Electronic Commerce Research and Applications 12 (2013) 299–308

Contents lists available at ScienceDirect

Electronic Commerce Research and Applications

journal homepage: www.elsevier.com/locate/ecra

Trust fraud: A crucial challenge for China’s e-commerce market

Yu Zhang *, Jing Bian, Weixiang Zhu
School of Information Science and Technology, Zhejiang Sci–Tech University, Hangzhou 310018, China

a r t i c l e i n f o a b s t r a c t

Article history: Currently, China’s e-commerce market is growing at an unprecedented pace, however, it is faced with
Available online 13 December 2012 many challenges, among which the trust fraud problem is the biggest issue. In this article, we use Taobao
as an example and conduct a thorough investigation of the trust fraud phenomenon in China’s e-com-
Keywords: merce market. We present the development history of trust fraud, summarize its unique characteristics,
China and explore the reasons why so many sellers commit fraud. We further propose a dynamic time decay
e-commerce market trust model that aims to deter trust fraud by raising its cost and promote the growth of small and med-
Fraud
ium-sized sellers. The model utilizes detailed seller ratings as the data source, and incorporates a trans-
Taobao
Time decay
action amount weight, a time decay coefficient, and three trust factors in the calculation of trust. We test
Trust fraud the model on real transaction data from Taobao, and the experimental results verify its effectiveness. Our
Trust model proposed trust model yields a practical approach to online trust management not only in the Taobao mar-
ket but also for other e-commerce platforms.
Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction uses stars to denote seller feedback scores. As a seller’s feedback

score increases, his star will change color.) On Taobao, each level
In recent years, China’s e-commerce market is developing very includes a scale of 1 to 5. For example, a trust score of at least 4
rapidly. The total e-commerce sales increased to RMB5.88 trillion earns a seller one heart, and a trust score of at least 251 earns a
(US$944 billion) in 2011, a year-on-year increase of 129.2%, which seller one diamond. If a seller’s trust score reaches 10,001, her trust
accounts for 4.32% of total retail sales in China (The Beijing News level will be upgraded to one crown. If her trust score continues to
2012). According to an analysis by the Boston Consulting Group, grow and reaches 500,001, then she will receive one gold crown.
by the end of the year 2015, the size of China’s e-commerce market Thus, the higher the trust score, the more trustworthy will the sell-
will possibly surpass that of the U.S. market and become the er be. This feedback system enables buyers to examine sellers’ pre-
world’s next e-commerce superpower (Walters et al. 2011). vious transactions and provides valuable reference for buyers to
On the surface, China’s e-commerce market appears to be flour- make purchase decisions. Reputable sellers are rewarded and
ishing, however, it faces many challenges, among which the trust may enjoy premium prices and a higher probability of future sales
fraud problem is the biggest issue. In this article, we utilize Taobao (Brown and Morgan 2006).
as an example to carry out an in-depth look at this problem, which Although there are sellers on eBay who manipulate trust by
has been ignored before. Taobao now is in a prime position, dom- purchasing or exchanging positive feedback ratings (BlackHat-
inating China’s consumer-to-consumer (C2C) market with a 90.3% World 2010, 2012), the feedback system has still proven to be quite
market share, followed by Tencent’s Paipai with a 9% market share. effective (Brown and Morgan 2006). Taobao has met a big chal-
To manage trust among buyers and sellers, Taobao has adopted lenge using a similar feedback system though. Compared to eBay,
a similar feedback system to eBay. On Taobao, each buyer can offer Taobao’s trust fraud problem is much more serious. Here’s why:
a feedback rating to the seller after completing a transaction. A po-
sitive rating raises a seller’s trust score by one point. A negative rat- 1.1. Trust fraud is widespread on Taobao
ing lowers a seller’s trust score by one point. A neutral rating does
not affect a seller’s trust score. The overall trust score is accumu- According to Taobao’s research, the highest percentage of de-
lated by adding these ratings together. tected fraud transactions accounted for about 47% of all the rated
Taobao makes use of ‘‘heart”, ‘‘diamond”, ‘‘crown”, and ‘‘gold transactions and the lowest percentage was nearly 9% during the
crown” to denote different seller trust levels. (In contrast, eBay period from October 17, 2008 to May 17, 2009 (Chen and Yang
2009). It is not only the small sellers who artificially boost their
reputations. Also there are super sellers who commit trust fraud
* Corresponding author. to make their businesses look prosperous so as to attract more
E-mail addresses: [email protected] (Y. Zhang), [email protected] (J. Bian), zhwx- customers. A seller will feel left behind if all of her competitors
[email protected] (W. Zhu).

1567-4223/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.elerap.2012.11.005
300 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308

artificially increase their reputations when the seller does not. A the period from June to December 2005 (Brown and Morgan,
gold-crown seller, with a score of 500,001 points on Taobao, con- 2006). Even at the end of year 2008, it was still easy and cheap
fessed to a reporter that half of his trust scores before getting the to manipulate trust on eBay (Dini and Spagnolo 2009).
gold crown title were all obtained by committing trust fraud (Mon- Nowadays, eBay does not allow sellers to include the word
eyWeek 2011). Estimates suggest that about 80% of Taobao sellers ‘‘feedback” or to make any reference to eBay feedback in a listing
have committed trust fraud during the process of their businesses, title except when that word is used to describe the item for sale
more or less (MoneyWeek 2011). (eBay 2012). As a result, the public feedback market has disap-
peared, however, trust fraud still exists. Fraudulent sellers make
1.2. Trust fraud breeds a huge gray industry posts in forums or websites outside of eBay to look for partners
who are willing to exchange positive feedback ratings. They dis-
Since many sellers hope to promote their reputations as quickly cuss fraud strategy in posts or through private email messages
as possible, some professional scammers see this as an opportu- (BlackHatWorld 2010, 2012). If fraudulent sellers commit fraud
nity. They have begun to offer services to artificially increase sell- in a careful and secret way, the results of their actions are very
er’s trust scores. The part-time jobs of scammers now have turned hard to detect.
into a large gray industry, in which sellers are the core, and buyers, Ba (2001) has pointed out that online trust can be established
scammers, trust fraud companies and logistics companies are all through a community responsibility system. In addition, Ba and
involved on the periphery. The interests of all sides are closely re- Pavlou (2002) have investigated the extent to which trust can be
lated and quite complicated. According to the available statistics, induced by proper feedback mechanisms, and how some risk fac-
there are about 1,000 active trust fraud companies that are operat- tors play a role in the process of trust formation. Later, Ba et al.
ing inside the Taobao ecosystem at present. Among them, there are (2003) proposed a trusted third party who assigns a digital certif-
three to five that manage to earn millions of RMB each year (Bei- icate to each participant in online auction communities. The ana-
jing Youth Daily 2009). Suppose that there are four million active lytical results demonstrate that a market participant should
sellers on Taobao and a quarter of them pay RMB500 (US$80) behave honestly to maximize her benefits.
annually to artificially boost their reputations. This creates Zacharia and Maes (2000) proposed a centralized trust model
RMB1 billion (US$160 million) in the trust fraud market. SPORAS, which introduces a reputation mechanism for loosely con-
nected online communities. Kamvar et al. (2003) presented Eigen-
1.3. Trust fraud corrupts China’s e-commerce Trust, a distributed and secure model to compute global trust
values based on power iteration. The algorithm aggregates the nor-
Since it takes little time and effort to increase reputation by malized local trust values by a weighted sum of all raw reputation
committing fraud, many fraudulent sellers rely on this shortcut ratings. In Ramchurn et al. (2003), a concrete computational trust
and gradually deviate from the principles of conducting honest model was developed, which took into account both the direct
business. They are anxious to achieve quick success and get instant and indirect interactions between agents.
benefits. A fraudulent seller’s reputation can increase extremely Wang and Vassileva (2003) proposed a Bayesian network-based
fast. In recent years, many sellers have committed trust fraud on trust model to combine different aspects of trust. Yu and Singh
Taobao to boost their reputations, which leads to reputation infla- (2000) presented a social mechanism for reputation management
tion: all sellers have high trust scores eventually. Buyers are al- which aims at avoiding interaction with undesirable participants.
ready aware of this feedback-based trust-enhancing strategy. Liang and Shi (2005a, 2005b) developed a trusted middleware for
They have doubts about the effectiveness of Taobao’s feedback sys- P2P applications by combining two models. PET, a personalized
tem as a result. Some of them do not trust the system any longer, trust model, employs reputation and risk evaluation, and M-CUBE,
which in turn reduced buyer-seller trade in this marketplace. Some a multiple-currency based economic model, laid a foundation for
particiants may stop shopping online and will not come back again. resource-sharing in untrustworthy peer-to-peer (P2P) computing
In this article, we mainly focus on the trust fraud problem in environments.
China’s e-commerce market and propose a new dynamic trust Xiong and Liu (2003, 2004) proposed PeerTrust, an adaptive
model. This article is laid out as follows. Section 2 provides a liter- trust model for quantifying and comparing the trustworthiness of
ature review. Section 3 discusses the trust fraud phenomenon, and peers. The model is based on a weighted sum of five factors includ-
analyzes its characteristics and the reasons behind its develop- ing feedback records, feedback scope, credibility, transaction con-
ment. Section 4 presents our proposed dynamic trust model, which text and community context. Bizer and Oldakowski (2004)
aims to decrease the chances of online trust fraud and promote the outlined a trust architecture that has trust policies combining rep-
growth of small-sized and medium-sized sellers. Section 5 pre- utation, context and content-based trust mechanisms.
sents our experimental results and proves the effectiveness of Other related work has touched on trust in recommender
our proposed trust model. Section 6 concludes. systems. Donovan and Smyth (2005) have indicated that the trust-
worthiness of users plays an important role in guiding recommen-
2. Literature Review dations and they presented two computational trust models. The
experimental results showed that their models managed to im-
There has been a lot of research on the e-commerce trust across prove the predictive accuracy during recommendation. Andersen
various disciplines, including Economics, Management, Computer et al. (2008) used an axiomatic approach from the theory of social
Science and Sociology. Brown and Morgan (2006) and Dini and choice and proposed a recommender system in which agents can-
Spagnolo (2009) have shown that fraudulent sellers artificially en- not induce others to share their opinions by lying or modifying
hance their reputations by trading positive feedback ratings on their trust links. Massa and Avesani (2004) have advocated a
eBay. These sellers listed items such as e-books, jokes or riddles trust-aware collaborative filtering algorithm to increase the cover-
for a very low price, like a penny, inclusive of shipping fees. The age of recommender systems while providing more accurate rating
words ‘‘positive feedback” may be included in the title of an auc- predictions.
tion or hidden in the text of its description, which suggests that Compared with previous works, our contribution is threefold.
this is a sale or an exchange for feedback ratings. More than First, we thoroughly investigate trust fraud in China’s e-commerce
6500 listings on eBay involving feedback trade were identified over industry, and reveal the seriousness of this problem. Second, we
give a detailed introduction on the historical evolution of trust
 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308 301

fraud methods and summarize their distinctive characteristics. dresses and fake accounts. Usually, brushers asked Taobao sellers to
Third, we propose a new dynamic trust model to lower the chances list virtual products in their storefronts, such as cards to add mo-
of trust fraud. We will verify its effectiveness through experiments bile phone minutes, electronic documents and Q coins (the virtual
using real transaction data. Our proposed trust model will have money used to buy virtual items in the online instant messaging
application value in the real world, and can help to promote the community known as QZone). Then, brushers used those fake ac-
management of trust in China’s e-commerce industry. counts to hypothetically purchase virtual products that did not re-
quire logistics. After the fake transactions were made, the software
automatically left positive ratings and good comments for the sell-
3. Trust fraud on Taobao ers. As a result, the sellers’ trust scores grew very rapidly. The sec-
ond-generation trust fraud method has obvious regularities
We next will give a detailed introduction about trust fraud on though, such as buyers’ accounts sharing similar naming rules,
Taobao, by first presenting the history of trust fraud’s develop- and the interval between purchases and confirmations are typi-
ment. Then we will summarize its distinctive characteristics. We cally very short. As a result, this method was also easy to detect,
also will explore the reasons why so many sellers have chosen to and so it is rarely used any longer.
commit fraud on Taobao.
3.1.3. Third-generation trust fraud: trust fraud between two sellers
3.1. History of trust fraud development The third generation of trust fraud method involves cooperation
among fraudulent sellers. They set up online group chats using in-
During the first few years after its foundation, Taobao was busy stant messaging software and communicate with each other with-
competing with its biggest rival – eBay China. At that time, Taobao in messaging groups. Then they work together to obtain
mainly concentrated on attracting more sellers and buyers to its undeserved trust scores. The principle of this method is as follows.
marketplace. The trust issue was not its main focus. Nowadays, Within the group, seller si purchases a RMB10 product from seller
Taobao is dominating China’s e-commerce market, however, trust sj, and seller sj purchases an equal value product from seller si. They
fraud is flourishing and Taobao members are at risk of developing do not actually deliver the two products. The RMB10 transaction
a culture of pursuing short-term financial gain at all cost (Epstein total transfers from seller si to seller sj, then comes back to seller
2011). si. Both seller si and sj incur no financial cost other than their time;
When Taobao was just starting up, if a seller’s trust level was however, both sellers’ respective trust scores increase.
one diamond, then he was regarded as trustworthy enough in most When analyzing the transaction records among a group of sell-
buyers’ opinions. At this time, quickly obtaining at least a one-dia- ers, it is easy to detect this kind of trust fraud since almost every
mond reputation on Taobao was the primary objective of new sell- seller within the group has transaction connections with each of
ers. Trust fraud on Taobao, as a result, is usually called brush the other sellers. To hide from Taobao’s detection, a fraudulent sell-
diamond in Chinese. The words vividly describe the phenomenon er needs to join new instant messaging groups every day to have
that a new seller with few or no transactions surprisingly turns many different fraudulent partners with various IP addresses. He
into a one-diamond seller overnight, just like a magician uses a will not know much about those partners though. So he also will
magic brush to create diamond out of nothing. In fact, the im- not be sure whether his partner will be careful enough to change
proved reputation is not acquired through real transactions but his IP address or clear all cookies when committing fraud with
by fake trading. Scammers who live on artificially-boosted reputa- him each time. If any of his fraudulent partners does not follow
tions for sellers are called brushers in Chinese. the so called ‘‘safety rules” for fraud, then he might be detected
Although Taobao was founded four years later than eBay, trust and punished by Taobao, and all his previously-obtained unde-
fraud techniques developed much faster on Taobao. In the mid- served positive ratings will be confiscated. Once one fraudulent
2000s, Taobao became aware of this problem and began to put in seller is caught, many others who previously have had fake trans-
a lot of effort to detect online trust fraud. Now it has become a com- actions with him will have a difficult time to escape discovery.
petition though. Professional brushers try their best to find system Aside from being caught by Taobao, there are also other risks
vulnerabilities and continue to upgrade the techniques they use. associated with using an instant messaging group for trust fraud.
As a result, trust fraud methods on Taobao have evolved across five For example, suppose that seller si purchases a product from seller
different generations. Each generation has surpassed the preceding sj, and then seller si confirms the transaction and seller sj receives
one, and it gets harder and harder for Taobao to detect such fraud. the payment. It is then seller sj’s turn to purchase seller si’s product,
however, he may refuse to pay. Due to the lack of guarantee and
3.1.1. First-generation trust fraud: trust fraud by the sellers themselves the fact that their agreement is not protected by law, seller si can
This is the most elementary way for fraudulent sellers to obtain do nothing in this situation. There are a great many scammers
undeserved positive ratings. Basically, they set up multiple ac- who specially cheat such sellers.
counts on Taobao. Then they use these accounts to purchase items
from themselves, and in the end leave themselves positive ratings 3.1.4. Fourth generation trust fraud: trust fraud among multiple sellers
and good comments. Using this method, fraudulent sellers make To avoid detection by Taobao, a more advanced cooperation
themselves look trustworthy. It is very easy for Taobao to detect pattern among multiple sellers emerged as the fourth-generation
these fraudulent sellers since their multiple accounts use the same trust fraud method. Suppose there are five sellers within the group.
IP address and the cookies also indicate multiple ‘‘buyers” have Seller s1 hypothetically purchases a product from seller s2, seller s2
logged in from the same computer. hypothetically purchases a product from seller s3, seller s3 hypo-
thetically purchases a product from seller s4, seller s4 hypotheti-
3.1.2. Second-generation trust fraud: trust fraud by machines cally purchases a product from seller s5, and at last seller s5
The first-generation method is usually used by sellers to pro- hypothetically purchases a product from seller s1. The larger the
mote their own reputations. From the second generation onward, group, the less likely the group’s activities will be detected by Tao-
trust fraud on Taobao became a new business. Professional brush- bao. Since multiple sellers take part in fourth-generation trust
ers started making a living by committing fraud on Taobao full- fraud activity, this method is often used among people who are ac-
time. This was also the beginning of large-scale online trust fraud. quainted with one another already. If this is the case, the risk of
Brushers began to utilize computer software to generate fake IP ad- being cheated by other sellers is greatly reduced, however, the risk
302 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308

of being discovered by Taobao will increases since group activities 3.2.1. Good organization
among acquaintances tends to leave clues that can be followed. Trust fraud activities on Taobao are well organized and are not
one man’s act. Among all the brush diamond companies, nearly ten
3.1.5. Fifth-generation trust fraud: trust fraud through a third-party of them have more than fifty employees. There is a clear division of
platform labor among these employees too. Some are in charge of technical
At the end of 2006, the fifth-generation method involving the support, some are responsible for advertising and promotion, and
first third-party brush diamond platform emerged. The basic princi- some are in charge of management.
ple of the platform is as follows. Sellers all log into the platform
with registered user names. Suppose seller si wants to obtain a po-
3.2.2. Professionalism
sitive feedback rating by falsely selling an RMB10 product, then si
Many people who work in this gray industry are professional
publishes this task on the platform. The platform deducts RMB10
software engineers. They develop brush diamond platforms, update
from si as a deposit. Another member u takes over the task and pur-
functions and maintain security, etc. Since they choose brush dia-
chases the product from si’s storefront on Taobao using his own
mond activities as their ‘‘careers”, they are motivated to try their
money. Then he confirms the receipt of the product and offers si
best to discover system vulnerabilities and make use of them to
a positive rating on Taobao. Seller si receives RMB10 from u for
earn money. Over time, they tend to adopt more subtle means to
the fake transaction. Afterwards si’s RMB10 deposit is transferred
commit fraud, and so it gets harder and harder for Taobao to detect
to u’s account on the third-party platform. Meanwhile, member u
their activities.
also gains bonus points from the platform. Money only is trans-
ferred from one account to another and is returned, however, both
the buyer and seller will have no financial losses. When there are 3.2.3. Blatant advertisements
enough active members on this platform, it is hard for Taobao to Brush diamond companies constantly send spam advertisements
detect such trust fraud activities, since the process of making fake to sellers through instant messaging software. An example is Aliw-
purchases is very similar to real transactions on Taobao. The angwang, which offers free services for online business communi-
so-called ‘‘buyers” and ‘‘sellers” come from all over China and there cation to buyers and sellers on Taobao and Alibaba. A second
actually are real cash flows that occur among them. example is QQ, a popular free instant messaging application in Chi-
If member u himself owns an online shop, then he may do the na. When you search for the term ‘‘Taobao brush diamond” in Chi-
same thing as seller si does and gets undeserved trust scores for nese using the two most popular search engines in China, Baidu
free. Certain bonus points will be deducted from u’s account for returns about 2,710,000 hits, while Google returns around
publishing a task on the platform. So if a seller wants to get posi- 5,070,000. Baidu topped China’s search engine market with a
tive feedback ratings for free, he must accept tasks and fulfill them 78.3% share in the fourth quarter of 2011, while Google China
first. In this way, he can obtain the desired bonus points for pub- ranked second with a 16.7% market share (China Knowledge Online
lishing his own task. If u is a professional brusher, then he will earn 2012), indicating the scope of the potential for fraud. For both
money from the platform by serving other sellers to artificially search engines, the advertisements of brush diamond platforms
boost their reputations. The platform earns money from sellers dominate the top few positions until this article go to press. It is
who want to get undeserved trust scores but do not want to take interesting to note that brush diamond companies can publicly
over other sellers’ tasks. These sellers just pay a fee to the third- place such advertisements in China without fear of reprisals
party platform and the platform will be in charge of all the rest against them.
of the work. Present, a one-diamond reputation costs a seller about
RMB150–RMB300, while a one-crown reputation costs a seller 3.2.4. Business terrorism
about RMB3500–RMB7000 (Tianshan Net 2012). In recent years, Taobao took measures to combat trust fraud by
‘‘All for one and one for all” is the motto of brush diamond plat- modifying its rules and deleting sellers’ fake feedback ratings. But
forms. Ironically, these platforms which aim to commit trust fraud, the brush diamond companies and fraudulent sellers rebelled
also claim that they will guarantee the safety of customer deposits, against Taobao in various ways. For example, they maliciously
as well as the brush diamond process. It is fairly that the executives made massive online orders from large sellers but refused to pay,
of these platforms escape and take away all their customers’ and they made crank calls to Taobao’s customer service center,
deposits though (Guangzhuo Daily 2011). The sellers who still have which seriously interfered with the firm’s normal work (Beijing
deposits on the platforms lose heavily. Their interests are not pro- Business 2009). There also are recorded instances of violence
tected by law, since their behavior is in itself improper. against Taobao to force it to compromise. This is not just a simple
Despite these concerns, brush diamond platforms still currently protest, but rather it has become a form of business terrorism.
enjoy a booming business. There could be three reasons for this
phenomenon: first, there is no clear and direct law against brush 3.3. Why sellers commit trust fraud
diamond behavior, so it is very hard to trace and punish those pro-
fessional brushers. Second, unlike eBay, Taobao is totally free with- Studies on eBay have shown that the market rewards a seller
out any insertion or transaction fee. Therefore, brush diamond who has accumulated a lot of positive feedback ratings (Resnick
activities on Taobao essentially involve a zero-cost business. Third, et al. 2003). When a seller’s trust score achieves 50 or greater, he
there are still a certain number of sellers who want to get unde- will see an additional 10% increase in views and prices (Gilberg
served trust scores. As brush diamond companies have a market 2012). Based on our experiments on Taobao, seller trust levels also
and manage to make money, they continue to exist and develop. appear to have a positive correlation with their recent transaction
amounts (Zhang 2012). Profit, not surprisingly, is the fundamental
3.2. The characteristics of trust fraud on Taobao reason for trust fraud.
During our investigation, we interviewed several sellers who
Trust fraud is not a problem that is specific to Taobao. As the had engaged in trust fraud activities. On the one hand, they were
world’s largest C2C e-commerce trading platform, eBay also has worried that their behavior might be discovered and punished by
trust fraud problems of its own (Nickson 2010). Compared to eBay Taobao in the future. On the other hand, they felt they had no
though, Taobao’s problem is more serious and has its own unique choice but to commit fraud. At present, there are more than 6.35
characteristics. million sellers involved in typical online activities, and there is
 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308 303

fierce competition among them.1 When a buyer searches for a prod- model; large-scale sellers with high trust scores should not be
uct on Taobao, there will be up to 100 pages of items that may be able to rely on their previous reputations. Sellers’ trust scores
retrieved. Usually, an overwhelming majority of buyers will choose will change along with their recently obtained feedback ratings.
a seller from among the ones listed on the first few pages. A combi- In this way, we are able to encourage the development of small
nation of factors such as a seller’s trust level, sales volume and pop- and medium-sized sellers.
ularity will determine her ranking in the search results. As a new
Make trust hard to build, but easy to lose. There is a saying
seller or a small seller, she will not have any of the aforementioned attributed to the famous Japanese thinker, Daisaku Ikeda:
advantages. Therefore, it will be very difficult for such sellers to com- ‘‘Trust is difficult to earn and it is easily lost – the trust built
pete with large sellers. And, even if small sellers sell a product at a over a decade can be shattered in an instant by one offhanded
very cheap price, many buyers still will be afraid to purchase any remark or deed” (Ikedaquotes.org 2012). Currently on Taobao,
products from them. Without a high trust score, the seller will only it is easy for sellers to gain a positive feedback rating after a
sell a few products. Without good sales, small sellers will not be able transaction. According to our analysis, the ratio of positive rat-
to get a very high trust score. This situation creates a vicious circle. ings is more than 99.5% among all the valid feedback ratings,
One seller we interviewed told us that he was not intending to com- based on randomly-chosen observations from a much larger
mit fraud or cheat buyers. He just wanted to let them have a chance data set containing over 1.3 billion feedback ratings. For super
to notice his storefront, and this enabled him to make a living. This sellers, a negative feedback rating has little effect since it only
view is widely shared among sellers in the competitive marketplace leads to losing one point from the total trust score. In addition,
of Taobao. this negative impression is likely to be counteracted by the
To sum up, due to the lack of a sound social credibility system, many positive ratings that will follow. A more effective trust
trust fraud in China’s e-commerce industry is quite complex and model should reflect the trust characteristics as we know them
has its unique characteristics. Some of the problems we discussed in real life – that trust is hard to build but easy to lose. The sell-
are unprecedented and have not been addressed in prior e-com- ers should be required to have good long-term performance to
merce research. In addition, there are no ready-made or obvious obtain a high trust score, and they should lose a lot once they
solutions to them at present. act deceptively in a transaction.

Improve the authenticity of feedback ratings. The authentic-
4. A new dynamic trust model for China’s C2C e-commerce ity of feedback ratings is the primary basis for accurate trust
evaluation. Even if there were no brush diamond activities on
We next will present a new dynamic trust model to attempt to Taobao, current feedback ratings still would not truly reflect
combat the trust fraud problem on Taobao. We will introduce sev- buyers’ opinions about transactions. Resnick and Zeckhauser
eral directions for improvement to the current C2C e-commerce (2002) found that buyers and sellers reciprocate and retaliate
trust models. We also will propose a new dynamic trust model that in regard to feedback ratings. According to a survey conducted
incorporates a transaction amount weight, a time-decay coefficient on Taobao, 44.2% of respondents reported that the biggest prob-
and several additional trust factors. lem related to feedback is that sellers will continue to harass
them if they offer a negative feedback rating. Under this situa-
4.1. Improvement directions for current trust models tion, buyers are not willing to offer negative feedback ratings,
so the authenticity of feedback cannot be guaranteed. A more
Based on our study of trust fraud on Taobao, we propose a num- effective model of trust should make use of anonymous feed-
ber of directions for improvements to the current trust models. back ratings instead of ones that are associated with the names
of others, so that sellers cannot see which of buyers is giving

Raise the cost of trust fraud. The traditional method to elimi- them which ratings. As a result, buyers will feel free to express
nate trust fraud is to identify those sellers who have defrauded their true feelings without worrying about threats of reprisals
and then punish them. However, this is clearly a temporary or harassment. In addition, sellers will have the necessary
solution: it only cures the symptoms, not the disease. A better incentives to provide good products and services each time they
method is to build a more reasonable trust model and control are involved in a transaction.
trust fraud by improving regulations and rules. The basic strat-
egy is to raise the cost of trust fraud, so that more and more 4.2. A new dynamic trust model
brushers think it will be unprofitable, and as a result they will
quit (Chen and Yang 2009). A trust system that only reduces We now propose a dynamic trust model for Taobao. It includes
the lure of fraud without eliminating it may be sufficient to acti- the following parameters.
vate Yhprum’s Law – ‘‘everything that can work will work”
(Resnick et al. 2003). The founder of the brush diamond industry
Transaction amount weight. Various transaction amounts should
and previously the largest brush diamond company in China was be treated differently in the trust calculation.
WebGle. Today, WebGle serves as a platform that provides
Time decay coefficient. Trust should reflect sellers’ recent trading
Internet third-party management services. It exited the brush activities.
diamond business in March 2009. High service costs and low
Trust factors. These include quality, service and shipping speed,
margins were the main reasons for WebGle’s decision to exit which should be considered and treated differently according to
(Beijing Youth Daily 2009). the buyer’s opinions.

Build a dynamic trust model. According to research in Sociol-
ogy, trust is dynamic and non-monotonic (Abdul-Rahman and We utilized real transaction data records from Taobao to deter-
Hailes 2000). Thus, current models that view trust as a static mine the above mentioned parameters. The sample data set con-
accumulated amount should be replaced by a dynamic trust tains general information about 3,261,269 transactions. The
general information about a transaction includes the buyer ID,
1
the seller ID, the product price, the sales volume and the total va-
This number was obtained based on statistics from Taobao’s backend database.
Some non-active sellers haven’t logged into Taobao for a long time, some have closed
lue. There are 15,976 sellers. The general information about a seller
their online shops, and others are under investigation for alleged fraud. We removed includes his trust level, trust score and the opening time of his on-
these kinds of sellers to obtain the number of active sellers reported here. line shop. The data set also contains 2,308,704 feedback ratings.
304 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308

Buyers can leave a positive, negative, or neutral rating for sellers.

There are also 3,306,633 detailed seller ratings. The detailed seller
rating system is based on a 1–5-star scale. 5 stars is the highest rat-
ing, and 1 star is the lowest. A detailed seller rating reflects three
specific aspects of the transaction: item description, seller’s service
and shipping time. The time span is from March 9 to June 10, 2010,
covering three months in total. To build a new trust model, we first
present how to determine the three above-mentioned parameters:
transaction amount weight, time decay coefficient and trust
factors.

4.2.1. Transaction amount weight

Currently on Taobao, no matter how much the trading amount,
one positive rating only gets a seller one point. Suppose a buyer bk
purchases a RMB10,000 (US$1,605) product from seller si and pur-
chases an RMB1 (US$0.16) product from seller sj. If the buyer bk is
satisfied with both transactions, sellers si and sj obtain the same
one point. In fact, the risks associated with the two transactions Fig. 1b. Transaction amount distribution in sub-ranges of (1–100].
differ greatly. In most cases, buyers will carefully select reliable
sellers from whom to purchase valuable products. They will not
the proportion of the last sub-range (90–100] is a little more than
carefully scrutinize sellers before purchasing cheap products
that of the sub-range (80–90].
though. Therefore, an effective trust model must recognize the
We can set the transaction amount weight as shown in Table 1
appropriate weight to apply based on each transaction’s total
according to the distribution of transaction volumes and the Pareto
value.
principle to establish the basic system settings.
We propose to set a transaction amount weight m according to
First, we exclude the transactions of virtual products and the
the risk associated with each transaction. The transaction risk itself
transactions whose amounts are less than or equal to RMB1 from
is a quite complicated issue, which may involve many factors dur-
trust calculation. Previously, many brush diamond companies uti-
ing the transaction process. We can take a roundabout approach to
lized these kinds of transactions for trust fraud because the cost
evaluate the risk by observing the buyers’ behavior. Buyers’ online
was very low. Since the majority of transactions on Taobao are
behavior indirectly reflects their assessment of risk. As we all
for small amounts, we mainly consider how to set the trust weight
know, avoiding risk is a matter of human nature. If the majority
for transactions whose amounts are between RMB1 and RMB1000.
of buyers choose to do something, then the risk is likely to be small.
We divide the range (1–1000] into three ranges according to the
Conversely, if only a small number of buyers choose to do some-
distribution of transaction amounts. (See Figs. 1a and 1b again.)
thing, then the related risk is likely to be larger.
We then set the weight using the various aspects that are dis-
Fig. 1 shows the transaction amounts distribution of records in
cussed below.
sample data set. The x-axis denotes the range of transaction
amounts and the y-axis denotes the proportion of each range.
4.2.2. Time decay coefficient
The general trend is that the lower the price of products, the more
Trust is a dynamic concept. The closer the transaction happens
transactions will be observed. The higher the price, the fewer
to the current time, the greater the reference value should be for
transactions will be observed. As can be seen from Fig. 1a, most
trust calculation. The trust score should not be a simple accumula-
transaction amounts on Taobao are below RMB100 (US$14.6),
tion of feedback ratings though. Instead, it should be adjusted con-
which accounts for nearly 81.5% of the transaction records in the
tinuously as time passes. Thus, we introduce a time decay function
sample data set. We subdivided the range (1–100] into 10
for our proposed trust model in this section.
sub-ranges (see Fig. 1b) The percentage decreases gradually from
We utilized the power function to obtain the time decay coeffi-
(1–10] to (90–100], from one sub-range to another except that
cient, as shown below:

dtj ¼ abðtc tj Þ ð1Þ

90.00%
Here tc denotes the current time and tj denotes a time in the past
80.00% when a buyer offered a feedback rating. dtj is the time decay coeffi-
cient for a specific time tj. The variables a and b are parameters of
70.00%
the time decay function. We are able to obtain different decay ef-
60.00%
fects by adjusting the two parameters (Zhang and Zhu 2011). The
values of a and b can be determined by combining empirical knowl-
50.00% edge and practical experience. Fig. 2 shows the curves of different

40.00%

30.00%
Table 1
Transaction amount weight.
20.00%
Transaction amount Weight
10.00% 61 0
(1–100] 1
0.00% (100–200] 2
(
(0-1] (1-100]
( ((100-200] ((200-1000] ((1000- +∞) (200–1000] 3
>1000 4
Fig. 1a. Transaction amounts distribution in different ranges.
 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308 305

combinations of a and b. When the base a is determined, the smal-

ler the b, the slower trust decays.
When parameter b is determined, the larger the base value of a,
the faster trust decays. We can obtain various decays effects from
drastic drops to gentle declines by tuning the two parameters. This
function is simple to create and modify. Below, we chose one seller
as an example to illustrate the time decay effect. Here, we let a = 2
and b = 0.1. The related time decay coefficient can be calculated as:
dtj ¼ 20:1ðtc tj Þ .
Fig. 3 shows the result, where the x-axis denotes the number of
weeks and the y-axis denotes sellers’ trust scores.
For the period of sixteen weeks, the blue line denotes the sell-
er’s sales each week. For example, the seller’s transaction volume
is 44 in the fifth week and is 26 in the sixth week. The red line
shows the cumulative trust score after time decay under normal
sales circumstances. To clearly illustrate the effect of the time de-
cay function, we also simulate two extreme situations: one where
the seller is good at first, however, he becomes worse and worse
later in sales and service. The other one is just the opposite: the
Fig. 3. The effects of the time decay function.
seller behaves badly at first, however, he gets better and better
with time. The green line represents the former, while the purple
line represents the latter. As can be seen, if a seller relies on his ori-
ginal good reputation and does not provide as good a product or 4.2.3. Trust factors
service as before, his cumulative trust score will grow very slowly In our proposed model, we mainly utilize detailed seller ratings
or will even decrease. Meanwhile, if the seller continually im- that contain three trust factors: quality, service and shipping
proves product quality and service, his trust score will grow rap- speed. The importance of these trust factors in the trust calculation
idly. Using the time decay function, sellers will be incentivized to differs. According to a survey on Taobao buyers (Taobao 2011),
have continuous good performance. It will also punish sellers about 71.4% of buyers pay close attention to product quality, about
who begin to behave badly after obtaining a high trust score 35.7% of buyers are concerned about service and about 28.6% of
(Zhang and Zhu 2011). buyers pay attention to the shipping speed. According to the above
Fig. 4 illustrates the effect of weighted time decay on sellers of information, we set the trust factor weights, as shown in Table 2.
various trust levels from ‘‘1 heart” to ‘‘1 crown.”
The first column on the left shows the weighted method with- 4.3. A dynamic trust model
out time decay. All the remaining columns represent the trust cal-
culation using different time decay coefficients. The x-axis We let si (1 6 i 6 N) denote a specific seller. We let B denote the
indicates the decay circle r, while the y-axis indicates the percent- set of buyers who have had transactions with seller si, and let bj
age of sellers with various trust levels. For example, the second col- (1 6 j 6 H) denote a specific buyer. We use m to denote the trans-
umn from the left denotes the result when r = 730 days, and the action amount weight. We utilize wq to denote the quality weight,
rightmost column denotes the result when r = 7 days. As the decay ws to denote the service weight, and we to denote the shipping
circle diminishes, the number of low-trust sellers increases and the speed weight.
t
number of high-trust sellers decreases. Next, let qbkj ! si represent the rating that bj gives to si at time tk
t
From these experimental results, we can come to the following for the product quality. In addition, let pbkj ! si be the rating that bj
t
conclusions. First, using the time decay function, the overall com- gives to si at time tk for the service. Finally, assign ebkj ! si as the
position structure of sellers is stable. Second, high-trust sellers can- rating that bj gives to si at time tk for the shipping speed.
not rely on their previous trust scores forever. Thus, the time decay We utilize dDt to denote the time decay coefficient correspond-
function can lower the market access threshold for new sellers. ing to the current time tc, where Dt = tc  tj. We also let NG be the
total number of rated transactions that seller si has up until the
present time tc. Note that buyer bj may have multiple transactions
with seller si and we use transaction record IDs to distinguish
among the different transactions. Therefore, a seller si’s cumulative
dynamic trust score at time tc can be calculated using the following
formula:
X
NG
t t t
T si ¼ m  ðwq  qbkj ! si þ ws  pbkj ! si þ we  ebkj ! si Þ  dDt
j¼1

ð2Þ
In addition, the average dynamic trust score per transaction can
be calculated as follows for a recent period of time:

T si
T si ¼ ð3Þ
NG
The value T si for a recent period of time will be an effective param-
eter to evaluate sellers’ current trust status. We will further explain
the details of this finding based on our experimental results in the
Fig. 2. Time decay curves. following section.
306 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308

Fig. 4. The effect of weighted time decay on sellers of various trust levels on Taobao.

Table 2 Table 3
Trust factor weights. Large-scale sellers’ average trust scores calculated using the dynamic trust model.

Trust factor Weight SellerID A–Trust Q–Ave S–Ave L_Ave T Si

Quality 50% 
7897 1157197 3.70 3.57 3.47 3.62
Service 30% 3344 484663 3.80 3.79 3.70 3.78
Shipping speed 20% 5008 244071 3.91 3.86 3.78 3.79
32 332266 4.20 4.22 4.03 4.17

5. Experimental results

In this section, we utilize real transaction records from the ori-

ginal sample data set for the experiment. First, we preprocessed Table 4
Small-scale sellers’ average trust scores calculated using the dynamic trust model.
the data and removed records that contained a null value in any
field. In the sample of detailed seller rating records, there were SellerID A–Trust Q–Ave S–Ave L_Ave T Si
314,927 fields containing null values. Then, we associated the re- 3372 260 16.48 16.74 16.48 16.56
cords in different tables to obtain all the needed information for 3463 91 15.92 15.92 15.92 15.92
each transaction, including seller ID, buyer ID, transaction total, 1330 555 15.84 15.84 15.53 15.78
price, detailed seller ratings and transaction time. In this way, we 5736 1331 15.97 15.10 15.97 15.71

were able to identify 1,224,635 transaction records to create the

cleaned data set for our experiment.
The cleaned data set shows the trading conditions for 9960 sell- We ranked all the 8003 sellers according to their dynamic aver-
ers in a three-month period. age trust scores. From the ranking, we observed an interesting phe-
Among these sellers, some had few transactions during that nomenon: many super sellers who already had trust levels such as
period and their records were not sufficient in number for trust ‘‘2 crowns” or ‘‘5 crowns’” have a low dynamic average trust score.
evaluation. This led us to select 8003 sellers from the cleaned data In contrast, a lot of small-sized and medium-sized sellers have a
set who made greater than or equal to 10 successful transactions. rather high dynamic average trust score.
We calculate the dynamic average trust score using Eq. (3) for Currently on Taobao, many super sellers rely on their existing
these sellers. reputations and feel it is not necessary to provide good products
Tables 3 and 4 show the average trust scores for several super or services any longer. They may receive a certain percentage of
and small sellers. negative ratings due to bad products or services, however, it has
We only present a small part of the experimental results. In little effect on their trust status since a negative feedback rating
both tables, the first column, SellerID, denotes the user ID of a sell- only leads to losing one point of their total trust scores. Moreover,
er, with the last four digits of their IDs omitted for privacy. The sec- this negative impression will be counteracted by the many positive
ond column, A_Trust, denotes the cumulative trust score of a seller. ratings that will follow. (Note that more than 99.5% of buyers offer
This is the trust score that appeared on his online shop’s home- positive feedback ratings to sellers even if they are not truly satis-
page, when we obtained the sample data. The third column, fied with the transactions.) Using our proposed model, the advan-
Q_Ave denotes the average score for product quality given by the tages of the super sellers will disappear and their average trust
buyers per transaction. The fourth column, S_Ave, denotes the scores will decrease gradually due to time decay. They will eventu-
average score for service given by buyers per transaction. The fifth ally end up at the back of the pack, as a less-preferred seller.
column, L_Ave, denotes the average score for shipping speed given In contrast, some small-sized and medium-sized sellers may
by buyers per transaction. The rightmost column denotes the aver- manage to have high dynamic average trust scores. They will stand
age trust score calculated using our dynamic trust model. out from millions of other sellers due to their continuous good per-
 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308 307

formance. As long as a seller provides good products and services,

he will get a high dynamic average trust score. This will be true no
matter whether the seller is large or small. This also is fair for all of
the sellers and will help to promote healthy market competition.
To examine the effect of our trust model, we browsed the
homepages of sellers’ storefronts on Taobao. For example, a small
seller who gets a high average trust score of 16.6 sells top-grade
electrical equipment. The average price of these products is above
RMB10,000 (US$1605), so the transaction amount weight is 4 in
the trust calculation. From the buyers’ comments, we can tell that
this seller only sold genuine branded products and provided very
good services in support of the transactions. As a result, the seller
received high detailed seller ratings in all the three categories.
Another example is a large-scale seller who sells all kinds of
snacks. The seller’s cumulative trust score was 1,157,197 when
we obtained the data from Taobao. Through reading buyers’ com-
ments, we could see that a certain percentage of buyers thought
the snacks were not delicious. Some buyers wrote that the pack- Fig. 6. The average dynamic trust scores of sellers on Taobao.
ages were broken when they arrived and the delivery speed was
very slow. We also noticed that some buyers even pointed out that
there were many food additives in the snacks, and so they were not We also rank all the above sellers according to their average
very healthy. Overall, many buyers were dissatisfied with this trust scores obtained by our proposed trust model. (See Fig. 6.)
super seller. Meanwhile, the prices of most products sold in the The x-axis denotes their rankings among all the sample sellers,
seller’s storefront were under RMB100 (US$14.6), so the transac- while the y-axis denotes the average trust scores obtained by our
tion amount weight is set to 1 in the trust calculation. As a result, dynamic trust model. There is a slight slope in the middle part of
this seller’s average trust score was rather low. the curve, which indicates that a majority of sellers belong to the
Currently, there is another problem that Taobao is dealing with: middle class among all the sellers. It gets harder and harder for
the trust score represents the accumulation of buyers’ feedback sellers to promote when they reach a score of about 12, so there
ratings after each successful transaction, so a seller’s trust score is a steep increase at the end of the curve. The highest average trust
grows continuously and it will not have an upper limit. Our model score of sellers in the sample is 16.6. It is very difficult for sellers to
can solve this problem, as the calculated average trust score will reach the full mark of 20.
have a defined upper limit. The highest detailed seller rating for
each aspect is 5. Suppose the transaction amount weight is 4, 6. Conclusions and future work
which is the highest value. Therefore, the final average trust score
will be 20, which is the highest average trust score that a seller can Based on the experimental results we have reported, our pro-
obtain using our dynamic model. It will be very hard for sellers to posed dynamic trust model supports the following conclusions:
reach this upper limit due to the time decay included in the trust
calculation. Sellers need to perform well consistently to have a
The closer that a transaction occurred relative to the current
high trust score. time, the larger the weight of its feedback rating will be in the
We select 5228 sellers from cleaned data set and ranked them trust calculation.
according to their cumulative trust scores, as shown in Fig. 5.
By adding the transaction amount weight, we have enlarge the
The x-axis denotes each seller’s ranking among all other sellers in original 5-star rating scale so it has a larger range of 20, which
the data set, while the y-axis represents their cumulative trust makes it easier for buyers to distinguish among different sellers.
scores. As can be seen from the results, the cumulative trust score
The average trust score in our proposed model has a defined
has no upper limit and the range of the trust score is extremely upper limit and will not grow in an unrestricted manner. All
large. sellers will seek to achieve the upper limit of the trust score,
however, it will be very difficult for them to achieve this goal.

The trust calculation rules in our model are easy to understand
and implement.

In the future, we will conduct more complete tests on our dy-

namic trust model using large-scale data from Taobao. Meanwhile,
we will consider more trust factors in the evaluation, such as the
history of the online shop, illegal behavior and customer com-
plaints, etc. We also plan to continue our in-depth cooperation
with Taobao. We further hope to develop a module based on our
proposed dynamic trust model, and apply it in one of Taobao’s on-
line applications to verify its effectiveness and efficiency.
To address the problems with trust in China’s e-commerce,
improving the trust model is an important internal solution, while
building a social credibility system in China is an external solution.
Social credibility, we believe, will have a great impact on the effec-
tiveness of a trust model in China’s e-commerce industry. That is
because Taobao’s problems are, in a very real sense, China’s prob-
lems (Epstein 2011). Each year, China’s economic loss caused by
Fig. 5. The cumulative trust scores of sellers on current Taobao. lack of trust reaches about RMB585.5 billion (US$94 billion)
308 Y. Zhang et al. / Electronic Commerce Research and Applications 12 (2013) 299–308

according to statistics of year 2011 (Qian 2011). Everyone in soci- Gilberg, T. How to supercharge your eBay feedback rating. ezinearticles.com, 2012.
Available at ezinearticles.com/94054.
ety must pay a high price to compensate for this. The government
Guangzhuo Daily (广州日报). Online shops hired people to conduct fake transactions
must intervene and enforce stricter rules, while reminding people for popularity: Brush Diamond Net platform collapses; hard for sellers to protect
of the value of trust in society. When the internal and external their interests (网店为赚人气雇人虚假交易, 刷钻网倒闭维权难). March 15, 2011.
solutions function together well, then we will be able to resolve Available at http://media.people.com.cn/GB/40728/14148900.html.
Ikeda, D. Words of wisdom. 2012. Available at www.ikedaquotes.org/attitude/
China’s fundamental e-commerce trust issues more effectively. attitude113?quotes_start=7.
Kamvar, S. D., Schlosser, M. T., and Garcia-Molina, H. The Eigentrust algorithm for
Acknowledgments reputation management in P2P networks. In Proceedings of the Twelfth
International World Wide Web Conference, Budapest, Hungary, May 20–24,
2003, ACM Press, New York, NY, 640–651.
This work is supported by the National Natural Science Founda- Liang, Z., and Shi, W. Pet: A personalized trust model with reputation and risk
tion of China under Grant No. 61100183, the Zhejiang Provincial evaluation for P2P resource sharing. In Proceedings of the 38th Annual Hawaii
International Conference of System Sciences, Hilton Waikoloa Village, Island of
Natural Science Foundation under Grant No. Y1110477, Zhejiang Hawaii, IEEE Computer Society Press, Washington, DC, January 2005a.
Provincial Commonweal Technical Project under Grant No. Liang, Z., and Shi, W. Enforcing cooperative resource sharing in untrusted P2P
2013C33063, and the 521 Talents Project of Zhejiang Sci-Tech Uni- computing environments. Mobile Networks and Applications, 10, 6, 2005, 971–
983.
versity. The authors would like to thank Fang Ma from Taobao Soft- Massa, P., and Avesani, P. Trust-aware collaborative filtering in recommender
ware Co., Ltd. for her helpful comments. They also would like to systems. Lecture Notes in Computer Science, Vol. 3290. Springer, New York, NY,
give thanks to Miss Tanya Hosking and Miss Laurie Sorensen for 2004, 492–508.
MoneyWeek. Gray industry cluster parasitism: Four complicated interest chains on
proofreading assistance.
Taobao, (理财周报灰色产业群寄生 : 淘宝越大四大利益链越盘根错节) May 3, 2011.
Available at http://finance.stockstar.com/MS2011050300000899.shtml.
References Nickson, C. eBay feedback scam. SafeFromScams.com, May 11, 2010. Available at
www.safefromscams.co.uk/EBayFeedbackScam.html.
Abdul-Rahman, A., and Hailes, S. Supporting trust in virtual communities. In R. Qian, W. H. (钱卫华). Quality credit white paper: China loses 585.5 billion yuan each
Sprague (ed.), Proceedings of the 33rd Hawaii International Conference on System year caused by trust fraud. (质量信用白皮书:我国每年因失信损失5855亿). Beijing
Sciences, Maui, HI, January 2000, IEEE Computer Society Press, Los Alamitos, CA, Times (京华时报), April 4, 2011. Available at www.chinanews.com/cj/2011/
2000. 04-11/2962399.shtml.
Andersen, R., Borgs, C., Chayes, J., Feige, U., Flaxman, A., Kalai, A., Mirrokni, V., and Ramchurn, S. D., Jennings, N. R., Sierra, C., and Godo, L. A computational trust model
Tennenholtz, M. Trust-based recommendation systems: An axiomatic approach. for multi-agent interactions based on confidence and reputation. In Proceedings
In Proceedings of the 17th International World Wide Web Conference, Beijing, of the Second International Joint Conference on Autonomous Agents and Multiagent
China, 2008, 199–208. Systems, Melbourne, Victoria, Australia, July 14–18, 2003.
Ba, S. Establishing online trust through a community responsibility system. Decision Resnick, P., and Zeckhauser, R. Trust among strangers in Internet transactions:
Support Systems, 31, 3, 2001, 323–336. Empirical analysis of eBay’s reputation system. In: M. Baye, and J. Maxwell
Ba, S., and Pavlou, P. A. Evidence of the effect of trust building technology in (eds.), The Economics of the Internet and E-Commerce (Advances in Applied
electronic markets: price premiums and buyer behavior. MIS Quarterly, 26, 3, Microeconomics), 11, 2002, 127–157.
2002, 243–268. Resnick, P., Zeckhauser, R., Swanson, J., and Lockwood, K. The value of reputation on
Ba, S., Whinston, A. B., and Zhang, H. Building trust in online auction markets eBay: a controlled experiment. Experimental Economics, 9, 2, 2003, 79–101.
through an economic incentive mechanism. Decision Support Systems, 35, 3, Taobao. Research report on taobao buyers’ rating experience (淘宝买家关于‘‘评价”的
2003, 273–286. 体验调研报告) (2011).
Beijing Business (北京商报). Taobao’s integrity self-inspection system goes online The Beijing News (新京报). China’s e-commerce development report: Total e-
while encounters malicious revenge (淘宝上线诚信自查系统打假遭攻击). July 27, commerce sales was nearly RMB6 trillion in 2011, May 30, 2012. Available at
2009. Available at http://tech.sina.com.cn/i/2009-07-27/05063295967.shtml. www.aliresearch.com/?m-cms-q-view-id-72721.html.
Beijing Youth Daily (北京青年报). Revealing the black industrial chain of the Brush Tianshan Net (天山网) 2012. E-Commerce companies are fewer now in 2012. (2012
diamond industry (揭秘刷钻黑色产业链). Tech.163.com, August 5, 2009. 电商淘汰进行时), March 2, 2012. Available at http://www.tianshannet.com.cn/
Available at http://tech.163.com/09/0805/06/5FUCNVQT000915BF.html. fashion/2012-03/02/content/6623113.htm.
Bizer, C., and Oldakowski, R. Using context- and content-based trust policies on the Walters, J., Kuo, Y., Jap, W., and Hsu, H. The World’s Next e-Commerce Superpower:
semantic web. In Proceedings of the Thirteenth International World Wide Web Navigating China’s Unique Online Shopping Ecosystem. Technical Report. The
Conference, New York, NY, May 17–24, 2004, ACM Press, New York, NY, 228– Boston Consulting Group, Boston, MA, November 2011.
229. Wang, Y., and Vassileva, J. Bayesian network-based trust model. In Proceedings of the
BlackHatWorld.com. 3-way feedback exchange. May 26, 2010. Available at www. IEEE/WIC International Conference on Web Intelligence, Halifax, Canada, October
blackhatworld.com/blackhat-seo/ebay/203348-3-way-feedback-exchange. 16, 2003, IEEE Computer Society Press, Los Alamitos, CA.
html. Xiong, L., and Liu, L. A reputation-based trust model for peer-to-peer ecommerce
BlackHatWorld.com. eBay feedback change anyone? September 16, 2012. Available communities. In Proceedings of the 2003 IEEE International Conference on E-
at www.blackhatworld.com/blackhat-seo/ebay/483376-ebay-feedback- Commerce, Newport Beach, CA, June 24–27, IEEE Computer Society Press, Los
exchange-anyone.html. Alamitos, CA, 2003, 275–284.
Brown, J., and Morgan, J. Reputation in Online Markets: Some Negative Feedback. Xiong, L., and Liu, L. Peer Trust: supporting reputation-based trust for peer-to-peer
Working paper, Department of Agricultural and Resource Economics, University electronic communities. IEEE Transactions on Knowledge and Data Engineering,
of California, Berkeley, CA, February 2006. 16, 7, 2004, 843–857.
Chen, Z., and Yang, J. Credit fraud control and credit system optimization on C2C Yu, B., and Singh, M. P. A social mechanism of reputation management in electronic
marketplaces. In R. Sprague (ed.), Proceedings of the 42nd Hawaii International communities. In M. Klutsch and L. Kerschberg (eds.), Cooperative Information
Conference on System Sciences, Hawaii, HI, 2009, IEEE Computer Society Press, Agents IV: The Future of Information Agents in Cyberspace, Lecture Notes in
Washington, DC, 2009. Artificial Intelligence, Vol. 1860, Springer, New York, NY, 2000.
China Knowledge Online. Baidu remains dominant in China’s search engine market, Zacharia, G., and Maes, P. Trust management through reputation mechanisms.
January 31, 2012. Available at www.asiatoday.com/pressrelease/baidu- Applied Artificial Intelligence, 14, 9, 2000, 881–907.
remains-dominant-chinas-search-engine-market. Zhang, Y. Research on trust issue of current Chinese C2C e-commerce: Problems and
Dini, F., and Spagnolo, G. Buying reputation on eBay: do recent changes help? solutions. In Proceedings of 2012 IEEE 11th International Conference on Trust,
International Journal of Electronic Business, 7, 6, 2009, 581–598. Security and Privacy in Computing and Communications, Liverpool, UK, IEEE
Donovan, J. O., and Smyth, B. Trust in recommender systems. In R. St. Amant, J. Riedl, Computer Society Press, Washington, DC, 2012, 1423–1428.
and A. Jameson (eds.), Proceedings of the 2005 International Conference on Zhang, Y., and Zhu, W. A dynamic trust model for e-commerce based on time decay
Intelligent User Interfaces, San Diego, CA, January 10–13, 2005, 167–174. and weighted transaction information. In Proceedings of the Fourth E-Commerce
eBay. Feedback Manipulation Policy, San Jose, CA, 2012. Available at pages.ebay. Retailers and E-Commerce Ecological Academic Seminar, Xi’an, China, 2011, 138–
com/help/policies/feedback-manipulation.html. 154 (in Chinese).
Epstein, G. Alibaba’s Jack Ma fights to win back trust. Forbes, March 23, 2011.