Recommending trusted online auction sellers using social network analysis

Available online at www.sciencedirect.com

Expert Systems with Applications Expert Systems with Applications 34 (2008) 1666–1679 www.elsevier.com/locate/eswa

Recommending trusted online auction sellers using social network analysis Wang Jyun-Cheng b

a,*

, Chiu Chui-Chen

b

a National Tsing-Hau University, 101, Section 2 Kuang-Fu Road, Hsinchu 30013, Taiwan National Chung-Cheng University, 117, Chian-Kuo Road, Section 2, Ming-Hsiung, Chia-yi 62142, Taiwan

Abstract The reputation system currently used by major auction sites to recommend sellers is overly simple and fails to take into account the collusive attempts by some sellers to fraudulently increase their own ratings. This paper presents a recommendation system that uses trading relationships to calculate level of recommendation for trusted online auction sellers. We demonstrate that network structures formed by transactional histories can be used to expose such underlying opportunistic collusive seller behaviors. Taking a structural perspective by focusing on the relationships between traders rather than their attribute values, we use k-core and center weights algorithms, two social network indicators, to create a collaborative-based recommendation system that could suggest risks of collusion associated with an account. We tested this system against real world ‘‘blacklist’’ data published regularly in a leading auction site and found it able to screen out 76% of the blacklisted accounts. This system can provide warning several months ahead of officially released blacklists to help guard against possible seller collusion and can be incorporated into current reputation systems used to recommend trusted online auction sellers.  2007 Elsevier Ltd. All rights reserved. Keywords: Online auction; Social network analysis; Center weight; Information asymmetry; Recommendation system

1. Introduction The reputation mechanism often used by the most popular online auction hosts is effectively a recommendation system for trusted sellers. The system is basically built on a simple mechanism, one that reports the cumulative result of bidder evaluations of the seller over time based on simple valence of positive, neutral or negative, along with communication comments. It makes use of a private ordering system to limit reviews by those who have traded with seller after the transaction has been completed. Successful private ordering systems in the past were based on the verifiability of geographic proximity and frequency of repeated transactions (Ellickson, 1991; McMillan & Woodruff, 2001), or the existence of guaranteed surrogates or resources in the vici-

*

Corresponding author. E-mail address: [email protected] (J.-C. Wang).

0957-4174/$ - see front matter  2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.01.045

nity of transactions (Greif, 1989, 1993, 1994). However, the characteristics of online auction markets that allow registering with pseudonyms and multiple accounts create a mean to misbehave and avoid the consequences of a negative reputation. The current reputation mechanism is reported to have a disproportionately greater amount of positive feedback than negative or neutral feedbacks for fearing of revenge rating (Resnick & Zeckhauser, 2002). As a result, some studies have concluded that online reputation mechanisms have not been proven to be beneficial to the public (Bolton, Katok, & Ockenfels, 2004). Because online auction transaction are paid-up-front then delivered, the buyer is vulnerable to fraud. In fact, there is growing concern that online auction markets are becoming a major source of Internet frauds (Barnes, 2002; Freedman, 2002; Griggs, 2003; Warner, 2003; Wilke & Wingfield, 2003). The most severe cases reported, those involving multimillion dollar losses, often involve bidders who have been misled by the current oversimplified online auction

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

inflated-reputation system (Chua, Wareham, & Robey, 2002; Han, 2003; Wilke & Wingfield, 2003). There are three defects in the current online auction reputation system (Baron, 2001): (a) the information asymmetry among the sellers and buyers, (b) the moral hazard of misusing reputation credits for pseudonym on Internet, and (c) the tendency to not give negative feedback after completion of the transaction. Such defects create a market structure that makes it difficult to verify transactions, creates difficulties in recouping losses due to transaction frauds, and provides a low-risk setting for malicious sellers. The large increase in reported frauds in online auction sites has clouded this market. An improved information structure that can keep online auction reputation system at a low cost has become critical, if such an auction site is to be successful. The approaches used by traditional recommendation systems, regardless whether they are content-based, collaborative, or hybrid of these two, are mostly used for recommending favored products and seldom used for evaluation of traders’ reputation. In fact, those systems are subject to constraints as to their usefulness in evaluating traders’ trustworthiness. The information asymmetry of the merchandise and accounts opened under pseudonyms affect the ability of content-based systems to profile sellers’ accounts. The collaborative recommendation systems are widely used among popular auction sites. In this kind of systems, collective opinions are aggregated as one reputation score reflecting the trustworthiness of each trader’s account. Here again, the given condition of anonymity and ease of maintaining multiple identities or collusive accounts increase the risk of relying on such a reputation score. The various heuristic-based collaborative recommendation systems are also limited by pseudonym and multiple identities on Internet and can only subscribe the information from transaction as inputs. Several studies (Fawcett & Provost, 1997; Goldsborough, 2002; Li, Liu, Wu, & Zhang, 2006; Lucking-Reiley, Bryan, Prasad, & Reeves, 2001; Resnick, Kirwabara, Zeckhauser, & Friedman, 2000; Resnick & Zeckhauser, 2002; Shah, Joshi, Sureka, & Wurman, 2002; Snyder, 2000; Turban, 1997; Wang, Hidve´gi, & Whinston, 2001a; Wheeler & Aitken, 2000) have focused on resolving such problem caused by the use of oversimplified recommendation mechanisms. Their solutions can be categorized into those (a) equilibrating the auction cost structures, or (b) introducing third party intermediators as appraiser to balance the information asymmetry benefits in current recommendation mechanism. However, those implementations have not been adopted by current auction houses for the additional cost and controversy of intermediator roles, and the concerns of losing the competitive niches in the highly elastic nature of online auction market. In this study, we seek to improve current problematic reputation mechanisms so that it may serve as a more meaningful reputation merit indicator for the prospective traders and will integrate smoothly under the simple yet

1667

well-accepted online auction reputation systems. To do this, we introduce the social network analysis approach to analyzing the underlying structure of the accumulated reputation score and its corresponding transactional network. We demonstrate how the social network measurements, k-core and center weights, can effectively filter out the malicious sellers. It cannot only serve as a meaningful and effective indicator for auction-goers to better appraise the risks associated with a reputation score, but also increase the cost of maliciously profiting from the information asymmetry by certain traders. To our knowledge, this study represents the first attempt of incorporating social network analysis information as part of the source of heuristic-collaborative recommendation for online auction reputation. Hereon, we refer to our study as a relationship-based recommendation mechanism for reputation evaluation. We used real world blacklist data that were suspended fraudulent accounts collected from the largest online auction site in Taiwan, the Yahoo Taiwan Inc., as the dependent variables in our evaluation of the proposed recommendation mechanism. The developed mechanism successfully, in monthly average, identified 76% of the fraudulent accounts posted by the online auction host, also referred to as ‘‘blacklisted account(s)’’ hereafter. Of those identified fraudulent accounts, 75% of them are detected with leading time longer than 30 days. This paper is organized as follows. Section 2 reviews the related literature. Section 3 describes the construction of online auction transaction indicators and the possible use of these indicators to recommendation system and the discussion of the results of our findings. The last section concludes with limitations, future research and the final comments. 2. Review of the literature This study focuses on resolving opportunistic abuse of online auction system’s oversimplified reputation system. To establish a context and the target problems for this study, we first reviewed previous research devoted to online auctions and their recommendation systems, summarized findings and contributions to of research devoted to recommendation systems, and review select aspects of social network analysis, which is used to analyze user activity. Note that our discussion of recommendation systems is mostly under the context of reputation. Thus the term recommendation systems can be interchangeable with reputation system in this paper, unless specified otherwise. 2.1. Online auction and the defects of its recommendation system The functions of the auction were identified as coordination, price determination, allocation, and a highly visible distribution mechanism (Klein, 1997). Online auctions reduce the barriers of traditional auction limitations, by

1668

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

relaxing such constraints as the same physical location and the synchronization of time while charging a proportional listing fee and commission fee (Turban, 1997). It also provides a more efficient market mechanism with the power of online searching, multimedia presentation techniques, the low cost of communication, as well as the relatively low listing and commission fees (Klein, 1997). Our study focuses on the auction used in 85% of the online auction markets (Lucking-Reiley, 1999): the open bidding firstprice online auction (English auction). In the traditional auction market, the recommendation system is developed through the actual appraisal of the product, the frequent and repetitive transactions among participants, accumulation of a reputation by the auction house, assessment from trusted subjects, and existence of established brands (Matsubara & Yokoo, 2002). The various types of recommendation systems do little to reduce information overload in recommending trustworthy sellers. Online auction portals try to reduce this overload and establish reputation by providing simple recommendation systems that allow buyers to give feedback about their transaction experiences with a seller over time. For example, after each transaction, buyers using eBay are given the opportunity to evaluate the seller. They choose among three possible valences: 1, 0, and –1, representing positive, neutral, or negative evaluation. They are given the opportunity to post additional remarks about the seller or the product as well. Possible buyers can find cumulative evaluations for the seller he or she is considering. Amazon and Yahoo have the same evaluation design but use a greater rank scale (1–5) or use semantic expressions instead of valences (Yahoo! Inc., 2004). Positive feedback on these systems has been found to be disproportionately higher than negative feedback (positive, 99.1%; neutral 0.3%; negative 0.6%) (Resnick & Zeckhauser, 2002). The current feedback mechanisms in those auction hosts provide little differentiating merit as recommendation references to help users select transaction partners from among the great number of sellers. Because of the simplicity of such recommendation system, two studies also showed that the reputation credits provide few price premiums for the sellers. For every one percent of higher reputation merit, there is only 0.03% of price premium for the same auctioned product (Ba & Pavlou, 2002; Lucking-Reiley et al., 2001). This self-regulated collaborative recommendation system has not been accepted and internalized by the traders as the trust mechanism of online auction society (Bolton et al., 2004). To make matters worse, the system fails to discourage the premeditated use of the simple meriting system by con-artists to deceive the sellers (Freedman, 2002). The online auction environment, using the Internet for communication, along with the nature of anonymity, the low level of medium richness, and the speed of transmission, limits the ability of these simple recommendation approaches to discourage such fraudulent behavior and yields many opportunities for creating noise, making such deception possible. This oversim-

plified online auction recommendation system with its low entry fees to the market creates a safe haven for the online con-artists. The most frequently used auction fraud scheme is bidder collusion in which different roles were played by the same player to extract higher bids. This scheme requires a certain amount of overhead to set up in the traditional auction market. Such schemes are made easier online by the simplicity of the current recommendation system and ease of creating new e-mail account identities. As far as most online auction houses go is to verify the validity of sign-on email account (Matsubara & Yokoo, 2002). The anonymity of the Internet reduces the trouble and cost of bidder collusion. Unlike traditional auction hosts, who are usually held responsible for the delivery and the authentication of the auction object, the Internet auction portals claims only to serve as the platform providers and shuns the role of intermediator in any transaction disputes. The online auction house listing fees are significantly much lower than those of other trading markets. The much lower level costs of entering and exiting the system may encourage, not deter, fraud and collusion. Such online recommendation mechanisms and low entry and exit costs make shilling an attractive scam (Wang, Hidve´gi, & Whinston, 2001b). In order to resolve the problems created by the oversimplified online auction recommendation mechanism, some have tried to reduce the profit gained from information asymmetry by directly or indirectly changing current price/cost structures and by introducing a third party to validate information. In the first (Wang et al., 2001a), mathematic/economic models were used to provide new pricing schedules or transaction structures to balance the lack of online market equilibrium caused by information asymmetry and the skewing of transaction costs. Other studies (Ba, Whinston, & Zhang, 2003; Changchien & Lu, 2001; Chua & Wareham, 2002; Resnick et al., 2000; Snyder, 2000) have suggested using intermediary to solve trust issues associated with an asymmetrical information market. Neither approach was accepted by online auction houses, who feared undoing the online auction market niche and increasing the low accessing cost provided by Internet information technologies. Online auction survives and prospers because information technology provides its own niche: a high degree of anonymity, relaxed legal constraints, lower barriers to entry and exit, and the limited media richness. The beauty of the simplicity of the system is accompanied by information asymmetry. Consequently, current online auction structure of business practices and simple reputation system provides the opportunity for the con-artist to manipulate the system for purposes of deception (Chua & Wareham, 2002). 2.2. Recommendation systems The study aims to provide a sound recommendation system for trustworthy sellers in online auction market, one

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

that would resolve the ill-gained profits made possible with information asymmetry of the product and collusion accounts. We review the literature to illustrate the specific problems related to the categorization of recommendation systems and boundaries between them. Use of online ratings is omnipresent and can be traced back to 1990 (Adomavicius & Tuzhilin, 2005). Ratings examples can be found for books (Linden, Smith, & York, 2003), movies (Miller, Albert, Lam, Konstan, & Riedl, 2003), news product items (Billsus, Brunk, Evans, Gladish, & Pazzani, 2002), and commerce servers (Peddy & Armentrout, 2003) as well as Google’s PageRank (Page, Brin, Motwani, & Winograd, 1998). These rating systems use various approaches, including information retrieval, information filtering, and rating synthesis. In one comprehensive overview of related recommendation system studies (Adomavicius & Tuzhilin, 2005), three differences were found, including how recommendation is approached, the techniques used and the problem targeted. Recommendation systems, differentiated by the sources of recommendation data, are categorized into three types (Balabanovic & Shoham, 1997): (a) content-based recommendation systems, which are those based on similarity of the items to the user’s previous preference profile, (b) collaborative-based recommendation systems, which are based on the general tastes of similar users’ profile, and (c) hybrid recommendation systems, which make use of a combination of the former two approaches. The techniques used in both two approaches can be heuristic-based or model-based. Heuristic-based techniques usually study contextual information from transactional behaviors to profile the rating. Model-based techniques are based on probability, statistic, or machine learning models from underlying data. They include Bayesian classifiers, linear regressions, clustering, decision trees, and artificial neural networks, all used to create various levels of recommendation ratings. The problem target is modeled as the utility function (the rating) of the item/object in the specific profile, generally that of the user or the context (Adomavicius & Tuzhilin, 2005). Two previous approaches or techniques may be used to retrieve information or filter the rating item/object and context/profile of the user construction to estimate utility of the rating. The content-based approaches are limited by the availability of specifically defined objects with features that count toward the rating, and require enough data to build new user’s profiles to provide accurate recommendations. The resource limitation constrains the practical application of attribute collection in the generation of significant and relevant ratings (Shardanand & Maes, 1995). The speed of transaction and the variety of auction objects make the information retrieval and filtering techniques inapplicable. Information asymmetry makes reliability of retrieved data questionable. The reliability of the profile information is weakened by the use of fictitious account names and lack of accountability to goes with that anonymity. Therefore, information retrieved from content-based cannot be used

1669

reliably to construct a recommendation system for trusted traders in the online auction market. For the heuristic-based approaches, like those used by Amazon and MovieLens, rating information is based on previously rated instances. Heuristic-based approaches attempt to locate the users with similar preferences/tastes. This use of stereotyping usually works well in finite categories of specific objects, such as recommendation of books, movies, news, or restaurants (Billsus et al., 2002; Burke, 2000; Linden et al., 2003; Miller et al., 2003). In the online auction market, the wide range of the auction object and information asymmetry affects the possibility and reliability of stereotype building. The same problems exist in building user stereotype preference models. Like the others, this oversimplified rating mechanism disregards or overlooks the impact of information asymmetry and lack of accountability associated with the use of fictitious account names and leaves lots of room for intentional fraud. To meet the challenges associated with the use of such online auction market practices, we chose to use transactions to analyze the contextual information. While some researchers using heuristic-based approaches have proposed conceptual models that utilize context information taken from electronic transactions, they did not specifically address the application of their models to online auction market (Adomavicius & Tuzhilin, 2005; Aggarwal, Wolf, Wu, & Yu, 1999; Hogg et al.). In the approach taken by Aggarwal et al. (1999) the graph-theory was used to determine user social paths to create a profile of the relationship between a pair of users. The profile, which could be used as a recommendation index, was ‘‘predictability.’’ The study presented as a conceptual model with simulation data for concept illustration and testing. Adomavicius and Tuzhilin (2005) proposed an extension model in which user interaction and the products being purchased were weighted to create recommendation ratings. In another study, Hogg and Adamic (2004) briefly describe the concept of using electronic social position to rate and filter. Another study, Zacharia, Moukas, and Maes (1999), advocating the use of social context and investigating problems related to information asymmetry and the use of fictitious account names, proposed the weighting of the evaluator’s reputation points in rating mechanisms, a process similar to PageRank in Google’s webpage rating system, which utilizes collaborative ratings based on social interactions. However, in the proposed system, the accounts with higher ratings would remain high even if recent transactions were found to be fraudulent. Although their model has not been tested with real transaction data, they suggest utilizing social context information taken from previous transactions to make recommendations for certain purposes. As can been seen, the studies above have tried to use interaction based relations to make up for a lack of demographic data about the user or certain object/item categories in earlier collaborative rating systems. Because the servers of the transaction portal keep a lot of historical transaction data, interactions can be analyzed and social

1670

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

relationship can be discovered, very likely making it the most applicable approach for reconstructing relationship networks in online auctions. However, the studies we have reviewed so far, have attempted to re-develop the interaction rating techniques from a conceptual model that involves the exploring relationships without exploiting related sociology theories. In our study, we attempt to incorporate the well-developed methodologies from social network analysis to examine the relationship networks by the transactional data. 2.3. Social network analysis (SNA) SNA attempts to quantify the interaction among the members of a community in order to profile its structure and its members. This approach has been developed and practiced in the domain of social studies for decades for social scientists to investigate social group dynamics. The basic components of an SNA study are the actor or node, as the source of action, and the connection or link, as the relationship developed among nodes. The nodes can be individuals, organizations, or communities (Borgatti, 1998) and the links can be a single or multiple types of relations or shared characteristics among the actors for an understanding the effects of the social structure on individual members as well as on the community as a whole (Freeman, 1979; Wellman, 1996). A great deal of network relations information has been learned and many constructs have been developed through decades of accumulated research in SNA research (Fischer, 1977; Johnson & Milardo, 1984; Mitchell, 1969; Scott, 2002; Whitten & Wolfe, 1973). Our study uses SNA as a theoretical framework and takes into two major SNA measurements – centers and k-core. Pertinent information is described as follows: 1. Relations are the key focus of SNA. There are two basic types of relations, group and network (Garton, Haythornthwaite, & Wellman, 1997; Wellman, 1996). Group relations have the characteristics of a close relationship, a narrow range, and low mobility, as is found in teams established to work on certain projects. Network relations have a lower density of member interaction, a broader range, many varieties of relations, and high mobility. Relations are the channel for exchange of resources, and it is the transaction in this context. The function of the network can be categorized by the type of the resources exchanged, such as trust, suggestions, etc. Relations are measured through such measurements as distance, density, frequency, centrality etc. Centrality is the measurement adapted from mathematic graph theory to analyze the relation status and dynamic of the network and group. Centralization is the structure indicator for the network, group, and the individual/node position status related to other nodes. Centrality can be measured in terms of degree, closeness, density, closeness, and betweenness (Hanneman, 1998; Scott, 2002).

Degree of centrality is the measure of point centrality by calculating the number of nodes one node is adjacent to (Nieminen, 1974). Closeness is the measurement of global centrality in term of the ‘‘closeness’’ of all the nodes in the group or network by measuring the path distance (Freeman, 1979; Freeman, 1980). Betweenness specifies the node’s status in the group and onto other nodes (Freeman, 1979). In this study, we use one measurement of centrality, which is called ‘‘Centers’’, to find the centers in the partitioned subgroup network extracted from original transaction networks. The concept of ‘‘Centers’’ follows a ‘robbery’ algorithm. Nodes that have higher degrees (are stronger) than their neighbors steal the weights of linking from adjacent nodes. In the beginning of the process, each node’s initial strength (weight) is calculated based on its degree of linking. Later the strength (weight) of each node is compared with that of its neighbors. The ‘‘stronger’’ neighbors steal weight from a ‘weaker’ node. The stronger they are, the more weight they steal. Two nodes steal the same weight from each other when they are equal in strength (Batagelj & Mrvar, 2005). This algorithm provides a better result in differentiating the local center nodes (whose resultant weights are greater than 0) from periphery nodes (resultant weights is zero) than degree centrality calculation. We called such measurement as ‘‘centered-weights’’ or ‘‘CW value’’ in rest of this paper. 2. Subgroup observation forms the focus of this research. Subgroups can be identified within the network and can provide valuable information in understanding the influence of the subgroup on the network as a whole and on individual characteristics and status within the subgroup, among subgroups, and the entire network. A subgroup can be identified by the measurements of component, clique, k-plex, and k-cores. K-core is a maximal sub-graph in which each node is adjacent to at least k other points. It is also thought to be an essential complement to the measurement of density, which may not capture many of the features of the global network (Scott, 2002). The mathematic definition of k-core is (Batagelj & Zaversnik, 2002): Let G = (V;L) be a graph. V is the set of vertices and L is the set of lines (edges or arcs). We will denote n = jVj and m = jLj. A sub-graph Hk = (W; LjW) induced by the set W is a k-core or a core of order k iff "v 2 W : degH(v) = k, and Hk is the maximum sub-graph with this property To illustrate, in Fig. 1, the subgroup, which includes (A, I, G, B, M), is the 2-core group. It provides the comparison to the other subgroup of 1-cores, e.g. (B, K, L, J), or (M, C, O, P, Q, N). The underlying structure revealed by the k-core subgroup measurement is used in this research as the indicator of abnormality in the context of Internet auction transactions. In SNA developed measurements, component,

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

Fig. 1. SNA k-core subgroup example, (the center nodes: A, M, I, G, B formed a 2-core).

clique, k-plex, and k-core are the indicators used to portrait the cohesiveness of subgroups. Each measure has a different level of cohesive intensity. The component is used to identify the differences among subgroups while showing little information within the subgroup. The clique forms the strongest subgroup cohesion by imposing all the nodes within the subgroup to be connected. This requirement, however, is not suitable for use in comparing the structure differences among clique subgroups. The k-plex has fewer connection requirements than the clique does, but there is not as much literature describing this measurement and it applications. K-core, however, is frequently used to differentiate the strong ties among subgroups (Swan, 2001). The level of k-core clustering measures provides more structure information than the measurement of clique does, and the ratio of decreasing number of nodes between different k-core levels reveals the structure details within the group (Reffay & Chanier, 2003). The greater the number of nodes of high k-core value, the stronger the cohesiveness of these nodes. The core/periphery (C/P) ratio is an indicator derived from the k-core. It is the ratio of the k-core members divided by the periphery members/nodes adjacent to the core members. As suggested in Liao (2005), the utilization of social science methodologies can be implemented in expert systems for obtaining new methodologies and understanding. The concept of SNA has been adapted and proved to be able to provide valuable insight into the study of the subjects in different group, organization, and community contexts. Examples of its utility can be found in organizational knowledge sharing (Tsai, 2002), the information it conveys regarding organizational structure and working relationship (Cross, Rice, & Parker, 2001), the study of agent system in Internet communities toward trust-building (Sabater & Sierra, 2002), and the organizational subgroup structure identification through email log (Tyler, Wilkinson, & Huberman, 2003). Several scholars have proposed using the SNA approach for analyzing criminal network (Potter, 1994; Sparrow, 1991; Xu & Chen, 2002). In fact, the structure of one criminal organization was analyzed through SNA to illustrate the network core, the network peripheries, the network activities pattern, and the roles related to

1671

the connections of analyzed actions (Granovetter, 1973; Williams, 2001). The practical application of SNA methods can be seen in the use of document data to identify criminal territory allocations in the US (Potter, 1994), locate terrorists and their collaborators network structure in the 911 attack identification (Carley, 2002; Krebs, 2002), and uncover insurance collusion or conspiracy through nodeto-node information recursive analysis (Little, Johnston, Lovell, Rejesus, & Steed, 2002). Our study represents, as far as we know, the first attempt to use the social network perspective in recommendation rating systems. This approach makes possible a technically and economically feasible recommendation system to overcome the major problem of online auction frauds. Our approach to using information technology to improve the current recommendation system use by online auction houses was based on one that was proposed and tested previously (Wang & Chiu, 2005). Basing our research on that previous reported approach, we further developed an actual recommendation system that can be incorporated into online auction systems. Next section will illustrate the detail analysis process of our research to construct an applicable system for online auction markets. 3. Building collaborative recommendation system with relationship-based heuristic technique In a previous study (Wang & Chiu, 2005), k-core was found to be effective for identifying the inflated-reputation accounts which tend to receive negative feedback comments. Derived from transaction histories, the k-core indicators reveal the cohesive of the account’s trading relationship as well as the reputation structures. In other words, frequently engaged accounts will form subgroups that can be captured by the k-core indicators. The target is to separate the malicious traders with the normal accounts. In this section, we illustrated the process of our study, from the data collections, relationship construction, SNA measurements building and testing based on recall and precision ratio of fraudulent accounts’ identification. The process represents our approach to making recommendations for the selection of the online auction partners. 3.1. Data source and data collection process Similar to the research (Min & Han, 2005), the user’s time-variant pattern data are collected and analyzed as collaborative filtering sources. Our data are collected from the leading online auction web host in Taiwan, Yahoo! Taiwan, which has 75% of Taiwan’s auction market shares. In 2004 Taiwan’s annually auction trading amount is around U$ 6 billion and Yahoo’s auction account for 4.6 billion (TWINC, 2005). Part of the reason of using the Yahoo! Site is that a ‘‘blacklist’’ of confirmed fraudulent accounts is released on a regularly basis. Starting at February 1st, 2005, Yahoo Taiwan started to post the suspended accounts for having engaged in illegal business practices or

1672

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

violated the web site’s code of transaction. This blacklist would become the source for us to test whether an account is fraudulent. We collected the account’s reputation history started from the beginning of the first blacklist page till May 26th 2005, a total of 2042 accounts posted. Some of the accounts were posted accompanied by the alias used by the account. Those accounts are grouped and considered as the same user with multiple identities. After the blacklist accounts collected, we proceeded with studying the fraudulent accounts’ transactional behavior patterns by rebuilding its transactional interaction patterns on a monthly basis. Transaction logs are kept for three months backward, whereas each account’s evaluation records are kept since an account’s was created. Thus we reconstruct the blacklisted accounts transaction network based on the evaluation logs using a self-developed web clawer. There are total 15,736 evaluation web pages gathered from Yahoo Taiwan. For all the blacklisted accounts, only 1904 accounts (out of 2042) have evaluation records. The self-developed parsing program parsed through these pages to extract the evaluation-given accounts, the auction items, the evaluation valence. . . etc. Three database tables are built from the identified data. They are: the profile of blacklisted accounts, the dated evaluation transaction logs, and the additional auction information with evaluation comments. There are total of 205,826 auction evaluation records acquired in the period of July 24th, 2001 to July 20th, 2005. Of these records, there are 143,944 unique paired accounts evaluation logs. The repeated transactions between the two accounts could only give feedback to each other once, so that we eliminated the later transaction and kept only one relation link. Beside the blacklisted accounts, from these evaluation logs, there are additional 112,491 accounts identified in the transactional network. 3.2. Constructing the interaction structure from transactions To simplify the process and calculations, we targeted those accounts that had transactions with more than two

blacklisted accounts for a 2-core subgroup, which requires at least two linkages for any given accounts. There are 92,223 accounts dropped from the retrieved logs for having only one transaction. We analyzed the social network measurements from the rest of the 20,258 accounts to track the network linkages among them, the data covering the period starting as early as October 2001 to June 2005 for 44 months. The linkages started from 6 transactions and 11 accounts on the first month, October 2001 and accumulated to 53,788 linkages among all 20,528 accounts on June 2005. Next, for every monthly transaction network, 2-core subgroup partitions of these networks were calculated and generated to identify the cohesive subgroups in these networks. Because the scale of these networks, we use the program Pajek 1.10 (Batagelj & Mrvar, 2005) for the SNA measurements processing. The subgroup process of k-core provides substantial power to extract the cohesive accounts from mass amount of transaction recorded. Fig. 2 exhibits the original transactional network on November 2002 with 1272 accounts at the left panel, and at right panel reduced to 179 accounts belonged to a cohesive 2-core subgroup network. Four colors/shapes of accounts are identified in the figure: (1) the red/square nodes are the blacklisted accounts that were successfully identified by our mechanism, (2) the yellow/diamond nodes are the blacklisted accounts that our mechanism failed-toidentified, the type II errors, (3) the green/triangle nodes are the ones not in the blacklisted accounts but falsely identified by our mechanism, the type I errors, and (4) the blue/ circle nodes are not in the blacklisted accounts and our mechanism also concurs. The first 2-core cohesive subgroup in the data appeared at June 2002 with 4 accounts linked by 4 transactions that are extracted from 291 accounts with 304 transactions. And, for these 4 accounts, two of the accounts were not posted as the blacklist account till 2005/3/10 and 2004/ 11/30. One of these blacklisted accounts had posted as the alias account and the main account of this alias account was identified in the 2-core subgroup on February 2004. From these result, the 2-core subgroup provides a great fil-

Fig. 2. November 2002 total transaction network (1272 accounts) vs. 2-core network (179 accounts). The red/square nodes: accounts posted in the blacklisted and successfully identified, the yellow/diamond nodes: accounts posted in the blacklisted and failed-to-identified, the type II errors, the green/ triangle nodes: accounts not posted in blacklisted and false-identified, the type I errors, and the blue/circle nodes: account not posted in blacklisted and successfully identified as normal accounts.

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

tering mechanism as illustrated in Wang and Chiu (2005) from the early days of online auction transactions with the precision ratio of 50%. However, as the relationship aggregated from cumulative transactions, the accounts formed 2-core subgroups have grown rapidly and precision ratio deceased. Using the data in Fig. 2 as example, on November 2002, for the total 179 accounts in 2 cores (the one with yellow/diamond and red/square shaped nodes), there are only 54 accounts (the red/square shape nodes in Fig. 2) were posted as fraudulent accounts in the blacklists. Those accounts were not posted as blacklist until from 2004/5/10 to as late as 2005/3/10. The 2-core subgroup relationship provides the strength as the early warning indicator. However, the type II error, the false-identified ratio, is as high as almost 70% on this month and the precision ratio also dropped dramatically as the transactions aggregated at later months. The results of 2-core subgroups’ account number, the blacklisted account number, the blacklisted accounts that joined before the month, the recall ratio, and the precision ratio of each month is exhibited in Fig. 3. There are two Y-axes in the figure, the first 3 numbers use the Y-axis at the left, and the two ratios use the right Y-axis. The X-axis represents each month. The calculation formula and its process will be illustrated later in this section. As Fig. 3 exhibited, the precision ratio, successfully identifying the blacklisted accounts, has drop substantially from 50% on June 2002 to 7.3% on June 2005. To be an effective recommendation mechanism, the further processes are imperative. There are two approaches to accommodate the dropping of 2-core’s precision ratio addressed in Wang and Chiu (2005). One is to further partition the core network to higher level, e.g. 3-core or higher. The other one is to calculate the core/peripherals ratio of the subgroup’s network.

1673

We have pursued to the highest core partitions as high as 10-core on the extracted data. The similar figures identified in Fig. 2 are provided for 3-core subgroup relationship network exhibited in Fig. 4. As Fig. 4 exhibited, by lowering the recall ratio, the precision ratios are higher than Fig. 3 and the average precision ratio also increased to 24% from 16.5% at 2-core subgroup. However, by comparing the difference of blacklisted account numbers between Figs. 3 and 4, the increase in precision ratio also raises the failed-to-identified type II error to be greater than 50%. The second approach is to calculate the core/peripherals (C/P) ratio of each core subgroup. As shown in Fig. 2, the 2-core subgroups on November 2002 have only two 2-core different subgroups in the graph: a four-account small 2-core at the upper right area, and the rest of 175 accounts formed a big 2-core subgroup. The C/P ratio index for one subgroup is the same for all 175 accounts by definition. This process cannot provide further differentiation power to increase the precision ratio. So, we looked for other social network measurements for remedy. The result of 2-core subgrouping has high recall ratio as well as the high type I error, false-identified number of blacklisted accounts. A desirable outcome would be to increase the precision ratio but not at the expense of increasing failed-to-identified ratio. The centrality measurements provide the information of network structure information for each node. After testing various types of centrality measurement, e.g. linkage degrees, closeness, and betweenness, we find that the ‘‘Centers’’ algorithm provides the most applicable discriminating power for improving the precision ratio. The center measurement starts with one account’s linking degrees as initial weights and uses ‘‘robbery’’ algorithm to rob the linkage/degree

Fig. 3. The recall and precision ratios of the 2-core subgroup network accounts.

1674

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

Fig. 4. The recall and precision ratios of the 3-core subgroup network account.

Table 1 Results of the three methods at one month before the accounts blacklisted

2-core 2-core with centers

Type-I err (%)

Type-II err (%)

Recall%

Precision%

92.7 6.6

32.9 70.8

67.1 29.2

7.3 93.4

weights from neighbors that has less weights. All the weights originally associated with the ‘‘weak’’ nodes in the graph/network will be reduced to zero. The algorithm strengthens the stronger node’s weights among the neighbor nodes, which in turn identify the relative local ‘‘center’’

nodes within the network. This center algorithm generates two groups of accounts from 2-core subgroup, one group of accounts were ‘‘robbed’’ to have centered degree weight drop down to zero. The other group has the centered degree weights greater than zero and is identified as potentially suspicious accounts. As shown in Table 1 and Fig. 5, the application of center weights resulted in dramatic improvement to the Type I error rate. In particular, the Type I error rate for June 2005 was reduced to an impressively low value of 6.2%, from the previous best value of 82.2% in Method-II. Only 42 honest actors were incorrectly identified by the method. At the same time, the Type II error rate did increase to a 70.8% level, due to its omission of 1445 fraudulent actors

Fig. 5. The no. of actors identified and error rates.

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

out of a total of 2042. However, this still represents its capability of detecting 29.2% of fraudulent actors in tuning the system to a relatively loose setting, which is a significant number (597 fraudulent actors were correctly identified). We can see that one distinctive feature of this method is its extremely high reliability in correctly identifying fraudulent actors.

4. Results and discussion From the data process and analysis results above, by combined with online auction transaction network 2-core subgroup partition and centrality index of centered-weights algorithm, the average precision ratio to filter out the fraudulent accounts stands at 76.0% with standard deviation of 11.5%. This index can serve as a warning alarm to help the auction-goers avert the probable problem transaction partners. Some interesting results are worth noting. First, we examine closely on the figures of two types of errors. The first one is the failed-to-identified, or the type II error. They are mostly caused by the account who worked alone without collusion, i.e. those in blacklist but no alias accounts. This is the limitation of this research design, the system is to identify the traders who use alias to inflate their reputation or transaction records that maliciously manipulating the information asymmetry accompanied by exploiting the oversimplified reputation system of online auction market. For those lone blacklisted accounts, they may just join the market and acted recklessly to cheat and re-enter by another identity. To identify such types of behavior is beyond the scope of this research. We further inspected 228 of zero center weight (CW) blacklist accounts who have no alias and found that about one-fourth, 44 accounts, were posted as fraudulent accounts within 30 days after they joined the market, and one third of them, 76 accounts, recognized as fraudulent accounts within 2 months. Of the same 228 accounts, about one third, 85 accounts, of their reputation credits are below or equal to zero and 205 of the accounts’ reputation credits are below 30. Adding the power for filtering out calculated fraudulent accounts proposed in this study to the current reputation mechanism, the online auction goers would have access to better information in selecting transaction partners. To be a viable recommendation system, the indicators should alarm to the general traders before the actual fraudulent transaction took place. Further examination of the results of identified blacklisted accounts by 2-core and CW value gives encouraging evidences that our proposed mechanism can provide leading warning signals. For 834 blacklisted accounts identified by 2-core and CW value, there are 200 days in average ahead of the time when those accounts are posted as fraudulent accounts. There are almost 75% of accounts were identified 30 days prior to posted time. Moreover, there are 49% accounts are identi-

1675

fied 6 months in advance and only 4% of accounts failed to give advance warning. The power of providing leading warning signals can adversely lead to type I error, the false-identified accounts by 2-core and CW value. Earlier warning signs more likely to be erroneous and contribute more to the false-identified type I error rate. For example, for three years period that the data spanned, accounts flagged in the first 13-month have average error rate of 18%, and the last 12 month average error rate is only 7.6%. A further study on verifying the power of the leading indicators and finding possible enhanced mechanisms may be necessary. To provide a comprehensible index value to complement the current auction feedback evaluation reputation systems, we propose an index of ‘‘Closeness of Reputation Relationship’’ (CRR) as the fraudulent account indicator derived from this study. The model simply takes into account the two indicators we find useful in this study. The calculation of CRR is proposed as CRR ¼ 1  ððH k Þ

1=2

1=2 1

þ ðCWHk Þ

Þ

for H ek >¼ 2

¼ 0 for H k ¼ 1 Hk is the maximum k-core that the account participated, for k-core definition in Section 2. CWHk is the centeredweights value calculated from the account’s Hk-core subgroup. The selection of using square roots of Hk and CWhk is to smooth the increasing slope of CRR that better reflect the average precision ratio obtained in this study. From the constructed model, the highest possible value for CRR is 1 when the account has high k-core, cohesive subgroup, and high centered-weights which is the central hub of the subgroup. The lowest value is 0, when the account’s reputation credits are only sourced from star network. The result of CRR may be presented as percentage to stand for the concentration ratio of the account’s reputation sources when applied to current reputation system. The identified blacklisted account’s CRR in this study are located at the range from 64.67% to 98.8%. The CRR index can serve as an auxiliary indicator to help the auction-goers having better understanding and interpreting the seller’s/ buyer’s reputation credits and even as an early warning notice. By the presented model, we further processed all CRR indices for those identified blacklisted accounts. To illustrate the progressing of blacklisted accounts’ CRR, we calculated four months’ CRR for every account. We started at the month that the account was blacklisted and three months before the month. The distributions of all CRR indices to the account’s reputation credits at one month before being blacklisted were exhibited in Figs. 6 and 7. Most accounts’ CRR indices were higher than 0.6. Three types of marks were employed to denote the group status of the blacklisted accounts. Those accounts with square mark were blacklisted as a lone individual account without accompanying alias account ID. Those accounts with tri-

1676

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

Fig. 6. The CRR distribution to auction credits – 1 month before the account blacklisted.

angle and asterisk marks represent fraudulent accounts involving a primary account and some alias accounts. We used a triangle mark to indicate the primary account, which is identified by the highest CRR value within the posted alias group accounts. The asterisk marks represent the peripheral accounts in the alias group. In Fig. 7, the square-marked accounts (the lone individual accounts) were removed, leaving only the primary and peripheral fraudulent accounts only. We can observe from the figure that CRR index effectively separated the primary accounts from the peripheral accounts in blacklisted alias group. The primary account groups were mostly located at high CRR value and high reputation credit area. Those accounts artificially manipulated their reputations with the help from the peripherals, or the alias accounts. Our approach demonstrates that primary and alias accounts

can be differentiated through the analysis of social interactions. To verify the significant differences between primary and peripheral group accounts, ANOVA statistic test results were exhibited as Table 2. The results showed that two groups of blacklisted accounts were significantly differed by CRR in all 4 months at the level of 0.01. In Table 3, the 95% confidence intervals for the primary and peripheral blacklisted account groups were exhibited. The CRR for the primary accounts, in 95% confidence interval, have lower bound of 0.42 at three months before the blacklist being posted by Yahoo. This figure continued to increase and almost reached 0.5 at two months before posting time. The index of CRR provides useful discrimination power to help buyers understand seller’s reputation structure before the fraudulent acts occurred.

Fig. 7. The CRR distribution to auction credits – 1 month before account blacklisted, grouped only.

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679 Table 2 ANOVA test result for CRR values between primary and peripheral blacklisted accounts

Groups at 3 months before account blacklisted Groups at 2 months before account blacklisted Groups at 1 months before account blacklisted Groups at the blacklisted month

Sig.

through the perspective of social network analysis. It is interesting to note that while savvy users may find ways to exploit the online environment, the salvation can be provided from within the system itself.

60.755

0.000

5.1. Limitations and future researches

7.868

80.557

0.000

9.081

9.081

104.703

0.000

8.837

8.837

216.799

0.000

Between groups sum of squares

Mean square

6.289

6.289

7.868

F

Table 3 The CRR confidence interval Time

1677

Account group

95% Confidence interval for mean Lower bound

Upper bound

3 Month before blacklisted

Peripheral Primary

0.222523 0.421953

0.282310 0.526462

2 Month before blacklisted

Peripheral Primary

0.253240 0.480605

0.311949 0.580737

1 Month before blacklisted

Peripheral Primary

0.292546 0.540533

0.348483 0.633511

Blacklisted month

Peripheral Primary

0.421945 0.671613

0.459959 0.736107

5. Conclusion Using SNA indicators, k-core and centered-weights algorithm, we demonstrate a feasible technique to build the heuristic-collaborative recommendation system for trusted sellers from relationship-based transaction records. The examination of network structures from transaction or interaction logs has proved to be viable information for revealing the underlying social relationship structure of the traders. The problem originated from Internet’s nature, such as information asymmetry and ease of collusion under pseudonym, can be remedied by uncovering the underlying structure of the transactional network. The application of this mechanism can complement the current online auction feedback rating system to uncover the ill-intended traders exploiting the nature of online auction. The implications of our findings are twofold. First, by analyzing the transaction logging and utilizing computing power to construct social relationship structures, we can reconstruct the relationship profiles to supplement the lack of demographic data in the online environment. Secondly, by using the social network analysis, it is useful to find proper measurements to identify ill-intended users who left interaction footprints when forging their creditability with additional information process resources and activities. Under such a complicated new online business framework, the order and tenets can be discovered and understood

The first limitation is the scope of data collection. Because of resources and time constraints, the transaction data were constructed by taking the blacklisted accounts as centers and grow outward. The generalization of this research may require further large scale investigation for verification. Secondly, although this research attempted to make contact with online auction hosts, Yahoo Taiwan Inc., it did not result in positive feedback. So the data was collected from our self-developed web crawler to gather the information posted after the conclusion of the transaction. The cumulative transaction comments and credits may have been altered during the parsing and extracting process. If this study could have involved the cooperation of online auction hosts, the prediction capability of those indicators could be further verified. Thirdly, as discussed in previous section, such recommendation mechanisms have certain failed-to-identified ratio. It is still possible for the hostile fraudulent accounts to use multiple accounts inflating only once to its target alias account (a star network) without forming a circle linkage of k-core and CW value. However, such plot will require a much higher transaction costs and risks in conducting their activities. The nature of recommendation is not to be fool-proof but to provide helpful information given the information overloaded scenarios. The use of our recommendation mechanism will effectively increase the difficulty for reputation manipulation. Social network structures were parsed and developed from the transaction data obtained from the public web pages. The online auction hosts should have more detailed information to help verification, such as open account information, the access IP, ISP, etc. With such information inside their database and with their information processing power, there are opportunities to identify further social insights thorough SNA measurements and construct more streamlined mechanism for other problems, such as shilling, bid shielding, or shell auctions, etc. These would constitute the further research subjects extended from this finding. References Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), 734–749. Aggarwal, C. C., Wolf, J. L., Wu, K.-L., & Yu, P. S. (1999). Hosting hatches an egg: a new graph-theoretic approach to collaborative filtering. Paper presented at the Fifth ACM SIGKDD international conference on Knowledge discovery and data mining, San Diego, CA, United States.

1678

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679

Ba, S., Whinston, A. B., & Zhang, H. (2003). Building trust in online auction markets through an economic incentive mechanism. Decision Support Systems, 35(3), 273–286. Balabanovic, M., & Shoham, Y. (1997). Fab: content-based, collaborative recommendation. Communications of ACM, 40(3), 66–72. Ba, S., & Pavlou, P. A. (2002). Evidence of the effect of trust building technology in electronic markets: price premiums and buyer behavior. MIS Quarterly, 26(3), 243–268. Barnes, B. (2002). Art & money. Wall Street Journal, W.8. Baron, D. P. (2001). Private ordering on the Internet: The eBay Community of Traders (No. 1709). Graduate School of Business, Stanford University. Batagelj, V., & Mrvar, A. (2005). Pajek, Program for Large Network Analysis (Version 1.10). Ljubljana, Slovenia. Batagelj, V., & Zaversnik, M. (2002). An O(m) algorithm for cores decomposition of networks. Unpublished manuscript. Billsus, D., Brunk, C. A., Evans, C., Gladish, B., & Pazzani, M. (2002). Adaptive interfaces for ubiquitous Web access. Communications of ACM, 45(5), 34–38. Bolton, G. E., Katok, E., & Ockenfels, A. (2004). How effective are electronic reputation mechanisms? An experimental investigation. Management Science, 50(11), 1587–1602. Borgatti, S. P. (1998). What Is Social Network Analysis? (Retrieved 10.07.2003). Burke, R. (2000). Knowledge-based recommender systems. In A. Kent (Ed.). Encyclopedia of library and information systems (vol. 69, pp. 32). Marcel Dekker. Carley, K. M. (2002). Dynamic network analysis. In: R. Breiger & K. M. Carley (Eds.), Summary of the NRC workshop on social network modeling and analysis. National Research Council. Changchien, S. W., & Lu, T.-C. (2001). Mining association rules procedure to support on-line recommendation by customers and products fragmentation. Expert Systems with Applications, 20(4), 325–335. Chua, C. E. H., & Wareham, J. (2002, June). Self-regulation for online auctions: an analysis. Paper presented at the international conference on information systems, Barcelona, Spain. Chua, C. E. H., Wareham, J., & Robey, D. (2002, December). Anti-fraud mechanisms in internet auctions: the roles of markets, hierarchies and communities of practice. Paper presented at the international conference on information systems (ICIS). Cross, R., Rice, R. E., & Parker, A. (2001). Information seeking in social context: structural influences and receipt of information benefits. IEEE Transactions on Systems, Man, And Cybernetics – Part C: Applications and Reviews, 31(4), 438–448. Ellickson, R. C. (1991). Order without law. Cambridge: Harvard University Press. Fawcett, T., & Provost, F. (1997). Adaptive fraud detection. In Data mining and knowledge discovery (pp. 1–28). Boston, MA: Kluwer Academic Publishers. Fischer, C. S. (1977). Networks and places: Social relations in the urban setting. New York: Free Press. Freedman, D. H. (2002). What eBay isn’t telling you. Business 2.0, 3(8), 56. Freeman, L. C. (1979). Centrality in social network: I. Conceptual clarification. Social Networks, 1. Freeman, L. C. (1980). The gatekeeper, pair dependency and structural centrality. Quality and Quantity, 14. Garton, L., Haythornthwaite, C., & Wellman, B. (1997). Studying online social networks. (Retrieved 8.07.2003). Goldsborough, R. (2002). Personal computing: avoiding online auction fraud. Unpublished manuscript, Reston. Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78, 1360–1380. Greif, A. (1989). Reputation and coalitions in medieval trade: evidence on the Maghribi traders. Journal of Economic History, 49, 857–882. Greif, A. (1993). Contract enforceability and economic institutions in early trade: the Maghribi traders? coalition. American Economic Review, 83, 525–548.

Greif, A. (1994). On the political foundations of the late medieval commercial revolution: Genoa during the twelfth and thirteenth centuries. Journal of Economic History, 54, 271–287. Griggs, B. (2003). Identity theft, other cybercrimes show no sign of slowing in Utah. Knight Ridder Tribune Business News, 1. Han, K.-H. (2003, June 27). Brother and Sister partners on the net, Scam NT$ 5 Million on Yahoo auction. United Evenings. Hanneman, R. (1998). Introduction to Social Network. (Retrieved 3.07.2003). Hogg, T., & Adamic, L. (2004). Enhancing reputation mechanisms via online social networks. Paper presented at the 5th ACM conference on Electronic commerce, New York, NY, USA. Johnson, M. P., & Milardo, R. M. (1984). Network interference in pair relationships: a social psychological recasting of Slater’s (1963) theory of social regression. Journal of Marriage and the Family, 46(November), 893–899. Klein, S. (1997). Introduction to electronic auctions. Electronic Markets, 7(4), 3–6. Krebs, V. E. (2002). Mapping networks of terrorist cells. Journal of the International Network for Social Network Analysis, 24(3), 3–52. Liao, S.-H. (2005). Expert system methodologies and applications – a decade review from 1995 to 2004. Expert Systems with Applications, 28(1), 93–103. Li, X., Liu, L., Wu, L., & Zhang, Z. (2006). Predicting the final prices of online auction items. Expert Systems with Applications, 31(3), 542– 550. Linden, G., Smith, B., & York, J. (2003). Amazon.com recommendations: item-to-item collaborative filtering. IEEE Internet Computing. Little, B. B., Johnston, W. L., Lovell, A. C., Rejesus, R. M., & Steed, S. A. (2002). Collusion in the US crop insurance program: applied data mining. Paper presented at the Eighth ACM SIGKDD international conference on Knowledge discovery and data mining. Lucking-Reiley, D. (1999). Auctions on the Internet: what’s being auctioned, and how? (Working paper): Vanderbilt University. Lucking-Reiley, D., Bryan, D., Prasad, N., & Reeves, D. (2001). Pennies from eBay: the Determinants of Price in Online Auctions. Working paper, University of Arizona. Matsubara, S., & Yokoo, M. (2002). Defection-free exchange mechanisms based on an entry fee imposition. Artificial Intelligence, 142(2), 265–286. McMillan, J., & Woodruff, C. (2001). Private order under dysfunctional public order. Michigan Law Review, 98, 2421–2458. Miller, B. N., Albert, I., Lam, S. K., Konstan, J. A., & Riedl, J. (2003). MovieLens unplugged: experiences with an occasionally connected recommender system. Paper presented at the international conference of intelligent user interface. Min, S.-H., & Han, I. (2005). Detection of the customer time-variant pattern for improving recommender systems. Expert Systems with Applications, 28(2), 189–199. Mitchell, J. C. (1969). The concept and use of social networks. Manchester: Manchester University Press. Nieminen, V. (1974). On centrality in graph. Scandinavian Journal of Psychology, 15. Page, L., Brin, S., Motwani, R., & Winograd, T. (1998). The PageRank citation ranking: Bringing order to the web (technical report). Stanford, CA: Stanford University. Peddy, C. C., & Armentrout, D. (2003). Building solutions with Microsoft Commerce Server 2002. Microsoft Press. Potter, G. W. (1994). Criminal organizations: Vice, racketeering, and politics in an American city. Prospect Heights, IL: Waveland. Reffay, C., & Chanier, T. (2003, June). How social network analysis can help to measure cohesion in collaborative distance-learning. Paper presented at the computer supported collaborative learning conference (CSCL’2003), Bergen, Norway. Resnick, P., Kirwabara, K., Zeckhauser, R., & Friedman, E. (2000). Reputation systems. Communications of the ACM, 43(12), 45–48. Resnick, P., & Zeckhauser, R. (2002). Trust among strangers in Internet transactions: empirical analysis of eBay’s reputation system. In R.

J.-C. Wang, C.-C. Chiu / Expert Systems with Applications 34 (2008) 1666–1679 Michael Baye (Ed.), Advances in applied microeconomics: The economics of the Internet and e-commerce. Amsterdam: Elsevier Science. Sabater, J., & Sierra, C. (2002). Reputation and social network analysis in multi-agent systems. Paper presented at the autonomous agents and multi-agents systems, Bologna, Italy. Scott, J. (2002). Social networks: Critical concepts in sociology. Routledge Publication Inc.. Shah, H. S., Joshi, N. R., Sureka, A., & Wurman, P. R. (2002). Mining for bidding strategies on eBay. Paper presented at the WEBKDD 2002 Web mining for usage patterns and user profiles. Shardanand, U., & Maes, P. (1995). Social information filtering: Algorithms for automating ‘Word of Mouth’. Paper presented at the conference of human factors in computing systems. Snyder, J. M. (2000). Online auction fraud: are the auction houses doing all they should or could to stop online fraud? Federal Communications Law Journal, 52(2), 453–472. Sparrow, M. K. (1991). The application of network analysis to criminal intelligence: an assessment of the prospects. Social Networks, 13(3), 251–274. Swan, W. (2001). Social network analysis in construction project teams. (Retrieved 10.05.2003). Tsai, W. (2002). Social structure of coopetition within a multiunit organization: coordination, competition, and intraorganizational knowledge sharing. Organization Science, 13(2), 179–190. Turban, E. (1997). Auctions and bidding on the internet: an assessment. Electronic Markets, 7(4), 7–11. TWINC, T.N.I.C. (2005). Internet Broadband Usage in Taiwan (No. 0507). Taipei, Taiwan: Taiwan Network Information Center. Tyler, J. R., Wilkinson, D. M., & Huberman, B. A. (2003). Email as spectroscopy: automated discovery of community structure within organizations. (Retrieved 20.07.2003).

1679

Wang, J.-C., & Chiu, C.-C. (2005). Detecting online auction inflatedreputation behaviors using social network analysis. Paper presented at the North American Association for Computational Social and Organizational Science, Notre Dame, Indiana. Wang, W., Hidve´gi, Z., & Whinston, A. B. (2001a). Designing mechanisms for e-commerce security: an example from sealed-bid auctions. International Journal of Electronic Commerce, 6(2), 139–156. Wang, W., Hidve´gi, Z., & Whinston, A. B. (2001, October 1). Shill bidding in multi-round online auctions. Paper presented at the 35th Hawaii international conference on system sciences, Hawaii. Warner, M. (2003). eBay’s worst nightmare. Fortune, 147(10), 89. Wellman, B. (1996). For a social network analysis of computer networks: a sociological perspective on collaborative work and virtual community. Paper presented at the ACM SIGCPR/SIGMIS conference. Wheeler, R., & Aitken, S. (2000). Multiple algorithms for fraud detection. Knowledge-Based Systems, 13(2), 93–99. Whitten, N. E., & Wolfe, A. W. (1973). Network Analysis. Wilke, J. R., & Wingfield, N. (2003). Leading the news: fraud crackdown hits Web auctions – federal and state officials to make arrests, file suits EBay cooperates in probe. Wall Street Journal, A.3. Williams, P. (2001). Transnational criminal networks. In J. Arquilla & D. Ronfeldt (Eds.). Networks and netwars: The future of terror, crime, and militancy (pp. 61–97). Santa Monica, CA: RAND. Xu, J.-J., & Chen, H. (2002). Using shortest path algorithms to identify criminal associations. Paper presented at the 2002 national conference on digital government research. Yahoo! Inc. (2004). Yahoo! Help – Listing Fees. (Retrieved 6.03.2004). Zacharia, G., Moukas, A., & Maes, P. (1999, January 5–9). Collaborative reputation mechanisms in electronic marketplaces. Paper presented at the 32nd Hawaii international conference on system sciences, Electronic Commerce Minitrack, Maui, Hawaii.