Experimental Study of Scalability Enhancement for Reverse Logistics e-commerce

Transcription

1 Chapter 1 Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce Sarah M. Ryan K. Jo Min Sigurdur Oafsson Abstract: Key words: Methods of improving scaabiity in onine auctions incude imiting the number of bidding opportunities, providing price information to users, and recommending auctions that may be of interest to the users. We constructed an experimenta prototype auction system in the context of reverse ogistics for eectronics products. Experiments were designed to test the effects of the number of trading opportunities and the amount of previous price and bid information presented to users. The participants profits improved with the number of trading opportunities but showed mixed effects for increasing price and bid information. The induction of decision trees for an auction recommender is iustrated aong with the use of attribute seection to reduce the size of the tree. Business-to-business e-commerce, empirica research, recommender systems 1. INTRODUCTION In recent years, business to business e-commerce has emerged as an important means to enabe efficient exchange, reduce marketing costs for seers and procurement costs for buyers, and unify fragmented markets. Severa e-commerce portas have been created for we-defined markets such as automotive parts (Covisint.com, 2001). On the other hand, a fragmented market is characterized by arge, dynamic sets of buyers and seers who are 1

2 2 generay unfamiiar with each other. One exampe is the market for postretai and surpus items such as used products, returned and refurbished products, and surpus inventories of unsod products nearing obsoescence. In particuar, because of their short ifecyces, arge quantities manufactured and in use, and potentia generation of hazardous waste, eectronic products constitute one important category. More generay, reverse ogistics can be defined as the process of panning, impementing and controing the efficient, cost effective fow of raw materias, in-process inventory, finished goods and reated information from the point of consumption to the point of origin for the purpose of recapturing vaue or proper disposa. It aso incudes processing merchandise returned because of damage, seasona inventory, restocking, savage, recas and excess inventory. Our focus is on moving goods from their typica fina destinations for the purposes of capturing vaue or proper disposa. Remanufacturing and refurbishing activities as we as recycing for materia recovery are common ways of capturing vaue. Hazardous materia content may be of concern aso. Kokkinaki et a. (2001a) have identified severa ways that eectronic commerce coud hep to unify the markets in reverse ogistics. Kokkinaki et a. (2001b) presented a prototype eectronic marketpace for PCs incuding decision support for reuse, remanufacturing or recycing. The designers of an e-commerce system meant to unify a fragmented market must consider (1) how to encourage participation in the onine marketpace, and (2) how to ensure that the system can scae to hande increased traffic as participation grows. This paper focuses on the roes of exchange mechanisms and information deivery in achieving these two reated goas. Potentia participants wi be encouraged to trade in an onine marketpace if they perceive that the onine system makes exchange easier and/or more profitabe than exchanges outside it woud be. The system coud speed the identification of potentia trading partners, provide an efficient exchange mechanism, and deiver usefu information about recent prices of items whie reducing the need for time-consuming onine searches. By performing much of the time-consuming information processing offine or in periods of reduced traffic, these efforts to expand the market coud simutaneousy enhance the scaabiity of the system. The information deivery methods studied in this paper can be better understood by contrast to a popuar consumer-oriented onine marketpace. Suppose you drop your persona digita assistant (PDA) on the foor and crack the screen. After a ca to the manufacturer reveas a fairy steep repair price for your discontinued mode, you decide to seek a repacement. Since you own severa accessories for your od PDA, you are ooking for an identica used or surpus mode and, knowing the repair price, you have a

3 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 3 ceiing on the amount you are wiing to pay. A search for this mode on ebay.com yieds one thousand records of ongoing auctions, most of which are for accessories rather than the actua PDA. By aboriousy examining pages of auctions, you choose severa soon-to-end auctions and, by repeated page updates, watch them to earn the settement prices. Then you choose an auction to enter and start bidding. You can enter your maximum price and et the system bid up for you. In addition, you can ask to be notified when you are outbid. But you soon earn that in order to win your repacement PDA, you must cosey monitor the auction at its finish, which generay occurs at an inconvenient time, and repeatedy enter new bids. You are ikey to give up and pay the manufacturer s repair price. Suppose in contrast that you were a reguar buyer of used PDAs and that you coud participate in a system that proactivey notified you when an auction for your preferred mode began. This system coud inform you of current prices without repeated searches and carefu monitoring of auctions on your part. Armed with this reiabe price information, you coud enter a singe bid and then go about your business whie the auction was resoved, knowing that other potentia buyers were imited to a singe bid as we. Assuming you were satisfied with the auction outcomes, such a system woud encourage your repeated business and woud simutaneousy save itsef the burden of compying to numerous search requests, page refreshes, and bid updates (both automatic and manua). We woud expect that price information and imits on bidding opportunities woud significanty affect buyer and seer behavior, the webeing of participants and the market as a whoe. However, these effects have not been studied. In addition, the abiity of knowedge discovery methods to reiaby cassify auctions as interesting to a particuar buyer or not has not been rigorousy tested. This paper describes an empirica study of a simuated business to business e-commerce system. Experiments were conducted to test the effects of different price information dispays and different numbers of bidding opportunities. Decision trees for an auction recommender system were constructed using the data coected. The resuts of data anaysis and data mining for these initia experiments iustrate the promise of the recommender system and suggest severa intriguing questions about auction design to expore in further research. Internet auction has been extensivey examined as a way to aocate goods and services effectivey and efficienty (see e.g., Kumar and Fedman (1998) or Teich, Waenius, and Waenius (1999)). In this paper, we empoy an auction mechanism caed seaed bid doube auction (see e.g., Wurman, Wash, and Weman, (1998)). Under this auction mechanism, each of mutipe potentia seers and mutipe potentia buyers can submit a bid ony

4 4 once during the auction. Each bid is seaed so that ony the one that has submitted the bid has any knowedge of the bid throughout the auction process. This auction is, conceptuay, one of the simpest auction modes in use today. This mode reativey (reative to, for exampe, the Engish auction mode; see e.g., McAfee and McMian (1987)) induces participants to revea their true vauation. Considering such factors as the conceptua simpicity, current usage in Internet auction, and reative reveation of true vauation, it is an idea standard auction mechanism. That is, this mechanism can be used as a benchmark mode in the future when other auction mechanisms are comparativey studied. 2. DESCRIPTION OF EXPERIMENTAL PROTOTYPE For our experiment, we constructed a simuated auction environment. With emphasis on eectronics recycing, our business-to-business (B2B) e- commerce participants consist of three Manufacturers, three Demanufacturers, and three Recycers. Each participant is given a prespecified eve of cash and reevant inventory of items, and each participant s objective is to maximize the eve of cash at the end of the auction process. At the end of the auction process, any eftover inventory is worthess. The auction process consists of severa rounds, in each of which there is an open auction for each type of item. In each round, a participant can bid a price and quantity to se (of items the participant has) and a price and quantity to buy (of items the participant wants). Any cost and price factors that are not expicity discussed in the next section are negigibe. 2.1 Roes of Manufacturer, Demanufacturer, and Recycer Each Manufacturer produces coffeemakers by using one CPU and one unit of Pastic. The Manufacturer can se any number of coffeemakers at a fixed unit price of $20. The CPUs and the units of Pastic can be bought in any round of the auction process. At the end of each round, we assume that equa quantities of any CPUs and units of Pastic the Manufacturer has are immediatey transformed into coffeemakers. Finay, we assume that the Manufacturer is provided with an initia inventory of obsoete Computers, which can be sod in any round of the auction. A Demanufacturer can buy obsoete Computers in any round of the auction process. As soon as the Demanufacturer gets them, they are

5 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 5 immediatey disassembed into memory boards, CPU s and Shes. Each memory board is sod outside the auction system at $4 per unit. Aso, if the Demanufacturer does not manage to obtain a fixed quota of memory boards, then the Demanufacturer wi have to pay a cash penaty of $2 for each board the Demanufacturer is short after the ast round. We assume that the Demanufacturer is provided with an initia inventory of CPUs and Shes. Such inventory as we as the CPUs and Shes from the disassemby of the obsoete Computers can be sod in any round of the auction. Finay, a Recycer can buy Shes in any round of the auction process. As soon as the Recycer gets them, they are meted down instantaneousy into units of recyced Pastic. Each Recycer is provided with an initia inventory of Pastic. Again, we wi assume that the initia inventory of units of Pastic as we as the units of recyced Pastic can be sod in any round of the auction process. Figure 1 depicts transformations of products and materias by the three types of participants. Figure 2 iustrates the exchange of items within the onine marketpace and externay. 1 CPU 1 unit Pastic Manufacturer 1 Coffeemaker 1 Computer Demanufacturer 1 CPU 1 Memory Board 1 She Recycer 1 She 1 unit Pastic Figure 1. Reationships among products traded onine (bod) and sod externay.

6 6 B2B Marketpace Manufacturer Coffeemakers Computers CPUs Pastic Memory Demanufacturer Shes Recycer Figure 2. Reationships among participants in the onine marketpace 2.2 Seaed Bid Doube Auction (SBDA) Mechanism The auction process consists of either 3 rounds ( short process) or 6 rounds of SBDA ( ong process). In each round, for a pre-specified kind of product, each potentia seer or buyer can submit a bid consisting of the desired quantity and the corresponding price that the potentia buyer is maximay wiing to pay or the potentia seer is minimay wiing to accept. At the concusion of a pre-specified bid submission period for each round, the auction processing unit rank orders a buy bids according to the maximay acceptabe buying price (MABP) and a se bids according to the minimay acceptabe seing price (MASP). By matching rank ordered MABP s and MASP s as we as their respective bid quantities, the minimum MABP and the maximum MASP that maximizes the tota transaction quantity can be found. The settement price is the average of the minimum MABP and the maximum MASP. If the tota quantity demanded with the minimum MABP or higher is not equa to the tota quantity suppied with the maximum MASP or ower, which shoud be the typica case, the foowing priority is given: Among buyers, the buyers with higher than the minimum MABP have the priority. Among seers, the seers with ower than the maximum MASP have the priority. It is quite possibe that a buyer or a seer without priority may not be abe to buy or se a the quantity specified in the bid. We wi assume that such a buyer or seer is wiing to buy or se ess than the exact quantity specified in the bid. Finay, if two or more buyers submit

7 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 7 the same MABP or two or more seers submit the same MASP, then the buyer or seer who submitted the bid earier has the higher priority. According to the settement price and the reguations just described, the transaction records are reveaed to a the seers and buyers with winning bids. The auction process terminates after three rounds (short process) or six rounds (ong process) of such SBDA. 3. EXPERIMENTAL DESIGN We designed the experiments to test the effects of two factors: the information eve and the number of rounds in the auction. Each experiment consisted of one auction process. The information factor, INFO, had three eves. The MIN eve dispayed the settement price of each auction ony to the winning buyer and seer. The MED eve dispayed the settement price of each auction to a participants in the auction. The MAX eve showed each participant in an auction not ony the settement price but aso the price bids entered by his/her competitors (for a buyer, the other buy bids and for a seer, the other se bids). The number of rounds (K) in each experiment had two eves: K = 3 for a short process or 6 for a ong process. The eve of 3 was chosen as the minimum number of rounds necessary for Computers sod by the Manufacturers to be transformed into Pastic avaiabe for their purchase. Six rounds were deemed to be ong enough for the market to make maximum use of the starting inventories. The subjects for the experiments were undergraduate engineering students at Iowa State University. We recruited 27 participants and randomy divided them into three groups, each of which participated in five experiments over two sessions. The first session for each group began with training and some practice rounds. Each subject adopted a different roe (M, D, or R) for each experiment conducted in a given session. We conducted four experiments of each of the factor combinations 3*MED, 6*MIN, 6*MAX and one experiment of each of the factor combinations 3*MIN, 3*MAX, 6*MED. The combinations of factor eves were randomy assigned to subject groups and sessions. 3.1 Data Anaysis Onine auctions provide a weath of data from which we can study the behavior of participants, the evoution of prices, consensus about prices or the ack of it, and the overa economic outcomes for participants and the

8 8 market as a whoe. Let i be the index for participants where i M = { 1, 2, 3} for the Manufacturers, i D = { 4,5,6} for the Demanufacturers and i R = { 7,8,9} for the Recycers. The item type is denoted by j = 1 for Computers, j = 2 for CPUs, j = 3 for Shes and j = 4 for Pastic. The index = 1,, L identifies the experiment (process) and k = 1,, K refers to the auction round, where K is either 3 or 6. For each experiment = 1,, L, we observed vaues for the foowing variabes: p = se price bid by participant i for item j in round k q ijk ijk ( p > 0 if se bid, < 0 if buy bid ) ijk = se quantity bid by participant i for item j in round k ( q ijk > 0 if se bid, < 0 if buy bid ) r ijk = quantity sod by participant i of item j in round k t jk ( r ijk > 0 if sod, < 0 if bought ) = quantity that changes hands of item j in round k s jk = settement price of item j in round k jk ( s > 0 if the auction was successfu, = 0 if the auction faied ) d ik = cash gain for participant i in round k (oss if negative), which incudes revenue from saes of coffeemakers and memory boards. In the fina round, any penaties for Demanufacturers are appied aso. From these recorded vaues we computed aggregate measures of the auction outcomes. In the foowing, I [] is an indicator function that equas 1 if the quantity in brackets is true and 0 otherwise. T j K k = 1 jk = t is the tota quantity of item j exchanged 4 K F = I s jk = 0 is the number of auctions that faied j= 1 k= 1

9 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 9 i K k = 1 [ brokering by participant in round of expt. ] B = I i k, i = B T 1,,9. Brokering attempts were defined according to Tabe 1. 9 L = Bi is the tota number of brokering attempts by a i= 1 participants. Let I be the initia cash hed by participant i. Based on experience with i 0 preiminary tests, we attempted to equaize the expected gain in weath for the three participant types by setting I i 0 at successivey higher vaues for Manufacturers, Demanufacturers and Recycers, respectivey. In addition, each participant started with some initia stocks of items: for j = 1,, 4, I ij is the initia stock of item j hed by participant i. We vaued these initia inventories using estimates of their vaue over a experiments. The vaue of L K L K item j was estimated as w = 0 j s I s = 1 k= 1 jk = 1 k= 1 jk. Then the net gain (oss if negative) for participant i in the th experiment was computed as K 4 i = ik j ij i0 k= 1 j= 1. G d w I I In order to study the degree of consensus about item prices, we aso computed the squared deviation of the bids about the settement price in the successfu auctions: Vjk = pijk sjk if s 0. jk ip : ijk 0 ( ) 2 Tabe 1. Identification of suspected brokering behavior by participant and item type j = 1 Computer j = 2 CPU j = 3 She j=4 Pastic i M p > 0: norma p < 0: broker p>0:broker p < 0: norma p 0: broker p > 0: broker p < 0: norma i D p > 0: broker p > 0: norma p>0: norma p 0: broker p < 0: norma p < 0: broker p < 0: broker i R p 0: broker p 0: broker p > 0: broker p < 0: norma p > 0: norma p < 0: broker Because we observed some quaitative differences in the three subject groups (for instance, one group appeared quiet and studious, whie another

10 10 was takative), we first checked for any statisticay significant differences in the outcomes. First we anayzed the overa outcomes of the experiments. 4 For each of the response variabes T 1, T 2, T 3, T 4, Q = T T, j= 1 j B, and F, T we tested the nu hypothesis that the subject groups were identica using a nonparametric Kruska-Wais test (n = 15). The subject group showed a significant effect ony on Q T, the tota quantity of materia traded. Next we examined outcomes for individua participants. We anayzed the net gain or oss for each participant according to their type, G M, G D, G R ; as we as the number of brokering attempts by each individua participant. Of these six response variabes, with 45 observations, the subject group membership was significant ony for G M, the gain by Manufacturers. Despite these two warnings that the subject groups were not competey homogeneous, we fet that since homogeneity coud not be rejected for most of the performance measures, we coud group a the responses across subject groups for the remainder of the anaysis in reasonabe safety. In this preiminary study, the sizes of the data sets were reativey sma. The sampe sizes were L = 15 experiments for the responses T j, F, and B T and n = 45 for G P and B P, where P = M, D, or R (15 experiments with 3 payers of each type). Because assumptions of normaity seemed untenabe in most cases, we reied on nonparametric tests as much as possibe. To anayze settement prices and bid variation, we excuded any rounds in which the settement price was zero, i.e., the auction had faied. We ooked at the effect of the round number and found that if we aso excuded the ast round, in which a participants had the incentive to dump excess inventory and no one was motivated to buy Shes, the hypotheses that the round number had no effect on either settement price or bid variation coud not be rejected. Finay, in the bid variation anaysis we eiminated two outier bids (in two separate experiments) that were an order of magnitude arger than a the other bids. In the end, we had between 50 and 56 rounds of auctions, depending on the item, over which to anayze prices and bids. More detais of the statistica anaysis, incuding p-vaues for hypothesis tests, can be found in Ryan, Min and Oafsson (2002). 3.2 Resuts The number of trading opportunities (auction rounds in each experiment) coud be expected to affect the webeing of participants as we as the heath of the market. Based on the profit (gain) for each payer i, we make the foowing observations. When we compare the profits from the three round auction with those from the six round auction, for a three groups of payers (Manufacturers, Demanufacturers, and Recycers), the average profit is

11 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 11 substantiay higher from the six round auction. We beieve that this resut occurred because of the substantia expansion of the set of feasibe strategies in the case of the six round auction. The onger experiments aowed more materia to change hands and aso appeared to encourage brokering behavior, particuary by Recycers. However, we observed that, athough some brokering attempts were made by a groups of participants, in no case did a payer manage to se items in a ater round for a higher price than he had paid for the same items earier. With a coupe of exceptions, the number of rounds had itte effect on settement prices or bid variation. The mean settement price for CPUs was significanty arger for K=3 than for K=6 rounds. The mean variation of bids around the settement price was arger for CPUs but smaer for Shes in 3 rather than 6 rounds. The settement price information reveaed to traders and woud-be traders coud have effects on their behaviors and their eventua economic webeing. A trader who was abe to use this information effectivey might profit by brokering items she did not want or simpy by bidding more effectivey on items she did want. We originay considered the MED eve of information to be the typica amount of price information one woud expect when participating in an auction (for instance, this is the information reveaed by ebay). Surprisingy, we found that by some measures, participants tended to be worse off when presented with this information than when they received either ess or more information about past settement prices and bids. Generay speaking, the settement prices and bid variations were simiar under MIN and MAX information but different under the MED eve of information. Due to the sma sampe sizes, not a these effects were statisticay significant. Figure 3 iustrates the effect of information on settement prices and squared deviations of bids about the settement price for two different items.

12 12 Shes Settement Price with Max Info Pastic Bid Dev with Max Info Shes Settement Price with Med Info Pastic Bid Dev with Med Info Shes Settement Price with Min Info Pastic Bid Dev with Min Info Figure 3. Histograms of settement prices of Shes and variation in bids for Pastic for different information eves In particuar, the mean settement price for Computers took its smaest vaue under MED information, whie the settement prices for the other items were arger under MED information than under either MIN or MAX information. For each item, the mean bid variation was smaest under MED information. These resuts seem to suggest that by giving a payers information about the settement price, but not competitors bids, the auction administrator coud steer prices for items either higher or ower. Higher prices coud be expained by a conjecture that, in contrast to MIN information, MED information motivates a potentia buyer who faied to obtain items previousy to bid higher than the known settement price. Subsequent bids are steered to a narrow range just above the previous settement price. However, if this buyer aso earns the range of buy bids in the previous auction, he may fee more comfortabe bidding somewhat ower than the previous settement and, in any case, wi not be steered to such a tight interva. Seers woud be motivated to bid in the opposite direction; however, in our setup there is more pressure for Manufacturers and Demanufacturers to buy parts and materias than there is to se.

13 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 13 When we compare participant gain across the amount of information, the MIN information resuts in the highest average profit for a three groups. For the Manufacturer group and the Recycer group, the MAX information resuts in the second highest average profit. On the other hand, for the Demanufacturer group, the average profits under MAX and MED information differ by ony 2.7%. Finay, we observe that the tota average profit of a three groups under MIN information is substantiay higher than that under MAX information. The tota average profit of a three groups under MAX information, in turn, is substantiay higher than that under MED information. Mfg Gain/oss with Max Info Tota Materia with Max Info Mfg Gain/oss with Med Info Tota Materia with Med Info Mfg Gain/oss with Min Info Tota Materia with Min Info Figure 4. Histograms of gain/oss by Manufacturers and tota materia traded by information eve. These observations seem to suggest that the MED information case not ony eads to a reativey narrow range of settement prices, but aso resuts in a reativey sma amount of profit for a groups of payers. If the auction administrator is paid by a fraction of the profit, these resuts impy that it

14 14 shoud strive to provide either MIN or MAX information, and to avoid providing MED information (though providing MAX information may be necessary especiay if potentia seers and buyers demand it). However, the MIN information eve may be seen as more consistent with the seaed bid auction environment. The tota amount of materia that changes hands during the auction process is smaest under MED information. Figure 4 iustrates the gain/oss by Manufacturers and the tota materia traded by information eve. Finay, the amount of brokering by a participants and particuary by Recycers (who had the east reguar trading to do) was east under MED information. These outcomes suggest that settement price information aone somehow discourages trading and does not inspire confidence of traders in their abiity to successfuy broker items. 4. AUCTION RECOMMENDER In this section we consider what can be earned about the preferences of the auction participants using knowedge discovery and data mining techniques. In particuar, we are interested in what auctions might interest each user and shoud thus be identified or recommended. A recommender is a system that can identify which products and hence auctions are of interest to a particuar user, and dispay the information that is reevant for the user to participate in such an auction. Such systems have received considerabe attention in the iterature with eary recommender systems generay based on simpe nearest-neighbor coaborative fitering agorithms (Resnick et a., 1994; Shardanand and Maes, 1995). Later versions have used other earning methods, incuding Bayesian networks (Breese et a., 1998), cassification methods (Basu et a., 1998), association rues (Lin et a., 2000), regression (Vucetic and Obradovic, 2000), and inteigent agents (Good, 1999). The key to the successfu impementation of such systems is knowing the participants. The more intimate this knowedge, the more ikey it is that the system wi be abe to recommend the correct products. To gain as much knowedge as possibe about the behavior of the auction participants, a data set is constructed that describes the detais of their behavior for every auction as we as the characteristic of the auction itsef. The foowing attributes, which are shown aong with their possibe vaues, are used in this set:

15 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 15 Product Description 1) Item {Computer,CPU,She,Pastic} 2) ItemContainsCPU {T,F} 3) ItemContainsShe {T,F} 4) ItemContainsPastic {T,F} Round Information 5) Round {1,2,3,4,5,6} Information on Last Transactions to Occur 6) PriceForLastComputerAuction numeric 7) QuantityForLastComputerAuction numeric 8) PriceForLastCpuAuction numeric 9) QuantityForLastCpuAuction numeric 10) PriceForLastSheAuction numeric 11) QuantityForLastSheAuction numeric 12) PriceForLastPasticAuction numeric 13) QuantityForLastPasticAuction numeric Information on Last Transaction with Participants Invovement 14) TimeOfParticipantsLastComputerAuction numeric 15) PriceForParticipantsLastComputerAuction numeric 16) QuantityForParticipantsLastComputerAuction numeric 17) CategoryOfParticipantsLastComputerAuction {buy,se,none} 18) TimeOfParticipantsLastCpuAuction numeric 19) PriceForParticipantsLastCpuAuction numeric 20) QuantityForParticipantsLastCpuAuction numeric 21) CategoryOfParticipantsLastCpuAuction {buy,se,none} 22) TimeOfParticipantsLastSheAuction numeric 23) PriceForParticipantsLastSheAuction numeric 24) QuantityForParticipantsLastSheAuction numeric 25) CategoryOfParticipantsLastSheAuction {buy,se,none} 26) TimeOfParticipantsLastPasticAuction numeric 27) PriceForParticipantsLastPasticAuction numeric 28) QuantityForParticipantsLastPasticAuction numeric 29) CategoryOfParticipantsLastPasticAuction {buy,se,none} Information on Participant 30)ParticipantName {M1,M2,M3,D1,D2,D3,R1,R2,R3}

16 16 31)ParticipantType {Manufacturer,Demanufacturer, Recycer} Cass Attribute 32) ParticipateInAuction {yes,no} From the experiments, 2592 data objects or instances were obtained, one for each user for each auction of every round. Each of these instances has a vaue for a of the 32 above mentioned attributes, incuding the cass attributes, that is equa to yes if the user participated in the auction and no otherwise. The knowedge discovery task of the recommender system is to reate the cass attribute with the others, a of which are observabe before an auction must be recommended. 4.1 Constructing a Recommendation Before using data mining to construct a mode for recommendation, et s consider how a very simpistic mode might be constructed. As the basic roes of each type of participant are known, for exampe a Manufacturer needs to se Computers and buy CPUs and Pastic to make coffeemakers, one might be tempted to aways recommend Computer, CPU, and Pastic auctions to a Manufacturer and never recommend any of the She auctions. However, this simpistic mode ignores the facts that a Manufacturer s interest in Computers, CPUs and Pastic changes as the auction evoves, and that the Manufacturer might want to act as a broker for Shes as we as Computers. Since past behavior of the users is avaiabe, data mining techniques can be used to construct a more inteigent recommender. Participant Type Manufacturer Demanufacturer Recycer Item Item Contains Pastic Item Contains CPU Figure 5. Top of the recommender decision tree The reevant data mining probem for this context is cassification. Cassification is the process of buiding a mode that describes a priori specified casses or concepts based on data objects with known

17 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 17 cassifications, and then using this mode to cassify new data objects where the cass is unknown. Here the concept to be earned is whether or not a user wants to participate in an auction. The primary performance measure of cassification modes is its accuracy, but its scaabiity in terms of induction speed and use of memory is aso a critica issue. The most common approach for cassification is decision tree induction using agorithms such as C4.5 (Quinan, 1993) and CART (Breiman, Friedman, Oshen, and Stone, 1984). To buid a recommender for the auction system a decision tree mode is induced using the C4.5 decision tree agorithm. At the top the decision tree spits according to participant type (see Figure 5), so the three branches can be thought of as decision trees for a Manufacturer, Demanufacturer and a Recycer. Let s consider the tree in Figure 6, which shows the branch or decision tree for a Manufacturer. Ceary, this decision tree impies the need for a much more compicated recommendation mode than the simpistic mode above. As an exampe, et s read off the tree when a Computer auction shoud be recommended to a Manufacturer. The foowing rues can be directy inferred from the tree: If the time since the Manufacturer ast participated in a Computer auction is ess than or equa to one, then recommend the auction. Otherwise, if the time since the Manufacturer ast participated in a Computer auction is more than one, and the Manufacturer bought the Computers ast time, then recommend the auction. Otherwise, do not recommend the auction.

18 18 1 Buy Computer Time Since M s Last Computer > 1 Category of M s Last Comp. Auction Se ne CPU Se Item She Time Since M s Last She Auction > 0 Time Since M s Last Computer 1 Category of M s Last Pastic Auction Buy Time Since M s Last Pastic Auction 1 > 1 ne 0 (never) > 1 Pastic Buy 1 Time Since M s Last Pastic Auction Quantity of M s Last CPU Auction > 80 Price for M s Last CPU Auction 3 > 3 Category of M s Last Comp. Auction Se 80 ne > 1 Time Since M s Last CPU Auction 2 > 2 Price for M s Last CPU Auction 7.76 Price for M s Last Pastic Auction 8.25 Quantity for Last Computer Auction > > 7.76 > 8.25 Figure 6. Branch of decision tree for a Manufacturer Intuitivey, one might expain these rues as foows. If the Manufacturer participated in a Computer auction ast time (or has never participated), this Manufacturer is ikey to be sti trying to se Computers and thus the auction shoud be recommended. If the Manufacturer did not participate ast time, this indicates that they have aready sod a of their Computers. The auction is ony recommended if they acted as brokers and bought Computers before, in which case they are ikey to want to continue in their broker roes. In this case, the recommender discovers two interesting nuggets of information:

19 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce A Manufacturer may become disinterested in Computer auctions after a Computers have been sod and thus these auctions shoud no onger be recommended. 2. However, if a Manufacturer wishes to act as a broker and aso buy Computers, then Computer auctions shoud continue to be recommended for future rounds. This is an exampe of discovering structura knowedge. The decision tree not ony provides a predictive mode but aso adds new insights into the decision making process of Manufacturers. Finay, note that these rues give a cassification for each auction being either recommended or not recommended, but they do not give any confidence assessment of the quaity of the recommendation. w et s consider when a She auction shoud be recommended to this Manufacturer. The foowing rues can be read off the tree: If this Manufacturer has never participated in She auctions, then do not recommend the auction. Otherwise, if there has been more than one round since the ast participation in a Computer auction, do not recommend the auction. Otherwise, if the Manufacturer s ast participation in a Pastic auction was to se Pastic, then do recommend the auction. Otherwise, if the time since the Manufacturer s ast participation in a Pastic auction is ess than one round, recommend the auction. Otherwise, do not recommend the auction. A Manufacturer wi ony participate in a She auction to act as a broker, and these rues ceary attempt to detect brokering behavior. For exampe, uness this Manufacturer has brokered Shes before, She auctions are not recommended. Furthermore, other behavior such as the Manufacturer seing Pastic is taken as an indication of a wish to serve as a broker and thus She auctions are recommended. Finay, note that determining the interest in a Pastic auction is the most compicated of a and requires considering the quantity, price, and timing of previous CPU auctions, brokering behavior as indicated by buying Computers, settement price of previous Pastic auctions, and the amount of Computers sod most recenty. In particuar, note the importance of the detaied behavior regarding CPU auctions, which appears to refect the fact that CPUs and Pastic combine to make coffeemakers.

20 20 Item Contains CPU True Fase Time Since R s Last Pastic Auction 1 > 1 Time Since R s Last She Auction Time Since R s Last She Auction 1 > 1 1 > > Pastic Item Quantity For R s Last She Auction She 1100 Pastic Buy Item Category of R s Last Computer Auction ne She Se Figure 7. Partia branch for Recycer w et s consider the decision of recommending an auction to a Recycer. A partia decision tree for this decision is shown in Figure 7. te that the tree first spits on the question of whether the item contains CPU or not. If the item contains CPU, that is the item is CPU or Computer, then the auction is not recommended, which impies that an auction participation soey for the purpose of being a broker is never recommended. If the item does not contain CPU, that is the item is Pastic or She, then the auction may or may not be recommended and the next two nodes spit by determining the time since the ast Pastic and She auction in which the Recycer participated. For exampe, if the Recycer participated in a Pastic auction ast time but did not participate in a She auction ast time then the auction is not recommended if it is for Shes and is ony recommended for Pastic if the She quantity of the Recycer obtained from the ast auction is arge (1100 units). These rues appear to be discovering the situation that the Recycer is no onger interested in buying more Shes and is ony trying to se the current inventory of Pastic. We finay note that the branch corresponding to Demanufacturer is simiar in its discovery and, as indicated by Figure 5, starts by spitting according to whether the item contains

21 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 21 Pastic, that is, if the item is Pastic in which case it woud ony be of interest for brokering, or if it is Computer, CPU, or She, in which case it woud be of primary interest. Attribute Seection In this simpe system the number of attributes is ony 32 and is thus quite manageabe for any standard induction agorithm. However, it is cear that this number of attributes wi increase very rapidy. Let s first consider how many transactiona attributes there are regarding a singe item, such as a Computer: PriceForLastComputerAuction QuantityForLastComputerAuction TimeOfParticipantsLastComputerAuction PriceForParticipantsLastComputerAuction QuantityForParticipantsLastComputerAuction CategoryOfParticipantsLastComputerAuction Thus, for every new item there are six new transactiona attributes, resuting in a very rapid exposion of the number of attributes as the number of items handed by the system increases. Next, et s consider the disassemby information: ItemContainsCPU ItemContainsShe ItemContainsPastic Here there are ony three attributes regarding components of the items. However, this is a function of the extreme simpification of the prototype system. When deaing with reaistic products the number of component reated attributes can be expected to grow very rapidy. When the number of attributes becomes arge, the scaabiity of the system may be compromised as the knowedge discovery agorithms do not aways scae up in a satisfactory manner. Thus, it becomes imperative to use attribute seection to determine which attributes are essentia to inducing a high quaity recommender mode. Such attribute seection is commony used as a preiminary step preceding a earning agorithm such as decision tree induction and has numerous benefits. The resuting tree may be simper, which often makes it easier to interpret and thus more usefu in practice. It is aso often the case that simpe trees generaize better when they are used

22 22 for prediction. Thus, a smaer tree is ikey to score higher on interestingness measures and may score higher in accuracy. Finay, discovering which attributes shoud be kept often provides vauabe structura information and is therefore important in its own right. The iterature on attribute seection is extensive, and some of the methods appied for this probem in the past incude genetic agorithms (Yang and Honavar, 1998), various sequentia search agorithms (see e.g., Aha and Bankert, 1996; Caruna and Freitag, 1994), correation-based agorithms (Ha, 2000), evoutionary search (Kim, Street, and Menczer, 2000), rough sets theory (Modrzejewski, 1993), randomized search (Skaak, 1994), branch-and-bound (Naranda and Fukunaga, 1977), and the nested partitions method (Oafsson and Yang, 2001). These and other attribute seection methods are typicay cassified as either fitering or wrapper methods. Fitering methods produce a ranking of a attributes before the earning agorithm is appied. Wrapper methods use the earning agorithm to evauate subsets of attributes. As a genera rue, fitering methods are faster whereas wrapper methods usuay produce subsets that resut in more accurate modes. Using the NP-Wrapper attribute seection method deveoped by Oafsson and Yang (2001), the foowing eeven attributes were seected for the entire decision tree (with C4.5 used to evauate the attribute subsets): 1) Item 5) Round 11) QuantityForLastSheAuction 13) QuantityForLastPasticAuction 18) TimeOfParticipantsLastCpuAuction 21) CategoryOfParticipantsLastCpuAuction 24) QuantityForParticipantsLastSheAuction 25) CategoryOfParticipantsLastSheAuction 27) PriceForParticipantsLastPasticAuction 30) ParticipantName 31) ParticipantType

23 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 23 Computer CPU Item She Quantity for M s Last She Auction > 1 Time Since M s Last CPU Auction 1 Pastic Quantity for M s Last Pastic Auction > > 0 Price for M s Last Pastic Auction 2.9 > 2.9 Quantity for M s Last She Auction > 1 1 Figure 8. Decision Tree for Manufacturer after Attribute Seection te that attributes seected are from a variety of categories and the tota number of attributes is reduced to about a third of the origina size. With these attributes a new decision tree can be constructed. The branch corresponding to recommending auctions for a Manufacturer is shown in Figure 8. This tree is ceary much simper than the one constructed before (6 spit nodes versus 15 nodes before), yet its accuracy is about the same or 84.5% versus 84.7% estimated using 10-fod cross vaidation. The induction time for this mode is aso much shorter, or 1.3 seconds versus 4.3 seconds using a attributes, and when the mode is appied for cassification in rea time its simper structure wi resut in a faster response. The trade-off with respect to time is that considerabe computationa time is expended to seect the attributes. However, this can be done off-ine and therefore need not affect the scaabiity of the system adversey. 5. CONCLUSION In order to buid and sustain participation in an onine auction site whie aowing the system to scae robusty to a arge number of users, site administrators might consider providing users with decision support, such as

24 24 information about past prices and recommendations of current and upcoming auctions. Scaabiity can aso be enhanced by imiting the number of bidding opportunities for any given item. The resuts in this paper, based on imited experimentation with a simpe prototype auction system, suggest that some apparenty innocuous types of decision support can have unintended consequences for user behavior and profits, prices of items traded onine, and the quantities of materia traded. If the goa is to unify a fragmented market or generay to encourage exchange of items such as end-of-ife eectronics products, the methods used to achieve that goa may be sefdefeating. To foow up this initia study, much further research needs to be done to discover more about the effects of price information dispays, auction mechanisms (other than seaed bid doube auction), and extensiveness of auction opportunities. From the experimenta data, we were abe to iustrate the induction of decision trees for an inteigent recommender and to trim the set of reevant auction attributes it uses. The smaer attribute set provides greater scaabiity. We did not attempt to study the impacts of the recommender system on the outcomes for participants, prices or the market as a whoe. Our findings on price information suggest that it is dangerousy easy to ead the market in directions that benefit neither it nor its participants. Much further research, both anaytica and experimenta, is necessary to hep auction site administrators design effective decision support systems that achieve the goas of their sites. Acknowedgment: This work was supported by the Nationa Science Foundation under grant DMI REFERENCES Aha, D. W., and Bankert, R. L. (1996). A comparative evauation of sequentia feature seection agorithms. In D. Fisher & J.-H. Lenz (Eds.), Artificia Inteigence and Statistics V. New York: Springer-Verag. Basu, C., Hirch, H., and Cohen, W. (1998). Recommendation as cassification: using socia and content based information for recommendation. In Proceedings of the Nationa Conference on Artificia Inteigence. Breese, J., Heckerman, D., and Kadie, C. (1998). Empirica anaysis of predictive agorithms for coaborative fitering. In Proceedings of the 14 th Conference on Uncertainty in Artificia Inteigence. Breiman, L., Friedman, J., Oshen, R. and Stone, C. (1984). Cassification and Regression Trees. Wadsworth Internationa Group, Monterey, CA.

25 1. Experimenta Study of Scaabiity Enhancement for Reverse Logistics e-commerce 25 Caruana, R., and Freitag, D. (1994). Greedy attribute seection. In Proceedings of the Eeventh Internationa Conference on Machine Learning, pp New Brunswick, NJ: Morgan Kaufmann. Covisint.com (2001). Acceerating the pace of business. (October 29, 2001). Good, N., Schafer, J.B., Konstan, J.A., Borchers, A., and Sarwar, B. (1999). Combining coaborative fitering with persona agents for better recommendations. In Proceedings of the Nationa Conference on Artificia Inteigence. Ha, M. A. (2000). Correation-based feature seection for discrete and numeric cass machine earning. Proceedings of the Seventeenth Internationa Conference on Machine Learning, Stanford University, CA. Morgan Kaufmann Pubishers. Kim, Y. S., Street, W. N, and Menczer, F. (2000). Feature seection in unsupervised earning via evoutionary search. In Proceedings of the 6th ACM SIGKDD Internationa Conference on Knowedge Discovery and Data Mining. Kokkinaki, A. I., Dekker, R., de Koster, M. B. M., Pappis, C. and Verbeke, W. (2001a). From e-trash to e-treasure: how vaue can be created by the new e-business modes for reverse ogistics. In EURO 2001, the Operationa Research Conference, Rotterdam. Kokkinaki, A. I., Dekker, R., Lee, R., Pappis, C. (2001b). An eectronic marketpace for PCs in reverse ogistics networks. In 2nd European Conference E-COMM-LINE, Bucharest. Kumar, M. and Fedman, S. (1998). Internet auctions. In Proceedings of the Third USENIX Workshop on Eectronic Commerce, Boston, MA. Lin, W., Avarez, S.A. and Ruiz, C.. (2000). Coaborative recommendation via adaptive association rue mining. In Proceedings of ACM WEBKDD Liu, H. and Motoda, H. (1998). Feature Seection for Knowedge Discovery and Data Mining. Kuwer, Boston. McAfee, R. and McMian, J. (1987). Auction and bidding. Journa of Economic Literature, 25, Modrzejewski, M. (1993). Feature seection using rough sets theory. In P.B. Brazdi, editor, Proceedings of the European Conference on Machine Learning, pp Narendra, P. M., and Fukunaga, K. (1977). A branch and bound agorithm for feature subset seection. IEEE Transactions on Computers, 26(9), Oafsson, S. and Yang, J. (2001). Inteigent partitioning for feature reevance anaysis. Working Paper, Industria Engineering Department, Iowa State University, Ames, IA. Quinan, J. R. (1993). C4.5: Programs for Machine Learning. Morgan-Kaufmann, San Mateo, CA. Resnick, P., Iacovou, N., Suckak, M., Bergstrom, P., and Ried, J. (1994). Groupens: an open architecture for coaborative fitering of netnews. In Proceedings of ACM CSCW 94 Conference on Computer-Supported Cooperative Work, Ryan, S., Min, J. and Oafsson, S. (2002). Experimenta study of reverse ogistics e- commerce. In Proceedings of IEEE Internationa Symposium on Eectronics and the Environment, San Francisco, Shardanan, U. and Maes, P. (1995). Socia information fitering: agorithms for automating word of mouth. In Proceedings of ACM CHI 95 Conference on Human Factors in Computing Systems, Skaak, D. (1994). Prototype and feature seection by samping and random mutation hi cimbing agorithms. Proceedings of the Eeventh Internationa Machine Learning Conference, pp , New Brunswick, NJ: Morgan Kauffmann.

26 26 Teich, J., Waenius, H. and Waenius, J. (1999). Mutipe-issue auction and market agorithms for the Word Wide Web. Decision Support Systems, 26, Vucetic, S. and Obradovic, Z.. (2000). A regression-based approach for scaing-up personaized recommender systems in e-commerce. In Proceedings of ACM WEBKDD Wurman, P., Wash, W. and Weman, M. (1998). Fexibe doube auctions for eectronic commerce: theory and impementation. Decision Support Systems, 24, Yang, J., and Honavar, V. (1998). Feature subset seection using a genetic agorithm. In H. Motada and H. Liu, editors, Feature Seection, Construction, and Subset Seection: A Data Mining Perspective, Kuwer, New York.