A Study on Fraud Reviews:Incentives to Manipulate and Effect on Sales

Tao Yin,Wenqi Wang,Wenhua Shi＊

School of Economics and Management,Beijing University of Posts and Telecommunications,Beijing 100876,China

Abstract:With the broad reach of Internet,online reviews have become an important source of electronic Word-of-Mouth.Fraud reviews that are deliberately posted by businesses are a type of online reviews.This paper discusses the incentives of fraud reviews and the effect of fraud reviews on consumer behavior through empirical research.Using book download data at Amazon,we find that a book is more likely to manipulate fraud reviews when it has few online reviews posted by real consumers,higher proportion of negative reviews,longer average length of negative reviews,lower average rating scored by real users and higher price.And fraud reviews change the review environment and have a significant impact on the consumer purchasing decisions.More number,higher proportion,longer word count and higher promotion of rating of fraud reviews lead to higher sales.The results also show consumers can discern the manipulation of fraud reviews to a certain extent.

Keywords:electronic word of mouth; online reviews; fraud reviews; purchasing decision

I.INTRODUCTION

Nowadays,the influence of Internet has penetrated into all aspects of people's daily life,and more consumers are inclined to shop online.Consumers can easily access various types of online Word-of-Mouth (WOM) information when searching products.Electronic word-of-mouth (eWOM) as a special form of WOM whereby consumers share their experiences with others online has become a valuable source of information on products and gradually replaced the function of the offline WOM [1].Online user reviews as a representative EWOM in the business platform,have become an important source of information to consumers to make better decisions [2].Recent years,more researchers have turned to examining the effect of online reviews.Researches find online reviews have a directly positive impact on the product sales [3,4,5].Clemons [4] find that a very positive rating has a significant impact on product sales.Luca [6] examined consumer reviews on Yelp,a hotel reviews website and found that when consumer's rating on the hotel increased by one star,the hotel's revenue would increase by 5% to 9%.In the e-commerce shopping site,nearly nine out of ten consumers think they are influenced by positive reviews before making purchasing decisions [7].And two aspects of the effect of online reviews have been widely discussed:volume (the total number of user reviews) and valence (average rating of user reviews).A large number of reviews can make the product stand out and attract the user's attention [3,8].Review rating can provide information about quality of product to consumers and affect their attitudes towards uncertainty and selection of product utility [3,9,10].

This paper discusses the incentives of fraud reviews and the effect of fraud reviews on consumer behavior through empirical research.

Due to the high value and prevalence of online reviews,some organizations and individuals by interests or reputation exploit the lack of network information regulatory system and illegal marketing behavior to create enormous fraud reviews,which seriously endangers the online shopping order and environment.The earliest study of spam reviews is Professor Jindal [11].He proposed the concept of spam reviews and divided them into two types.One is the deceptive opinions intended to promote products or damage reputation.Another is non-reviews which contain no opinions about the product.In this work,promotional deceptive reviews are called fraud review,which are non-real reviews posted by suppliers,publishers or writer,used to increase the product reliability.

These days,reviews on online e-commerce sites are filled with massive fraud reviews.There is considerable evidence to prove this is a common phenomenon in the industry.For example,Hu [12] studied the website reviews of Amazon and Barnes & Noble,two online bookstores in the United States,and found that there was a lot of fraud information posted by publishers and sellers,which largely disrupted purchasing decisions of consumers.The New York Times [13] reports that companies hired workers of Mechanical Turk to publish fraud five-star reviews on Yelp,paying 25 cents for each review.Mayzlin et.al found that some of hotel reviews from Expedia and Tripadvisor were manipulated form the view of affiliation,ownership and management and structure of hotels [14].And fraud reviews are displayed in other fields.For example,Deming et.al [15] found that SUs could suffer from a new group cheating problem.Parts of users conspire to manipulate the auction by submitting untruthful bids.

The current researches on fraud reviews focus on the detection.For example,Liu [16] extracted the review features according to three aspects,and used the support vector machine to identify the quality of reviews.Jindal and Liu [17] applied logistic regression by 36 features of the review to identify spam reviews.But research based on effect of fraud reviews directly is almost a blank area.At the same time,the fraud reviews are not affected by reviews that posted by consumers who buy products in real terms,but whether users make purchasing decisions and decide to post feedbacks would be affected by fraud reviews.Fraud reviews affect the consumer's purchasing decisions and break the online market orders.We hope to discuss the incentives of fraud reviews form the view of reviews and price of the product and find the effect of fraud reviews on sales.

To do so,some representative reviews of books at Amozon.com are downloaded.And then mark fraud reviews by using three existing detective methods.Following previous literatures [18,19],panel data with lagged variables are used to directly study the incentives to commit fraud reviews and the effect of fraud reviews on books sales,helping the websites to regulate the goods and online reviews better and form a good market circumstances and orders.

II.THEORETICAL FOUNDATION

When consumers make purchasing decisions,they are in a situation of incomplete information,lacking full information on product quality,seller quality and other information.When acquiring product information,consumers need to spend cost of time to identify product quality.Therefore,the total cost of a product must include both the product cost and the cost of search [20].The user would turn to online reviews to minimize the uncertainty and reduce the search cost.Online reviews are usually in the product page,containing product-related information,greatly facilitating the access,reducing the search costs.User behaviors are largely based on online reviews.

A key determinant of search cost is the nature of a product.Nelson [20,21] divided the product into experience goods and search goods from the angle of economics.The evaluation of the former is mainly based on objective experience,and consumers mainly evaluate the quality of products and make purchasing decisions through comments.While the latter are those for which consumers have the ability to obtain information on product quality prior to purchase and evaluate the quality primarily through the objective attributes.Examples of experience goods are music,books and beer,and examples of search goods are cameras [20].With the prevalence of online reviews,the difference between two items is decreasing [22].But this classification is still widely used,because of differences between the access of product-related information and the processing methods [3,23].Studies have shown that consumers who buy experience goods tend to be more dependent on online reviews [22].

Two main theories of relevant literature explain users' decisions to share feedback online:psychological motivation theory and review environment theory.The first theory maintains that consumers share experience online for self-enhancement to to gain attention and enhance their images among others.The second theory holds that the user's decision to post comments depends on the existing comment environment,which is mainly applied in this study.Goes [24] confirm that prior product ratings would affect later consumer feedback.According to Xiao [18],existing user reviews would have a strong influence on subsequent reviews,moreover longer and more frequent reviews could eliminate consumer bias.According to review environment theory,we think that the manipulation of fraud reviews is to improve the existing review environment,enhance consumer recognition,and increase their purchasing behavior,by improving the characteristics of a period of review.

The current studies of the impact of online reviews can be considered based on two assumptions [12].One is reviews are written by real consumers,representing the authentic information of the product.Another is consumers can be wise to distinguish fraud reviews if the product has manipulations.In fact,fraud reviewers deliberately imitate the real consumers to publish fraud reviews for making transactions more authentic.It is difficult for consumers to distinguish fraud reviews from authentic information.From the consumer's point of view,the impact of fraud reviews on consumers that is considered as special online reviews is the same as the impact of online reviews.

The influencing factors of online reviews can be summarized from two aspects:the language features and the non-linguistic features.The first feature mainly includes the length of the review,the positive and negative of the review,the polarity of emotion,the variance of positive and negative emotions.The latter feature mainly includes the volume and the valence of reviews.The impact of existing research online reviews is focused on the non-linguistic features.Researches have shown that the amount of reviews can significantly improve market results [2,25,26].

A large number of reviews can be better to attract the attention of users and improve the user's purchase probability [8,26].

In this study,fraud reviews are considered as a deliberate release of online reviews and its influence among the consumers is the same as that of the normal users.Therefore,influence factors of online reviews concluded by previous works are used to analyze.And what situations the fraud reviews affect the consumer's buying behaviors.We hope to discuss the incentives of fraud reviews including reviews environments and the characteristics of the products and the impacts of fraud reviews on the consumer's buying behaviors.Previous studies have used proxy to represent the release of fraud reviews.For example,in the study of Hu [11],the variance of the adjacent review score have been used to indicate the degree of manipulation.Luca [6] have used reviews that are automatically filtered by Yelp as a proxy.Mayzlin [14] used the ratio of extreme reviews to total reviews to represent the manipulation where the 5 star reviews were to show manipulation.In this study,the number of fraud reviews which are detected using the existing identification methods are directly employed on behalf of a certain period of time to manipulate reviews.The choice of methods is described in the following section.Following previous studies [2,3,27],sales of products are used to represent consumer behaviors to purchase.Products with higher sales lead to more buying behaviors of consumers.The influencing factors and relationships are described in the next section.

2.1 Review valence and star ratings

Previous studies often use product star ratings as the research variable replacing the review valence.Reviews of rating above/equal/below 3 are defined as positive/neutral/negative user reviews.A very low rating (one/two star rating) indicates a negative view of the product,high ratings (five stars) reflects a positive view of the product,and a three-star rating reflects a moderate view.The star ratings are a reflection of satisfaction with the product to a certain extent.

The researches of star ratings on sales show mixed results.Chevalier [3] have found that ratings have a direct impact on consumer purchasing decisions by studying book reviews at Amozon.com and Barnesandnoble.com and the higher the average ratings,the higher the sales of the product.Moreover,one-star reviews are more pronounced than the five-star reviews.Similarly,Dellarocas [28] discusses the revenue of film and have found that the total revenue of the film could be effectively predicted trough the film audience's ratings and the film's income trajectory.Zhu [9] found that the average rating has a significant effect on sales by studying the consumer characteristics and the nature of the product in the video game market.In particular,this effect is more pronounced for books with lower prevalence.However,some scholars showed that the average ratings of reviews did not affect choices of consumers.For example,Chen [29],who used the same data set downloaded from Amazon,found that the valence did not affect the product sales.

For the mixed results of the valence,some scholars have begun to study its internal reasons.Researches have shown that whether ratings affect consumer choices depends on the specific contextual factors,including the popularity and the type of product [19].In this study,a single type book is chosen as a research object.Meanwhile,the products are filtered with a view to obtain a representative sample.So we follow the previous view that average rating has significant an impact on the sales of the product.

The lower average rating of product has a negative impact on consumers and reduces the purchasing desire.In such situation,businesses will manipulate more fraud reviews to increase the average rating and reduce the negative reviews on the degree of attention.Therefore,we hypothesize:

H1a.The lower average rating,the more number of fraud reviews.

Fraud reviews are suppliers,publishers or writers deliberately posted to improve the user's recognitions and increase the purchasing behavior,so they are extremely five-stars,which can significantly increase rating over a period of time and improve acceptance.But studies have found that high ratings are not always well.Mudambi [23] have shown that the relationship between ratings and usefulness of reviews is an inverted U-shaped for the experience goods.That is,reviews with more extreme ratings are less useful.A meta-analysis conducted by Purnawirawan [30] demonstrates that effects of online reviews are nor-linear and valence has a curvilinear effect on usefulness and a ceiling effect on attitudes.The reason for this phenomenon is that when the rating is too high,the consumers have a sense of distrust and reduce the credibility of the reviews.

According to the analysis,we hold that fraud reviews improve the existing comment environment.So the ratio of the rating after posting fraud reviews to the rating posted by real consumers is a variable in this study to discuss the impact of fraud reviews on rating of the product and the impact of this improvement on sales.When the proportion is high,the improvement is significant.But the higher proportion means manipulating more fraud reviews or the lower rating of normal reviews,which make this increasing too obvious and lead to decrease in sales instead.Therefore,we hypothesize

H1b.The greater improvement of fraud reviews on average rating leads to higher sales.But when the promotion reaches a certain level,it leads to a decrease in sales.

2.2 Review length

Consumers spend a lot of time and effort to assess their own choice,but a lack of information and motivation make purchasing decisions and actual purchase behavior with difficulty.When the information is highly relevant and clear,consumers have enough confidence to buy the product [25].

Review length containing descriptions of the information of product and quality refers to the number of words that an online review is made after the consumer purchase the product.Most studies suggest that the amount of words reflects the richness of the product.Reviews with longer length describe the product details more clearly and give consumers more information about the product.That help the consumers perceive the product and enhance the cognitive emotions thus consumers could make the purchasing decisions better [31].

The researches about the review length generally conclude that it significantly influences product sales and purchase behaviors of users.Chevalier et.al [3] have shown that consumers not only concerned with the average product rating,but also focus on the length and the content of the reviews.Moreover,Longer reviews stimulate consumers to read carefully,deepen or change their original view,improve the awareness of the product or service and eliminate consumer uncertainty.Mudambi [23] have found that the longer reviews are more useful for the consumers and strongly impact consumer's acceptance.

It is considered that different rating of reviews contains different information.The five-star reviews contain positive evaluations of the product,so the longer reviews are beneficial for the goods.In contrast,reviews with lower ratings include negative information and longer negative reviews make the attitude about product become inactive and reduce the purchasing decisions.Therefore,we think that the average longer words of negative reviews in the existing stage have passive impacts on consumers.Publishers and suppliers issue more fraud reviews to compensate for more negative information and increase consumer identity.Therefore,we hypothesize

H2a.The longer average negative review length,the more number of fraud reviews.

The length of the review is affected by product type and price and other factors.But fraud reviews are manipulated by sellers,so the average length is longer than most reviews posted by real consumers,which lengthen the average length.At the same time,the fraud reviews all contain the positive evaluation of the product,giving consumers favorable information and enhance recognitions of the reviews and products.That boost the purchase behaviors and have a significantly impact on sales.Therefore,we hypothesize

H2b.The longer average length of fraud reviews,the higher the sales of the product.

2.3 Review volume

The number of reviews reflects the total amount of information in the reviews to a certain extent.A large number of reviews give consumers further messages about product and increase product popularity.Chevalier [3] study the effect of online reviews in book market at Amazon.com and B&N.com through Differences-in-Differences Analysis,and find that review volume has a significant impact on sales.Similar studies have obtained the same results in different product areas.Liu [16] and Duan [2] have found that the volume has an effect on the box office in the field of cinema.Gu [32] verify that the product with massive number of online reviews has a forward sales ranking in the field of electronic products.Wu [8] have shown that the volume of review significantly affects the consumer's willingness in terms of consumer uncertainty about a product.Online reviews enhance consumer acceptance and consumer purchasing behavior by increasing the number.

Following prior researches that the volume affects sales namely a small number of reviews lead to the less sales,we propose that some businesses deliberately publish reviews to augment sales and enhance the focus on specific goods when the number of reviews is minor at a certain stage.Based on it,we make an assumption

H3a.The number of fraud reviews is manipulated when product has a less number of online reviews.

At the same time,reviews with different extremities have diverse impacts.Nga [10] have showed that positive reviews increase product sales while negative reviews reduce sales.Some studies have found that the impact of negative reviews is more pronounced than the positive [3,33].This is because positive reviews are common,so users are more concerned about another viewpoint.To a certain extent businesses manipulate fraud reviews for reducing the proportion of negative reviews.So it is considered that an incentive to commit fraud reviews is a large proportion of negative reviews.Based on it,we make an assumption

H3b.The greater the proportion of negative reviews,the more the number of fraud reviews being posted.

Based on the above analysis,after the product posting substantial fraud reviews,they provide consumers with greater information.Manipulations promote sales which affect consumers purchase behaviors in a certain period of time.Massive fraud reviews descend the sale ranking,which strengthen the desire to buy and improve the overall product sales.

The relationship between the number of fraud reviews and the total number of is also meaningful to discuss the effect.The proportion of fraud reviews is used as a variable in our research.High proportion contains more fraud positive information that increases purchase uncertainty and the sale.But the proportion exceeds a certain range means product less normal reviews that is,has less purchase behavior.In particular,when the ratio is the maximum,there is little real review meaning fraud reviews do not affect the consumers and the products only manipulate the purchase behavior.Therefore,we argue that fraud reviews affect the purchasing decision through the higher proportion.But this relationship is an inverted U-shaped,indicating extreme proportion is easier to distinguish.This leads us to hypothesize

H3c.The number and the proportion of fraud reviews affect sales of product.However,the extreme number and proportion of fraud reviews has a negative effect on sales.

2.4 Price

Price is an important attribute of the product itself; different prices of products affect the consumer attention and preferences to the product.Huang [22] have shown that consumers collect relevant nature attributes of product when they have willingness to product in which the price of the product plays an important role in their purchasing decision-making.Chevalier [3] has found that product with lower sales generally has higher price in the same situation.This is consistent with consumer psychology,that users are more cautious to purchase the higher price products which cost more.So the products positioning high prices of businesses in order to enhance their sales,will be more concerned about the quality of the comments,will manipulate more false comments.So businesses with the product positioning high prices will be more concerned about the quality of the comments and manipulate more fraud reviews in order to improve sales.This leads us to hypothesize

H4a.The product with higher price is manipulated more fraud reviews.

In fact,the price of the product fluctuates due to supply and demand,economic cycle,seasonal changes and other factors.In some cases,businesses need to raise the price,which may cause decline in sales according to the experience.So we argue that some businesses choose to improve the quality of online reviews to maintain or enhance sales.Therefore,we hypothesize

H4b.When the price of product rises,more fraud reviews are manipulated.

2.5 Identification of review spam

Recent years,more researches have focused on identifying review spam.Professor Jindal [11],who was the first one to study this area,used Logistic regression model to review spams of books,music,and other manufacturing products.Then Jindal [17] added the variables.A total of 36 features were extracted based on reviews,reviewers and review targets,including the characteristics of the content of the review text,emotional words of positive or negative reviews and the similarities between review and product features.

After that,more scholars have begun to carry out identification of research of fraud reviews.The detection could be concluded in three methods,machine learning,pattern recognition and classifier.Ott [34] collected online reviews of the top 20 hotels from the most popular ones on TripAdvisor.And then based on the knowledge of computational linguistics and psychology,the basic characteristics were analyzed from three aspects:text classification,psychology and genre recognition.Through these characteristics,a variety of classifiers were built to identify fraud reviews.Liu [16] applied support vector machine (SVM) to establish a machine learning model to distinguish the quality of the criticisms according features extracted from amount of information,readability and subjectivity of the product.Li [35] manually annotated a corpus and used semi supervised collaborative training algorithm to identify fraud reviews by crawling the online reviews on the Internet,which obtained the best effect in the semi supervised framework.

In order to get more accurate results,we hope to apply fraud reviews as real as possible to determine the effect on sales.Three detective methods are carried out to label the reviews.Every method is applied to label the review as a fraud review manipulated by business or a normal review posted by the real consumer.We consider reviews marked as fraud reviews by all three methods are fraud reviews,or reviews marked as normal reviews by all three methods are normal reviews.And we ignore reviews that are different labels by three methods.

After deeply studying and implementing detections mentioned above,we decide to implement methods proposed by Jindal,Li and Liu.Jindal uses logistic regression of 36 features.Li uses a semi-supervised collaborative algorithm to identify the method through a corpus.And Liu uses the SVM to identify by extracting characteristics of reviews.By and large,the three identification methods cover the existing detection theories.Meanwhile,they are widely cited and have higher recognitions.By simultaneously using these three methods to mark the reviews,we can get more accurate data,which contribute to the subsequent research.

It s a mercy, he said, that they didn t root up the tree on which I was perched, or I should have had to jump like a squirrel on to another, which, nimble though I am, would have been no easy job

III.RESEARCH METHODOLOGY

3.1 Data collection

We conducted this study in the online book market.According to the classification of Nelson [20],the book is a typical type of experience goods.The quality of a book is difficult to evaluate before adoption.Meanwhile,the online book market contains almost all kinds of bibliographies.Hence,it is challenging for online users to locate favorable products out of the abundant product selection in the current online book market.This motivates consumers to zealously resort to various WOM sources to make well-informed decision [32].Consumers with intentions to purchase books online are able to get the relevant WOM information and search the online review communications.Therefore,information about the qualities and contents of books obtained from online reviews and nature of books influence user choices.

In particular,we empirically examine our model by using the data collected on Amazon.com.Many previous researches examining the effect of reviews are through the Amazon online bookstore which is also the largest one in the world.For example,Chen [29] and Chevalier [3] both have found that online reviews from Amazon have a significant impact on product sales.Moreover,Hu's research [12,36] and Professor Liu Bing's interview prove that fraud reviews are a common phenomenon in Amazon book market.Therefore,we discuss that online users consult the relevant reviews to discriminate books and part of businesses deliberately released fraud reviews to enhance consumer recognition and visibility.Finally,data of online reviews on Amazon's books are object in this study according to the nature of the books,the impact of online reviews,and the Amazon Web sites own situation.

We aim at to generate a representative sample of sales.Given that multimedia big data is difficult to analyze [37] and the impossibility to obtain the company's internal specific data of sales,we approximated a random sample of sales.In order to get enough lifecycles of fraud reviews for discussion,we randomly collected a sample of books that were released after June 2016.We got a sample of 2159 books.We further choose books that have received more than 100 consumer reviews to filter those have lower sales.That is to select books that are more likely to be manipulated.Finally,the collected data contains 1502 observations.

We note the specific release time of book through the book's details page.Following the extant studies [19,26],we study on a weekly basis.At the same time,review manipulations decrease as time goes by according to study of Hu [36].We collect reviews for twenty-six weeks about each book after it is released rather than the reviews of chosen books over a fixed period of time.For each review,we collect the following data:rating,content and posted time.Then the reviews marked as fraud reviews or normal reviews are divided by week.After processing,we get the data we need.

Amazon did not disclose the specific sales of products,but the sale ranking is available.So we rely on ranking data rather than more conventional sales.The related websites are accessed to obtain the weekly price and sale ranking of books.Many similar studies use the same operation.We directly use the sale ranks to discuss the effect of the fraud reviews on sales.Schnapp and Allwine [38] have find that the relationship betweenln(sales)andln(ranks)is approximately linear.Chevalier and Goolsbee [39] further explored the linear relationship between sales and sales rankings and pointed out thatln(ranks)could take place ofln(sales).Since then many scholars use this linear relationship to research in the books,software,electronic products and other fields [3,27].

3.2 Variables

Following the literature [3,27],we use weekly sale ranks to capture online user choices of books and we apply a natural log transformation on ranks.The reason for the log transformation is that the log specification could better estimates the effect of a change in the independent variables on the percentage change in the dependent variable [3].And the relationship between log ranks and log sales is closer to liner,the estimated coefficients in our specifications and their standard errors would simply be scaled by a constant.We use the weekly number of fraud reviews to express the emergence of manipulated.The ln(Ranking)i,kandFraud-VOLi,kare dependent variables of incentive model and effect model.Explanatory variables include variables related to review valence,length,volume,and price,where the valence is expressed by the star rating,and the volume is expressed by the number of reviews.

In the effect model of fraud reviews on sales,square term ofFraudFractioni,kis defined asFraudFractionSQi,kin order to verify whether impact of fraud reviews proportion on sales is inverted U-shaped.Square term ofRatioi,kis defined asRatiosqi,kto discuss influence of the promotive stars of fraud reviews on sales.Considering fraud reviews are only part of the reviews and it is not the case that all products are manipulated,a dummy variableFraudDi,kis used in the study to indicate whether the product have the manipulation.If the product has not received a fraud review at week t,FraudDi,kis 1,otherwise its value is zero.We add the average length of the normal reviews and the price of the product as control variables.

Followed the prior studies [19,26],a oneweek lag between independent variables and dependent variables of book sale and WOM are used to better represent the actual decision-making process due to the dynamic process between the online WOM and online user selection [26].Meanwhile,the potential feedback between the dependent variable and the associated independent variable can be effectively controlled.Finally,we get a panel dataset containing 26 weeks of online WOM data to analyze the dynamic process of online reviews.The descriptive statistics for the main variables in the data set are included in table 2.

3.3 Analysis method

Previous have found prior reviews will have an impact on subsequent reviews.And the actual process of purchasing decisions shows that relationship between online reviews and the user purchase behavior is a certain lag effect.Panel data is mixed data of time series and cross section,that could control individual heterogeneity,increase the sampling accuracy of the estimator,eliminate the heterogeneous effects of time and space and avoid multiple collinearity problems.Panel data can also study of the dynamic adjustment process better.Based on the construction of variables and the analysis of dynamic processes,we use the panel data with lag variables to build the model.

In the incentives to manipulate fraud reviews model and the effect of fraud reviews on sale ranking model,the explanatory variables are different.In the first model,in order to validate the hypothesis,we use the number of reviews posted by the normal user,the proportion of negative reviews,the average length of negative reviews,the normal average rating,the change of price product between two weeks,and the price as explanatory variables.

Table II.Summary data.

In the second model,we use the number of fraud reviews,the proportion of fraud reviews,the square of the proportion,a dummy variable to indicate whether manipulate a fraud review,the average length of fraud reviews,and the promotion through the fraud reviews,the squared of this promotion as the explanatory variable,while the average length of normal reviews and the price are control variables.

The incentive of fraud reviews model is:

Table III.The incentives to manipulate fraud reviews.

The effect of fraud reviews on sale ranking model is:

IV.RESULT AND CONCLUSION

The regression results of the incentives to commit fraud reviews model are included in table 3.The analysis of the model indicates a good fit,with a highly significant likelihood ratio (p=0.000).

The corresponding assumptions can be verified by the above model results.To test Hypothesis 1a,we examined the interaction of average normal rating and the number of fraud reviews.NormalVALi,k-1(p＜0.000) is statistically significant and the coefficient is negative,which can be concluded that one of the incentive of fraud reviews is a lower average normal rating of product.Lower rating makes the intention to manipulate more obvious,and the businesses want to offset or cover the rating through posting more fraud reviews.

In H2a,we hypothesize the effect of negative review depth on the number of fraud reviews.We find strong support for Hypothesis 2a.NegWordi,k-1is a significant and its coefficient is positive.When product receive longer average length of negative reviews which have the bad impression of the products,the businesses would manipulate more fraud reviews to make sure consumers skim more positive reviews to eliminate the bad evaluations.

The regression results test H3a that we assume one of the incentive of fraud reviews is less number of reviews posted by real users of product.In this case,less information the consumer can get from the reviews.According to the attributes of the experience goods,consumers evaluate the product primarily by reviews.So more fraud reviews manipulated by businesses,which can eliminate uncertainty of consumers about products.

There are some correlations betweenNormalVALi,k-1andNegFractioni,k-1,so we split them into two regressions.The results show that the fraction of negative reviews effect the manipulation.When there are a high proportion of negative reviews,the willingness to buy of users will decline.In such case,sellers post more fraud reviews to reduce the proportion of negative emotions to promote purchasing behaviors.So H3b is supported.

The results also provide strong support for H4a,which is the same by Hu [13].Products with higher price are manipulated more fraud reviews to eliminate product uncertainty,increase purchasing behavior.This is because consumers are more cautious to buy high cost products.But the regression results did not confirm H4b,thePriChangei,kis not significant.We consider this is due to the small range of price changes of the product.In the collection of data,we found that,contrary to the imagination,the product prices in a certain period of time tend to be smoothly.So the slight fluctuations of price are the normal lifecycle of product that have not much impact on consumer purchasing behaviors.The sellers have less intention to manipulate the fraud reviews.

The results of the effect model have been shown in table 4.The analysis of the model indicates a good fit,with a highly significant likelihood ratio (p=0.000).

Column 1 presents the result of a regression in which only price is included.The price coefficient is positive,suggesting that sales ranks at Amazon.com get larger when price rise.

Since there are some correlations betweenFraudVOLi,k-1andFraudFractioni,k-1,we discuss them in two regressions with the control of other variables.The results are shown in column 2 and column 3 from table 4 respectively.FraudDi,k-1is statistically significant and the negative coefficient,indicating there is a significant relationship between fraud reviews and sale rankings for experience goods.

The top-selling book has a sales rank of one,and the lower sellers are assigned higher sequential ranks.Ratioi,k-1andRatioSQi,k-1are used to discuss the impact of rating promotion by fraud reviews on sale rankings.These two variables are statistically significant.The coefficient of the linear term is negative,and the coefficient of quadratic term is positive indicating our hypothesized “inverted-U” relationship between rating promotion by fraud reviews and the sales.When the range of this promotion is larger,the rating of normal reviews comments is low and the user experience is poor.So the manipulation of fraud reviews is easy to discriminate,and it reduces the credibility of online reviews.This also shows the consumers have a certain ability to identify the real quality from the reviews when the manipulation is palpable.The result conforms H1b.

Table IV.The effect of fraud reviews on sales.

The regression results test the H2b.The coefficient of ln(FraudWord)i,k-1is negative,which means the average length of fraud reviews has a positive,significant effect on sales.Fraud reviews with more words include favorable evaluations on the product,which can promote the consumer's willingness to buy.

In column 2,there is a significant relationship between bothFraudFractioni,k-1andFraudFractionSQi,k-1and sale rankings.The negative coefficient forFraudFractioni,k-1and the positive coefficient forFraudFractionSQi,k-1also indicates our hypothetical “inverted-U” relationship between proportion of fraud reviews and the sales.The signs of the coefficients are contrary to the common“inverted-U” because of the negative correlation between sales rank and sales.Reviews with high proportion of fraud reviews contain have a higher degree of positive evaluation,which will promote the purchase choices of consumers.However,when this proportion reaches a certain range,the product sales decline.This is the same as expected.We consider higher proportion makes the product look too perfect,and it reduces the credibility of reviews instead.The experimental results show that consumers are able to distinguish fraud reviews to some extent.When the manipulation is obvious,consumers can identify the real situation of the product.

In column 3,we discuss the effect of number of fraud reviews on sales in another regression.The result conforms to our assumption that the number has a positive effect on sales.For consumers,a certain amount of fraud reviews could form good impression on product due to the higher rates and the longer length of fraud reviews.But when companies publish more fraud reviews,consumers will have a certain ability to recognize reviews and the sales will decrease,which is consistent with result from theFraudFractioni,k-1.

A summary of the results of all the hypotheses tests are provided in table 5.

V.DISCUSSION AND LIMITATION

This study contributes to both theory and practice.We have established a theoretical framework to understand the content of fraud reviews.From the perspective of the characteristics of the reviews and the price of the product,the incentives to manipulate fraud reviews and the effect of fraud reviews on the consumer decision process in Amazon's book market are studied.The review features include review valence,volume and the review length.

Table V.Summary of findings.

As a special type of online reviews,fraud reviews are deliberately published by suppliers,publishers or writer.The study finds that the prior reviews and the piece of product have an impact on the manipulate fraud reviews.The incentives of fraud reviews are less online reviews posted by real consumers,higher proportion of negative reviews,longer average length of negative reviews,lower average rating scored by real users and higher price.In these situations,more fraud reviews would be manipulated by businesses.Then,the release of fraud reviews effect consumers.The number,proportion,average length of fraud reviews and the promotion on rating have effects on consumers to make purchasing decisions.Our results also show that consumers can judge the authenticity of online reviews and discriminate fraud reviews when the manipulations get to a certain level.

In this study,the relevant research of fraud reviews is given from new insights.It can be concluded the review environments have an impact on the manipulation of fraud reviews and the fraud reviews and normal reviews affect the purchase behavior of consumers at the same time.Fraud reviews in themselves promote the sales of products.

The result also verifies the conclusions by previous researches and it is a useful supplement.Research shows that the number of reviews,ratings,review length and price affect the willingness to buy.Reviews have great potential value to companies,including increased sales [3,4,29].The manipulation of fraud reviews similarly has a significant impact on sales.The fraud reviews provide consumers with fraud information about products,services and other useful messages.When this manipulation is in a certain extent,consumers cannot discern qualities of reviews.Fraud reviews increase the credibility of product from the quantity and content,thus the approvals are improved and the sales rise.But the products with manipulate fraud reviews are usually those with low ratings and poorly evaluations.The fraud reviews change unfavorable emotions and affect the reference value of reviews.The real product quality is concealed,which misleads the consumers' consumption judgments,and makes the potential consumers make risky consumption decisions.Products with manipulation often have worse feedback by real users.The fraud reviews from the number and the content change the actual experience of reviews,which increase sales.This behavior undermines the shopping environment of the electricity business platform,and damages the legitimate rights and interests of consumers.Although consumers can judge some fraud reviews and the sites have begun to attack the behavior of manipulation,using certain computer algorithms to screen out suspicious reviews.This study finds there are still many reviews manipulated.Consumers cannot effectively identify all the goods.It is hoped that this study will enable enterprises to pay attention to this manipulation,to better regulate the goods and online reviews,providing consumers with a good online shopping environment and order.

This study tries to meet the scientific principles in theoretical deduction and empirical analysis,but there are still some limitations due to various problems.According to the objects and results of previous researches,we solely choose the book market as the research object.However,in fact,many experience products have the phenomenon of manipulation,while the search goods also have fraud reviews.In this study,the sample selection is single,and the results can not reflect all products of the website.In future research,product type selection can be increased and get more accurate results from full site products.