Amazon.com. The data has sample helpfulness reviews collected between May 1996 and July 2014 [20, 44].

The data contains the review information about the two product categories, home and kitchen goods and cellphones and accessories. The former product category occupies the memory space of 138 MB, while later occupies the memory space of 411.5 MB, respectively. There are numerous items such as Phones and accessories, which are considered as a part of search goods category, while home and kitchen items are designated as experience products, making it challenging and expensive to learn about the quality of a product before using it [11].

Table 3 lists the various types of products along with the total number of reviews, goods, and reviewers (users) for each category. In dataset, there were nine column representing nine attributes such as reviewerID, asin, reviewerName, helpful, review Text, overall, summary, unixReviewTime, and reviewTime. The details of each attribute is explained as follows:

reviewerID: it represents the alphanumeric code, which is given by Amazon to their reviewers

asin: it is amazon standard identification Number(asin) that referred to as the alphanumeric product ID assigned to the products by Amazon.

reviewerName: It demonstrates the name of reviewer that is present in the list of Amazon product review database.

helpful: It indicates those cases of users for whom review was helpful.

reviewText: it refers to the review contents submitted by the reviewer’s regarding a particular product.

overall: it gives information about the star rating value of any product.

summary: it provides the review’s summary of any goods which is available on website for sale.

unixReviewTime: it refers as Unix time, when the review was published.

reviewTime: it refers to the date and time, when the review was created.

Table 3
Description of the dataset for Amazon reviews

Collection of Amazon reviews

Dataset Product type #reviews #products #reviewer Product categories

DS1 Cellphone and accessories 194,439 10,429 27879 Search goods

DS2 Home and Kitchen 551,682 28237 66519 Experience goods

Collection of Amazon reviews
DS1	Cellphone and accessories	194,439	10,429	27879	Search goods
DS2	Home and Kitchen	551,682	28237	66519	Experience goods

An example Amazon.com review is given in Fig. 5.

3.2 Preprocessing

DS1 and DS2 are two real-world review datasets, which have been utilized in this study to evaluate the performance of proposed methodology and other ML models under study for prediction of review helpfulness. A data cleaning process was used to eliminate unnecessary data with the aim of improving the qualitative features of the dataset. The applied cleaning method substantially improves the performance of the suggested mode [20, 38].

The various steps data cleaning approach can be summarized as follows:

The first step plays very vital role in cleaning the data. It includes the process of identifying and eliminating the duplicate reviews from the datasets.

In the second step, the redundant information such as blank text is eliminated from the dataset.

In third step, Reviews with a large number of votes are taken out, which often produces better categorization outcomes.

Therefore, only evaluations with minimum 10 total votes are included. The number of reviews present in both datasets before preprocessing and after preprocessing are listed in Table 4.

Table 4
Summary about the dataset1 (DS1) and dataset2 (DS2)

Dataset Number of reviews

Before processing After preprocessing

DS1 3,447,249 8775

DS2 4,253,926 59515

Dataset	Number of reviews
DS1	3,447,249	8775
DS2	4,253,926	59515

3.3 Feature set extraction

This section primarily focus on classification problem and extraction of feature/characteristic set that is required to build an effective model. Let’s assume R = { αⁿ, βⁿ} where αⁿ = { α_1, α_{2, … … . .} α_n } depicts the n-numbers of review features, while βⁿ indicates the levels of helpfulness. βⁿ ∈{0, 1} where ‘0’ and ‘1’ represents the review helpfulness and unhelpfulness, respectively. As we know that features plays crucial role in determining the performance of any classification model. Hence, feature set ought to be chosen wisely in order to ensure the better classification results. In this study, k number of features are used, where k ∈+I (universal set of positive integer number). Here each feature vector is represented by x_i that is comprised of five disjoint feature set [S_t, L_e, S_e, L_i, V, S_t, L_e, S_e, L_i, V] represents the structural, lexical, Sentimental, linguistic and voting feature set reviews respectively. Feature matrix, αⁿ is generated by integrating the various feature sets.

In this experiment, we have used 22 reviews features, which have been derived from various studies based on their characteristics [26 , 46]. These features have been grouped into different five disjoint feature sets based on their compatibility to each other. These feature can be well understood in the subsequent sections.

Table 5
List of feature set used in previous studies

Type Feature name Description of features Remark

Structural feature [SF] [10 , 52] Char_Len Denote the character count in the review text Character count

nWord Denote the total word count the review text Word count

nSent Refers to total sentence count in the review text Sentence count

WPS Mean value of word count in a sentence # of word / # of sentence

DFW Denote total difficult word count in the review text

UW Represent unique word count

Rating Indicate rating of any specific product Range [1, 5]

Avg_W_Len Represent mean word length in the review process

1Word Demote how many ties one length word appears in the review text One length word

2Word Indicate how many times two length word appears in the review text Two length word

Long_Word Indicate the frequency of long length word in the review text Long length word

Lexical feature [38 , 49] Fre Illustrate Flesch_reading_ease readability

Dcri Denote Dale_chall_readability

Sentimental feature [38 , 51] Neg Denote Negative polarity score of review text

Pos Refers to Positive polarity score of review text

Neu Refers to neutral polarity score of review text

Compound Represent summation of neg, pos and neu score Normalized between [–1, 1]

Linguistic Feature [26 , 50] Noun Indicate the frequency of noun word in the review text Noun count

Adj Indicate frequency of adjective word in the review text Adjective count

Verb Number of times verb word appears in the review text Verb count

Voting feature [46, 47] Vote Number of times helpful vote appears in review text

Total Total number of times vote word appears in review text

Helpfulness Represent the fraction of helpful vote to the total number of vote (#Helpful / #Total)*100

Type	Feature name	Description of features	Remark
Structural feature [SF] [10 , 52]	Char_Len	Denote the character count in the review text	Character count
	nWord	Denote the total word count the review text	Word count
	nSent	Refers to total sentence count in the review text	Sentence count
	WPS	Mean value of word count in a sentence	# of word / # of sentence
	DFW	Denote total difficult word count in the review text
	UW	Represent unique word count
	Rating	Indicate rating of any specific product	Range [1, 5]
	Avg_W_Len	Represent mean word length in the review process
	1Word	Demote how many ties one length word appears in the review text	One length word
	2Word	Indicate how many times two length word appears in the review text	Two length word
	Long_Word	Indicate the frequency of long length word in the review text	Long length word
Lexical feature [38 , 49]	Fre	Illustrate Flesch_reading_ease readability
	Dcri	Denote Dale_chall_readability
Sentimental feature [38 , 51]	Neg	Denote Negative polarity score of review text
	Pos	Refers to Positive polarity score of review text
	Neu	Refers to neutral polarity score of review text
	Compound	Represent summation of neg, pos and neu score	Normalized between [–1, 1]
Linguistic Feature [26 , 50]	Noun	Indicate the frequency of noun word in the review text	Noun count
	Adj	Indicate frequency of adjective word in the review text	Adjective count
	Verb	Number of times verb word appears in the review text	Verb count
Voting feature [46, 47]	Vote	Number of times helpful vote appears in review text
	Total	Total number of times vote word appears in review text
	Helpfulness	Represent the fraction of helpful vote to the total number of vote	(#Helpful / #Total)*100

3.3.1 Structural feature set (S_t)

This feature set comprises of text reviews, which possess 11 structural characteristics to define it. In the review text specific words, phrases, and paragraphs are specified by structural components of structural feature set. Structural feature set (S_t): {Char_Len, nWord, nSent, WPS, DFW, UW, Rating, Avg_W_Len, 1Word, 2Word, Long_Word} are used in our study. Each element of the feature are defined as follows:

Char_Len: it alludes to the total amount of characters in the review’s content.

nWord: it alludes to the total amount of words in a single review’s content.

nSen: it alludes to the total amount of sentences in a review’s content.

WPS: it represents the average numbers of word in a sentence.

DFW: it represents the challenging words, which are part of review & is beyond the list of 3000 well-known words.

UW: It represents the non-similar words in review.

Avg_W_Len: it denotes the average number of character in a word of review content.

1Word: it alludes to the total amount of words in the review content that are of one length.

2Word: it represents the number of word, which have two length in review text conducted by company.

Long_Word: it gives information about the presence of total number of word in text reivew, which have greater than two length.

3.3.2 Lexical feature set (L_e)

Another crucial aspect of review content that might affect how beneficial a review is its readability. The readability of a review content tends to be directly correlated with the lexical feature set. The number of user votes for helpfulness might rise when a review is simple to read. In this study, we employed two readability measures that have been identified in the literature to calculate review readability [26, 47]. Flesch Reading Ease readability score (FRE) and Dale Chall readability score (DCR) are two such examples. ${Lexical feature set (L}_{e}) = {FRE, DCR}$

One of the most common and well-known readability metrics is the Flesch Reading Ease (FRE) score. The following is the formula for the FRE score [46]: $• FRE = 206 . 835 - (1 . 015 * A) - (84 . 6 * B)$ (1)

Where,

A = Average word length of a sentence

B = Average syllables per word

Dale_Chall_readability score is another readability formula used in this research

The formula of Dale-Chall Readability is given as [47]: $• DCRI = 0 . 1579 * C + 0 . 0496 * A$ (2)

Where,

DCRI = Dale Chall Readability Index

A = Average word length of a sentence

C = Percentage of difficult words in a review text

3.3.3 Sentimental feature set (S_e)

In this work, sentimental feature set (S_e) comprises of four review features viz., {Neg, Pos, Neu, Compound}

Neg: Negative feature is assessed in the range of [–4,0].

Pos: Positive feature is evaluated in the range of [0,4].

Neu: Neutral sentiment is indicated ‘0’.

Compound: It represents the normalized score of the sum of Neg, Pos and Neu that ranges between –1 to+1.

While performing the experiment, we performed vocabulary- and rule-based sentiment analysis using the Valence Aware Dictionary and Sentiment Reasoner (VADER) tool in order to generate of the polarity ratings in the form of numerical representations. [20, 48] using python library. Here VADER transformed the sentiment scores of lexical features into emotion scores.

3.3.4 Linguistic feature set (L_i)

Linguistic feature set (L_i) is comprised of three linguistic features, which are represented as follows [26, 47]: $Linguistic feature set (L_{i}) = {Noun, Verb, Adj}$

After the individual linguistic feature {Noun, Verb, Adj} extracted for L_i set, each linguistic feature was assigned a score. The assigned score is the frequency of occurrence of individual linguistic feature in the sentence of a review text. The steps involved in computing the frequency of three linguistic characteristics for each and every review in the review dataset are described in algorithm 1. Computational cost of the proposed model is O (N*M), which shows its efficiency (where N represents total number of user generated text reviews, M denotes the total number of words in individual reviews). To simulate our experiments, we used hardware with the following specifications: 8GB RAM, Processor Intel P-5. On our system, the proposed algorithm took an average time of 33.05 and 251.06 seconds for datasets DS1 and DS2 respectively in 10 epochs.

Table 6
Pseudo Code of the scoring of linguistic feature set

Algorithm 1: Linguistic feature set score

Define:R = { α₁, α₂, . . . . . . . . . . A_n}: The set of text reviews in dataset.

POS: Part of speech.

Input: L_i [0, 0, 0]: Initial score vector of linguistic feature set [Noun, Adj, Verb].

WC = 0: Initial value of word count.

Output: Final score of linguistic feature set L_i.

Function Linguistic Score ( L_i)

Apply NLTK to compute POS of each review α_I in R

For each review in R

For each sentence S_j in review α_i do

For each word w_k in sentence s_j do

WC ← WC+1

If w_k belongs to Noun

Then increment L_i [Noun++, 0, 0]

End if

El-if w_k belongs to Adjective

Then increment L_i [0, Adj++, 0]

End if

El-if w_k belongs to Verb

Then increment L_i [0, 0, Verb++]

End if

End for

End for

For each Noun, Adjective and Verb in L_i score vector

Compute percentage w.r.t. WC

End for

Compute Z-score of L_i score vector of every review

Algorithm 1: Linguistic feature set score
Define:R = { α₁, α₂, . . . . . . . . . . A_n}: The set of text reviews in dataset.
POS: Part of speech.
Input: L_i [0, 0, 0]: Initial score vector of linguistic feature set [Noun, Adj, Verb].
WC = 0: Initial value of word count.
Output: Final score of linguistic feature set L_i.
Function Linguistic Score ( L_i)
Apply NLTK to compute POS of each review α_I in R
For each review in R
For each sentence S_j in review α_i do
For each word w_k in sentence s_j do
WC ← WC+1
If w_k belongs to Noun
Then increment L_i [Noun++, 0, 0]
End if
El-if w_k belongs to Adjective
Then increment L_i [0, Adj++, 0]
End if
El-if w_k belongs to Verb
Then increment L_i [0, 0, Verb++]
End if
End for
End for
For each Noun, Adjective and Verb in L_i score vector
Compute percentage w.r.t. WC
End for
Compute Z-score of L_i score vector of every review

3.3.5 Voting feature set (V)

Voting feature set (V) comprises of three features that is mentioned as follows:

Total: It represents all the votes that have been received.

Vote: It represents all the votes, which are treated helpful.

Helpfulness: It denote the fraction of the helpful votes and the total number of votes in [0, 1].

In literature [10 , 37], the review helpfulness is categorized into two classes either as a ‘1’ or ‘0’ based some threshold value (Φ) as follows: $Hi = {\begin{matrix} 1 (helpful), & if helpful ratio \geq Φ \\ 0 (unhelpful), & otherwise \end{matrix}$ (3)

Here H_i represents the helpfulness of a review I. Helpfulness threshold (Φ) is considered as important parameter on which classification result relies a lot. In our study, its value is set to 0.60 based on the experience of previous research studies [35 , 41].

Both the datasets (DS1, DS2) in Fig. 7 shows helpfulness scores (ratio) distribution (x-axis) versus frequency (y-axis). Helpfulness ratio (score) varies in the range of [0, 1].

In the Fig. 6, large amount of reviews in both datasets are distributed in the range of [0.9, 1], which point out to the density of helpfulness ratio lopsided towards the right.

Fig. 6

Helpfulness score distribution in DS1, DS2 datasets.

Fig. 7

Structure of CNN model.

4 Building the classification model

To build the CNN classifier, feature matrix of size N x F were used, where N & F denote the total number of reviews and total number of features used to train the model. Here, we set the value of F as 22. In this study, well-known machine learning algorithms such as Linear regression (LR), K-nearest neighbor (KNN), Gaussian Naive Bays (GNB), Linear Discriminant Analysis (LDA) were applied to classify the product reviews into helpful or unhelpful cases based on innovative feature sets. To build the classifier mode, both datasets (i.e., DS1 and DS2) were split into the ratio of 80 : 20 percent; wherein 80 percent of data was used for training purpose and remaining 20 percent was used to test the performance of the model. The proposed model was simulated using Python 3.7 and computed various performance measures like Precision, Recall, Accuracy, F1 Score, AP and AUC are some of the assess the performance of the developed system. The performance measures can mathematically be defined as [38 , 42]

Precision = \frac{Tp}{Tp + Fp}

(4)

Recall = \frac{Tp}{Tp + Fn}

(5)

Accuracy = \frac{Tp + Fp}{Tp + Tn + Fp + Fn}

(6)

Where:

T_p: True positive rate, T_n: True negative rate, F_p: False positive rate, F_n: False negative rate $F 1 = 2 * \frac{Precision * Recall}{Precision + Recall}$ (7) $Area under curve (AUC) = \int_{b}^{a} f (x) dx$ (8) $\begin{matrix} Average Precision (NP) = \sum_{k = 0}^{k = n - 1} ([Recall (k) \\ - Recall (K + 1)] * Precision (k)) [42] \end{matrix}$ (9)

Where, Recall (n)=0, Precision (n)=1 and n is number of threshold

4.1 Convolutional neural network architecture

This section focus on the development of CNN based deep neural network model that aims to classify review data into binary category as helpfulness and non-helpfulness. Proposed model was constructed using Keras library in addition to Tensorflow which were served as a backend.

Figure 7 illustrate the basic structure of CNN model. Figure 8 shows that the proposed CNN model. The proposed model comprises of two 1-dimensional convolution layers (Conv1D). There are several parts to each Conv1D, including a pooling layer (Pool), batch normalization, and a rectified linear unit (ReLU) activation function. In addition, the developed model also comprise drop out layer, sigmoid function, fully connected (fc) layer, and binary classification output layers. In the output layer, we used binary cross entropy as the loss function. This model has 22 input neurons at the input layer in which feature information was passed sequentially. As stated above, feature vector matrix has the dimensions of N * 22 that represents five feature sets. Feature information from first convolution layer was passed on to second convolutional layer (Conv1D 2), which had 64 neurons. Single neuron was used to display the output as 0 or 1 at the output layer. 1 indicated the binary class as while 0 indicated the class as unhelpful.

Fig. 8

Predictive model performance in percentage (%) of ML algorithms for DS1 and DS2.

The various tuning parameters, which we used during building the model are listed out in Table 7.

Table 7

Turing parameters of proposed model

S.No.	Parameters	Value
01	Optimizer(OPT)	Adam
02	Learning rate(LR)	0.0004
03	Beta_1(β1)	0.9
04	Beta_2(β2)	0.998
05	Epsilon(€)	0.00000001
06	Epochs	10
07	Loss function(LF)	Binary Cross
		entropy
08	Verbose(V)	1
09	Batch Size(BS)	15

5 Experimental results

This work illustrate the results of various experiments conducted with five machine learning algorithms using five sets of parameters. Subsequently the performance of all the algorithms understudy work evaluated and compared using the various performance matrices such as precision, recall, accuracy, AUC etc.

5.1 ML model performance evaluation

In this section, the performance analysis of various machine learning models for predicting reviews helpfulness is presented. The performance of all the underlying models (i.e., LR, LDA, KNN, GNB, and CNN) was validated using10-fold cross validation. In this experiment, using two datasets (DS1 and DS2), the significance and usefulness of suggested feature sets are also examined in terms of several performance metrics, including precision, recall, accuracy, F1 score, etc. Table 8 displays the performance of the predictive model using 10-fold cross-validation.

From the results given in Table 8, it can be observed that CNN method obtained better results than those of other competitive ML models for both the datasets. The proposed model obtained precision of 99.0% and 99.01% and accuracy of 99.00% and 98.97% for DS1 and for DS2, respectively. Similarly others parameters obtained highest result in proposed model in terms of recall, F1, AUC and AP i.e. 100%, 99%, 0.9999 and 0.9999 for DS1 and 100%, 99.89%, 0.9998 and 0.9997 for DS2 respectively. The results given in Table 7, demonstrate the superiority of CNN model with highest performance in terms of accuracy (98.99 for DS1, 78.67 for DS2), recall (100% for DS1, 100% for DS2), F1score (99% for DS1, 99.89% for DS2), AUC (0.9999 for DS1, 0.9998 for DS2) and AP (0.9999 for DS1, 0.9997 for DS2).

Table 8
Predictive model performance using 10-fold cross-validation

Dataset Model Precision (%) Recall (%) Accuracy (%) F1 (%) AUC AP

DS1 (Search goods) KNN 97.00 100 96.81 98.00 0.9961 0.9928

LR 87.00 100 87.41 93.00 0.9912 0.9912

LDA 97.00 99.00 96.92 98.00 0.9986 0.9986

GNB 98.00 68.00 70.83 80.00 0.9781 0.9769

CNN (Proposed model) 99.00 100 99.26 99.00 0.9999 0.9999

DS2 (Experience goods) KNN 98.00 100 98.47 99.00 0.9987 0.9974

LR 92.00 100 92.59 96.00 0.9990 0.9990

LDA 98.00 99.00 97.60 99.00 0.9994 0.9994

GNB 99.00 77.00 78.99 87.00 0.9879 0.9875

CNN (Proposed model) 99.01 100 98.97 99.89 0.9998 0.9997

Dataset	Model	Precision (%)	Recall (%)	Accuracy (%)	F1 (%)	AUC	AP
DS1 (Search goods)	KNN	97.00	100	96.81	98.00	0.9961	0.9928
	LR	87.00	100	87.41	93.00	0.9912	0.9912
	LDA	97.00	99.00	96.92	98.00	0.9986	0.9986
	GNB	98.00	68.00	70.83	80.00	0.9781	0.9769
	CNN (Proposed model)	99.00	100	99.26	99.00	0.9999	0.9999
DS2 (Experience goods)	KNN	98.00	100	98.47	99.00	0.9987	0.9974
	LR	92.00	100	92.59	96.00	0.9990	0.9990
	LDA	98.00	99.00	97.60	99.00	0.9994	0.9994
	GNB	99.00	77.00	78.99	87.00	0.9879	0.9875
	CNN (Proposed model)	99.01	100	98.97	99.89	0.9998	0.9997

The performance of the Gaussian Naive Bays (GNB) classification model was the worst of all the models analyzed. (i.e., 70.83% 78.99% for DS1 and Ds2). Figure 8 shows comparative analysis of KNN, LR, LDA, GNB, and CNN models for both the datasets in terms of bar chart. In proposed model, dataset DS2 delivers relatively better performance than dataset DS1 in terms of precision and F1. The tests’ findings overwhelmingly support the usefulness of the proposed features for the review helpfulness prediction in terms of precision, recall, accuracy, f-measure, AUC and AP metrics.

From the above discussion, it can be concluded that CNN model produced superior results in comparison to other competitive ML based models.

5.2 Performance comparison

Table 7 lists the results and outcomes of all the classifiers under study. Results shown in Table 7 declare the best predictive performance of CNN model for customer review helpfulness dataset. The various tuning parameters used in this study are given in Table 7. Both datasets (i.e., DS1 and DS2) were simulated for 10 epochs. Accuracy and loss curve yielded by CNN model for DS1 and DS2 are demonstrated in Fig. 9.

Fig. 9

Accuracy and loss of CNN predictive model for DS1 and DS2.

The proposed model achieved an accuracy of 99.56% and 98.87% percent for DS2 dataset (shown in Fig. 9), while it achieved loss of 1.45 and 0.53 percent during training and validation phase Using hybrid feature sets as an input, our model obtained the good performance with precision, recall, accuracy, F1, and AP score of 99.7%, 100%, 99.26%, 99%, and 0.9999, respectively for dataset DS1.

On the DS2, CNN model achieved the Precision score of 99.01%, recall score of 100%, accuracy score of 98.97%, F1 score of 99.89%, and AP score of 0.9994 using the extracted features. For each datasets (DS1 and DS2), in the Fig. 10 precision recall curve demonstrates how these hybrid feature sets affect classification outcomes. In both precision & recall curve (DS1) and (DS2), violet color curve indicates the maximum area covered by the suggested model. The developed model gives an AUC value of 1for both datasets DS1 and DS2, respectively. LDA obtained second highest AUC value of 0.999 for both datasets.

Figure 11 uses a precision recall curve to offer a distinct view of performance in terms of average precision (AP) for all various ML methods. The CNN model scored the highest average precision (AP) value of in precision recall curve (DS1) and (DS2) for both the datasets DS1 and DS2 respectively. LDA obtained the second highest score with average precision value of 0.999 for DS1, while LR and LDA models secured 2nd highest score with average precision value of 0.999 for DS2, respectively. GNB ML model obtained the minimum with average precision value of 0.977 and 0.987 for dataset DS1 and DS2 respectively. As a part predictive model for the classification of experimentation we also computed the precision-recall curve for the both dataset and concluded that CNN is a powerful helpfulness of user-generated text reviews.

Fig. 10

Precision recall curve for AUC for DS1 and DS2.

Fig. 11

Precision recall curve for AP (Average Precision) for DS1 and DS2.

In this case also, GNB model yielded the poor performance.

5.3 Comparison with best earlier models

In Table 9, the performance of our predictive model is evaluated and compared with current existing models contributed by other authors in recent years. The results presented in Table 8 demonstrate satisfactory performance of the proposed model in comparison to other models. The results presented in tabular form illustrate that CNN got improvement in precision by 29.79%, 15.01% and 16.4% with respect to [10 , 50] methods, respectively. The developed approach achieved F1 score of 99.00 and 99.89 for searches and experience based products, respectively. We also evaluated the research’s findings in terms of average accuracy (AP) with the numerical score of 0.9997 for experienced product (EP) and 0.9999 for search products (SP).

Table 9
Comparative analysis of the best earlier models

Model Dataset Features ML Algo. Precision Recall Accuracy F1 score AUC AP

Proposed Amazon Review CNN 99.00% (S) 100% (S) 99.26% (S) 99.00 (S) 0.9999 (S) 0.9999 (S)

Model 99.01% (E) 100% (E) 98.97% (E) 99.89 (E) 0.9998 (E) 0.9997 (E)

[10] Yelp Review, B-GBT 69.22% 68.25% 70.36% 68.73% 0.773 X

Business,

Reviewer

[26] Amazon Review RandF X X 81.33% (S) 87.21 (S) X X

77.02% (E) 77.87 (E)

[38] Amazon Review, DNN X X 84.54% (S) 92.1 (S) X X

Product 77.1% (E) 85.89(E)

[49] Amazon Review DT 77.0±7% 77.0±7% X 77.0±6% X X

[50] Social Review HQRM 80.1% (A) 64.4% (A) X 71.3% (A) X X

media 71.1% (B) 71.6% (B) 71.4% (B)

83.6% (C) 43.9% (C) 57.6% (C)

Model	Dataset	Features	ML Algo.	Precision	Recall	Accuracy	F1 score	AUC	AP
Proposed	Amazon	Review	CNN	99.00% (S)	100% (S)	99.26% (S)	99.00 (S)	0.9999 (S)	0.9999 (S)
Model				99.01% (E)	100% (E)	98.97% (E)	99.89 (E)	0.9998 (E)	0.9997 (E)
[10]	Yelp	Review,	B-GBT	69.22%	68.25%	70.36%	68.73%	0.773	X
	Business,
	Reviewer
[26]	Amazon	Review	RandF	X	X	81.33% (S)	87.21 (S)	X	X
						77.02% (E)	77.87 (E)
[38]	Amazon	Review,	DNN	X	X	84.54% (S)	92.1 (S)	X	X
Product						77.1% (E)	85.89(E)
[49]	Amazon	Review	DT	77.0±7%	77.0±7%	X	77.0±6%	X	X
[50]	Social	Review	HQRM	80.1% (A)	64.4% (A)	X	71.3% (A)	X	X
media				71.1% (B)	71.6% (B)		71.4% (B)
				83.6% (C)	43.9% (C)		57.6% (C)

B-GBT: Bagging Gradient Boosted Trees, HQRM: Helpful Quality-related Review Mining, S: Search Product(SP), E: Experience Product(EP), DNN: Deep Neural Network, (A): HQRM+C4.5, (B): HQRM+NB, (C): HQRM+KNN, DT: Decision Tree, RandF: Random forest, AUC Area under curve, AP: Average precision.

As is evident from the results, precision recall curve is near to 1 in our experiment. It indicate that CNN model has a powerful ability to classify the helpful and unhelpful problem. In [10] paper, authors used area under curve (AUC) to assess the predictive power of model and achieved the value of 0.773, although CNN model gives the score of 0.9999 and 0.9998 for the SP and EP datasets that seems to be better than the existing models. In addition CNN model produced the superior outcomes in terms of accuracy.

6 Conclusion and feature work

In this work, various competitive ML algorithms models CNN, KNN, LR, GNB, and LDA were developed in order to perform the classification of user generated text reviews dataset in terms of helpful and unhelpful reviews on two prominent Amazon datasets. In this study, models were trained using integrating the five set of features, which are known as voting, structural, lexical, sentimental, and linguistic characteristics. As discussed in above sections, features have limited scope and performance as an individual entity individual entity. Hence, in this study, we have integrated five set of features to analyze their classification performance in terms of various evaluation metrics such as precision, recall, AP, accuracy etc. To give better insight about the classifiers performance, the developed models generated a particular score. CNN achieved the accuracy of 99.26% and 98.97% for DS1 and DS2, respectively, while it achieved precision of 99% and 99.01% respectively for same datasets. The results produced by CNN model have consistently shown the superior performance over other state of art models, when these models were tested using different product categories available on Amazon. Based on the outcomes presented in results section, CNN model has proven its importance as a competitive classifier as compared to KNN, Linear regression LR, Gaussian Naive Bays GNB, and LDA for review helpfulness prediction problem.

The main limitation of the present work is that it requires manually generated feature set to solve classification problem. However, the present work can be further extended in near future to improve its scalability by using embedding and optimization techniques. The use of innovative features in related fields, such as ranking products based on helpful reviews and identifying prominent reviewers for their efficient ranking based on a hybrid set of attributes, is an additional promising study area.

References

Chakravarti

, Janiszewski

, Ulkumen

The neglect ofprescreening information, Journal of Marketing Research 43(4) (2006), 642–653.

, Huang

, Tan

C.H.

, Wei

K.K.

Helpfulness of online product reviews as seen by consumers: source and content features,, International Journal of Electronic Commerce 17(4) (2013), 101–136.

Guo

, Wang

, Wu

Positive emotion bias: role of emotional content from online customer reviews in purchase decisions, Journal of Retailing Consumer Service 52 (2020), 101891.

Murphy

Local Consumer Review Survey 2018, retrieved June 20, 2020. from https://www.brightlocal.com/research/local-consumer-review-survey/ (2020).

Alam

, Singh

et al. Distance-based confidence generation and aggregation of classifier for unstructured road detection, Journal of King Saud University–Computer and Information Sciences (2021). ISSN 1319–1578. https://doi.org/10.1016/j.jksuci.2021.09.020

Dafni

R.J.

, VijayaKumar

, Singh

, Sharma

S.K.

Computer-aided diagnosis for breast cancer detection and classification using optimal region growing segmentation with MobileNet model, Concurrent Engineering 30(2) (2022), 181–189. doi: 10.1177/1063293X221080518.

Huang

A.H.

, Chen

, Yen

D.C.

, Tran

T.P.

study of factors that contribute to online review helpfulness, , Computers in Human Behavior 48 (2015), 17–27.

Diaz

G.O.

, Ng

Modeling and prediction of online product review helpfulness: a survey, Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, Long Papers 1 (2018), 698–708. https://simpletexting.com/6-examples-of-good-customer-reviews/

Kumar

, Singh

, Tiwari

Path Planning for the Autonomous Robots Using Modified Grey Wolf Optimization Approach, Journal of Intelligent and Fuzzy Systems 40 (2021), 94–53-9470.

10.

Bilal

, Marjani

et al. Profiling reviewers’ social network strength and predicting the “Helpfulness” of online customer reviews, Electronic Commerce Research and Applications 45 (2021), 01026.

11.

Park

Y.J.

Predicting the Helpfulness of Online Customer Reviews across Different Product Types, Sustainability 10 (2018); 1735. doi: 10.3390/su10061735.

12.

Chua

A.Y.

, Banerjee

Understanding review helpfulness as a function of reviewer reputation, review rating, and review depth, Journal of the Association for Information Science and Technology (2014).

13.

Forman

, Ghose

, Wiesenfeld

Examining the relationship between reviews and sales: The role of reviewer identity disclosure in electronic markets, Information Systems Research 19(3) (2008), 291e313.

14.

, Liu

, Zhang

Do online reviews affect product sales? The role of reviewer characteristics and temporal effects, Information Technology Management 9(3) (2008), 201–214.

15.

Lee

, Choeh

J.Y.

The interactive impact of online word-of-mouth and review helpfulness on box office revenue, Management Decision (2018).

16.

Chevalier

, Mayzlin

The effect of word of mouth on sales: Online book reviews, Journal of Marketing Research, 43(3) (2006), 345–354.

17.

Mudambi

S.M.

, Schuff

What makes a helpful online review? A study of customer reviews on amazon.com, MIS Quarterly 34 (2010), 185–200.

18.

, Zhang

, Pavlou

Overcoming the J-shaped distribution of product reviews, Communications of the ACM 52(10) (2009), 144–147.

19.

Zhang

, Varadarajan

Utility scoring of product reviews, In: Proceedings of the 15th ACM international conference on information and knowledge management CIKM’06, (2006), pp. 51–57.

20.

Sharma

S.P.

, Singh

, Tiwari

Prediction of Customer Review’s Helpfulness Based on Feature Engineering Driven Deep Learning Model, International Journal of Software Innovation (IJSI) 11(1) (2023), 1–6.

21.

Huang

A.H.

, Yen

D.C.

Predicting the helpfulness of online reviews –A replication, Computer Interaction 29 (2013), 129–138.

22.

Cao

, Duban

, Gan

Exploring determinants of voting for the helpfulness of online user reviews: A text mining approach, , Decision Support Systems 50 (2011), 511–521.

23.

Otterbacher

Helpfulness” in online communities: a measure of message quality, In: Proceedings of the 27th SIGCHI Conference on Human Factors in Computing Systems, ACM (2009), 955–964.

24.

Lee

, Choeh

J.Y.

Predicting the helpfulness of online reviewsusing multilayer perceptron neural networks, Expert Systemswith Applications 41(6) (2014), 3041e3046.

25.

Wang

, Tanga

, Kimb

More than words: Do emotional content and linguistic style matching matter on restaurant review helpfulness? International Journal of Hospitality Management (2018).

26.

Krishnamoorthy

Linguistic features for review helpfulness prediction, Expert Systems with Applications 42(7) (2015), 3751e3759.

27.

Ramanan

, Singh

, Kumar

A.S.

et al. Secure blockchain enabled Cyber-Physical health systems using ensemble convolution neural network classification, Computers and Electrical Engineering 101 (2020), 108058. https://doi.org/10.1016/j.compeleceng.2022.108058

28.

Liu

, Park

What makes a useful online review? Implication for travel product websites, , Tourism Management 47 (2015), 140–151.

29.

Srinivas

, Singh

et al. Multi-modal cyber security-based object detection by classification using deep learning and background suppression techniques, Computers and Electrical Engineering 103 (2022), 1083. https://doi.org/10.1016/j.compeleceng.2022.108333

30.

Singh

, Alam

et al., Design of thermal imaging-based health condition monitoring and early fault detection technique for porcelain insulators using Machine learning, Environmental Technology, and Innovation 24(21) (2021), 102000.

31.

Iwendi

, Ibeke

, Eggoni

, Velagala

, Srivastava

Pointer-based item-to-item collaborative filtering recommendation system using a machine learning model, International journal of information technology and decision making [online] 21(1) (2022), 463–484. https://doi.org/10.1142/S0219622021500619

32.

Zhang

, Wei

, Chen

Estimating online review helpfulness with probabilistic distribution and confidence, Foundations and Applications of Intelligent Systems, Springer, (2014).

33.

Liu

, Huang

et al. Modelling and predicting the helpfulness of online reviews, Eighth IEEE international conference on data mining, (2008), pp. 443–452.

34.

Ngo-Ye

T.L.

, Sinha

A.P.

The influence of reviewer engagement characteristics on online review helpfulness: A text regression model, Decision Support Systems 61 (2014), 47e58.

35.

Ghose

, Ipeirotis

P.G.

Estimating the helpfulness and economic impact of product reviews: Mining text and reviewer characteristics, IEEE Transactions on Knowledge and Data Engineering 23(10) (2011), 1498e1512.

36.

Quaschning

, Pandelaere

, Vermeir

When consistency matters: The effect of valence consistency on review helpfulness, Journal of Computermediated Communication 20(2) (2015), 136e152.

37.

Chen

C.C.

, Tseng

Y.D.

Quality evaluation of product reviews using an information quality framework, Decision Support Systems 50(4) (2011), 755–768.

38.

Malik

M.S.I.

, Hussain

Helpfulness of product reviews as a function of discrete positive and negative emotions, Computers in Human Behavior 73 (2017), 290–302.

39.

Goldie

Emotions, feelings and intentionality, Phenomenology and the Cognitive Sciences 1 (2002), 235–254. https://doi.org/10.1023/A:1021306500055.

40.

Rafaeli

, Sutton

R.I.

Emotional contrast strategies as means of social influence: lessons from criminal interrogators and bill collectors, Academy of Management Journal 34(4) (1991), 749–775.

41.

https://www.cuemath.com/calculus/area-under-the-curve/

42.

https://hasty.ai/docs/mp-wiki/metrics/average-precision

43.

Mittal

, Iwendi

, Khan

, Javed

A.R.

Analysis of security and energy efficiency for shortest route discovery in low-energy adaptive clustering hierarchy protocol using Levenberg-Marquardt neural network and gated recurrent unit for intrusion detection system, Trans Emerging Tel Tech (2020), e3997. wileyonlinelibrary.com/journal/ett © 2020 John Wiley & Sons, Ltd. 1 of 16. https://doi.org/10.1002/ett.3997.

44.

Iwendi

, Khan

, Anajemba

J.H.

, Mittal

, Alenezi

, Alazab

The Use of Ensemble Models for Multiple Class and Binary Class Classification for Improving Intrusion Detection Systems, , Sensors 20 (2020), 2559. doi: 10.3390/s20092559 www-mdpi-com.web.bisu.edu.cn/journal/sensor

45.

Alsmadi

, AlZu’bi

, Hawashin

, Al-Ayyoub

, Jararweh

Employing Deep Learning Methods for Predicting Helpful Reviews, 11th International Conference on Information and Communication Systems (ICICS), IEEE-(2020).

46.

, Duong

, Nguye

, Cao

From Helpfulness Prediction to Helpful Review Retrieval for Online Product Reviews”, Association for Computing Machinery, ACM ISBN 978-1-4503-6539-0/18/12(2018).

47.

Singh

J.P.

, Irani

, Rana

N.P.

et al. Predicting the “helpfulness” of online consumer reviews, , Journal of Business Research 70 (2017), 346–355.

48.

Hutto

C.J.

, Gilbert

VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text, In: Proceedings of the Eighth International AAAI Conference on Weblogs and Social Media, (2014).

49.

Haque

M.E.

, Tozal

M.E.

, Islam

Helpfulness Prediction of Online Product Reviews, DocEng, Halifax, NS, Canada, (2018), 28–31.

50.

Jiang

, Liu

, Ding

, Liang

, Duan

Capturing helpful reviews from social media for product quality improvement: a multi-class classification approach, Int J Prod Res 55(12) (2017), 3528–3541.

51.

Bilal

, Marjani

et al. Profiling User’s Behavior, and Identifying Important Features of ReviewHelpfulness, Digital Object Identifier 10.1109/ACCESS,2989463(2020).

52.

Gang

, Taeho

Examining the relationship between specific negative emotions and the perceived helpfulness of online reviews, Information Processing and Management (2018), 0306–4573.

53.

Malik

M.S.I.

, Hussain

Exploring the influential reviewer, review and product determinants for review help fulness, Article in Artificial Intelligence Review, (2020).

54.

Malik

M.S.I.

Predicting users’ review helpfulness: the role of significant review and reviewer characteristics, Springer Nature (2020). 1003

55.

Qazi

, Syed

K.B.S.

, Raj

R.G.

et al.A concept-level approach to the analysis of online review helpfulness, Computers in Human Behavior 58 (2016), 75–81.

56.

, Zhan

Online Persuasion: How the Written Word Drives WOM Evidence from Consumer-Generated Product Reviews, Journal of Advertising Research 51(1) (2011), 239–257.

57.

Baek

, Ahn

, Choi

Helpfulness of online consumer reviews: Readers’ objectives and review cues, International Journal of Electronic Commerce 17(2) (2012), 99–126. https://doi.org/10.2753/JEC1086-4415170204

Collection of Amazon reviews
Dataset	Product type	#reviews	#products	#reviewer	Product categories
DS1	Cellphone and accessories	194,439	10,429	27879	Search goods
DS2	Home and Kitchen	551,682	28237	66519	Experience goods

Dataset	Number of reviews
	Before processing	After preprocessing
DS1	3,447,249	8775
DS2	4,253,926	59515

Integrated feature engineering based deep learning model for predicting customer’s review helpfulness

Abstract

Keywords

1 Introduction

Table 3 Description of the dataset for Amazon reviews Collection of Amazon reviews Dataset Product type #reviews #products #reviewer Product categories DS1 Cellphone and accessories 194,439 10,429 27879 Search goods DS2 Home and Kitchen 551,682 28237 66519 Experience goods

Table 4 Summary about the dataset1 (DS1) and dataset2 (DS2) Dataset Number of reviews Before processing After preprocessing DS1 3,447,249 8775 DS2 4,253,926 59515

3.3.2 Lexical feature set (Le)

3.3.4 Linguistic feature set (Li)

5.1 ML model performance evaluation

References

Table 3
Description of the dataset for Amazon reviews

Collection of Amazon reviews

Dataset Product type #reviews #products #reviewer Product categories

DS1 Cellphone and accessories 194,439 10,429 27879 Search goods

DS2 Home and Kitchen 551,682 28237 66519 Experience goods

Table 4
Summary about the dataset1 (DS1) and dataset2 (DS2)

Dataset Number of reviews

Before processing After preprocessing

DS1 3,447,249 8775

DS2 4,253,926 59515

3.3.2 Lexical feature set (L_e)

3.3.4 Linguistic feature set (L_i)