VI. Clcksream Bg Daa and Delvery before Order Makng Mode for Onlne Realers Yemng (Yale) Gong EMLYON Busness School Haoxuan Xu *, Jnlong Zhang School of Managemen, Huazhong Unversy of Scence &Technology Absrac Our research s nspred by a leadng onlne realer usng clcksream bg daa o esmae cusomer demand and hen shp ems o cusomers or hubs near cusomers by a mode of delvery before order makng (DBOM) mode. Usng clcksream daa o oban advance demand nformaon n order quanes, we negrae he forecasng wh a sngleem uncapacaed dynamc lo szng problem n a rollng-horzon envronmen. Usng he smulaed clcksream daa, we evaluae he performance of DBOM mode. 1 Inroducon and Leraure A leadng onlne realer, wh 10 bllon USD urnovers n Chna, uses bg daa of onlne clcks and daa mnng algorhms o esmae he expeced order quany n dfferen locaons ncludng collecon locaons, locker locaons and hubs, hen shps ems o he locaons by a mode of delvery before order makng operaonal mode. Amazon, anoher leadng onlne realer n USA, has laely announce a new dsrbuon mehod AS ( ancpaory shppng, see [1]), specfyng a mehod o sar shppng packages before cusomers really buy producs. Amazon AS mehod can predc he cusomer demand o oban he geographcal desnaon area nformaon by analyzng dfferen varables, ncludng hsorcal orderng behavor, wsh-lss, clckng daa. The packages are n rans or wang a a hub unl an order arrves, and hen shpped o he specfc locaon quckly. Inspred by hese new logscs modes usng bg daa, hs paper addresses an operaonal problem concernng he use of a knd of bg daa clcksream daa n a 1
specfc onlne realng envronmen. Lee e.al [2] defne he clcksream daa of onlne sores o be he pahs nformaon from vsors. Many researchers have suded he markeng benefs of usng clcksream daa, or raher clcksream rackng, n e- commerce sengs. For a dealed revew, see [3] and references heren. Dfferen from hs sream of research, we nvesgae he benefs of clcksream daa from an operaonal perspecve. Huang and Van Meghem [4] specfy he cos-savng effecs on nvenory managemen for a specfc company by usng clcksream rackng daa on s non-ransaconal webse. To our knowledge, exsng leraure has no suded he operaonal benefs of clcksream daa n an envronmen of onlne realng. A naural queson arses wheher or no onlne realers can use such clcksream daa o mgae he demand uncerany and mprove he nvenory managemen process. Agaz e al [5] ndcae ha delvery and afer-sales servce are becomng key compeve facors n oday s e-commerce. Hence, he mach of supply and cusomers demands s essenal for onlne realers o assure fas delvery and good servce. Noneheless, a rade-off exss beween he hghly guaraneed sock and demands uncerany. Unlke demand forecass n radonal offlne ransacon seng, whch s usually based on hsorcal daa, onlne realers can beer predc he fuure demands by furher usng clckng daa before cusomers placng orders (see [6]). In an onlne realng envronmen of sellng pershable or cusomzed producs (e.g., produce, assembled compuers and jewels), realers may expec a mely supply or fas producon, and face me-varyng demands. Based on he hsorcal radng daa, nvenory managers can forecas fuure demands and rea hem as deermnsc daa (e.g., he mean value), hen model he nvenory replenshmen processes as dynamc lo szng (DLS) problems. Based on hs applcaon seng, hs paper specfes he use of clcksream daa n an negraed dynamc nvenory conrol polcy. We frs develop an adapve forecasng mehod by he use of clcksream daa o beer predc he fuure demand paern. Then, we embed such advance demand nformaon no a DLS model n a rollng-horzon envronmen. Gven ha he clcksream daa evolves dynamcally, we updae he demand nformaon accordngly. Gallego and Özer [7] classfy advance demand nformaon (ADI) no observed and unobserved pars. The observed par of ADI s easy o oban for onlne realers when cusomers place orders onlne, snce hey are usually sasfed several perods laer. Ths s also he case n some radonal realng and producon sengs when he requremens of some producs or componens are released n advance. As for he unobserved par of ADI, of whch radonal realers has no nformaon, onlne realers can use clcksream daa o ge access. Alhough researchers exensvely nvesgae he value of ADI n operaonal managemen, few of hem explore o oban he ADI for onlne realers by usng he clcksream daa. Usng he smulaed clcksream daa accordng o real onlne realng envronmen, we examne he cos savng effec and fas delvery effec of our nvenory model. 2
2 Formulaon 2.1 Problem Descrpon We consder an onlne realer mananng s own sock for a ceran commody. The manager needs o develop a good nvenory conrol polcy o mnmze he relaed producon/purchasng cos and nvenory cos. Before applyng a ceran nvenory model, s necessary o denfy he demand paern. As usual, one can make use of he hsorcal demand daa o predc he fuure demands. However, s no ha accurae snce many unseen facors exs. A ypcal feaure of onlne purchasng s ha cusomers generae a large amoun of clcksream daa whch could be racked by onlne realers. Our problem s o explore he use of such bg daa n predcng a more accurae fuure demand paern. Ths ask can be handled by a suable algorhm n machne learnng heory. Specfcally speakng, we apply an on-lne algorhm no he forecasng process. Tha s, afer he learnng model predcs, he rue resul wll be revealed and ac as feedback o updae he algorhm accordngly. The algorhm hen adapvely makes a proper predcon. Whou any assumpon on he demand dsrbuon, we use he machne learnng algorhm descrbed above o mne he clcksream daa for predcng fuure demand. Then we ncorporae no a dynamc lo-szng model o solve he replenshmen problems n a rollng-horzon envronmen for hs onlne realer. The dynamc lo-szng models are wdely used by onlne realers snce hey wdely use ERP sysems conanng a MRP modular o make replenshmen plans (See [9]). 2.2 A Clcksream-based Adjused Rollng DLS We develop an operaonal decson framework o mprove he nvenory conrol polcy for onlne realers. We frs develop an adapve demand forecasng mehod, whch ncludes wo mnor seps. A he frs sep, we ncorporae he hsorcal demand daa of a ceran commody no a radonal forecasng algorhm o generae an nal demand paern. A he second sep, we apply an on-lne machne learnng algorhm, wnnow algorhm (see [8]), o adapvely forecas he demand of he neares fuure perod by usng he laes clcksream daa. Thereby, he predced demand paern s updaed. Ths process s dynamcally evolved as me goes on. Afer he predced demand daa s obaned, we ncorporae no a lo-szng model n a rollng horzon envronmen o dynamcally make he replenshmen plans wh an objecve of mnmzng he oal nvenory relaed cos. The overall decson framework s shown n fgure 1. 3
Fgure 1: Clcksream-based Adjused Rollng DLS 2.2.1 Usng Clcksream Daa n Demand Forecasng Managers esmae he demand paern before makng any specfc replenshmen decsons. Tradonal forecasng reles heavly on a demand hsory. Dfferen from he mode n convenonal realng, onlne realers can no only observe a demand hsory, bu also oban a clcksream hsory. Usng only he hsorcal demand daa o esmae he fuure demand may lead o devaons, snce a lo of flucuaons exs. As a resul, we desgn a dynamc procedure o adapvely forecas a more accurae demand. Based on he hsorcal demand daa, we frs apply a rend-adjused exponenal smoohng (TAES) algorhm o generae an nal predced demand. Then we use he clcksream daa wh an on-lne machne learnng algorhm o dynamcally updae he forecasng. 2.2.1.1 A rend-adjused exponenal smoohng algorhm Demand hsory provdes valuable nformaon for onlne realers o predc he fuure demand. In hs paper, we adop a TAES algorhm o generae he nal predced demand paern. The algorhm uses wo parameers, and, as coeffcens for he average demand and s rend, respecvely. The followng equaons are he forecasng 4
algorhm: A D (1 )( A 1 T 1) (1) T ( A A 1) (1 ) T 1 (2) F A T (3) 1 D : Demand n perod ; A : Exponenally smoohed average of he seres n ; T : Exponenally smoohed average of he rend n perod ; F 1 : Predced demand n perod ; : Smoohng parameer for he average (0 1) ; : Smoohng parameer for he rend (0 1) ; Usng hs forecasng mehod, we predc he demands of he followng plannng horzon from o N. We denoe hem by a demand vecor F ( F 1 F2 FN). In pracce, many convenonal realers, even onlne realers, jus fnsh he forecasng process here, whle hs s jus he nal predced demand n our forecasng mehod. In he followng, we use clcksream daa o updae and mprove he forecas. 2.2.1.2 A wnnow algorhm Wnnow s a ypcal on-lne machne learnng algorhm, whch s frsly developed by Llesone [8]. Based on he varables of he clckng examples, wnnow keeps learnng he weghs of each varable and makes a bnary predcon of wheher a vs/clck leads o a purchase. In an on-lne seng, once he algorhm makes a predcon, he real value s revealed hen and gves feedback o he algorhm. A smple verson of wnnow algorhm s as follows: Sep 1. Inalze each wegh of varable x o 1; Sep 2. Gven a clckng example x { x1 x2 x n }: n 1 1 x n oupu n 0 0 1 x oupu Sep 3. The weghs of he varables are updaed when he algorhm makes a msake: a). If he algorhm predcs 1 and he rue value s 0, hen p0 p 1; b). If he algorhm predcs 0 and he rue value s 1, hen qq 1; Sep 4. Go o 2. 5
Usng he hsorcal clcksream daa as he ranng se, we can oban an updaed vecor of he wegh of each varable. Applyng he wegh vecor o he laes clcksream daa as he es se, he wnnow algorhm can make a good predcon of hose clcks n perod leadng o purchasng n perod +1. Thereby, we can use hs nformaon o updae he demand of he neares fuure perod,.e. F. As me goes on, he predced 1 demand vecor F can be dynamcally updaed by combnng hese wo algorhms. 2.2.2 A rollng-horzon lo-szng model Replenshmen or producon plannng problems n onlne realng are usually solved n a dynamc, rollng-horzon paern. A frs, say n perod, a decson problem s solved o opmaly n a plannng horzon of gven lengh T. The manager hen wll mplemen he frs-perod decson for k perods n he resulng soluon. Aferwards, he sysem evolves o perod +k. Obanng he updaed demand nformaon, he manager has o make he nex decson. Ths process s repeaed under such rollng framework (see [11]). A he second sep of our operaonal decson framework, gven he updaed demand nformaon of a new forecas horzon obaned a sep one, we apply a sngleem uncapacaed DLS model o formulae he nvenory replenshmen problem. In a rollng-horzon envronmen, alhough we dynamcally oban a new demand vecor F for he nex forecasng horzon, we can regard any revew perod as he begnnng of a new forecas horzon when makng decsons. Usng he resul a sep one, we ge he predced neger demands of T perods,.e. F { F 1 F2 FT}. A any perod when we need o make decsons, we rese =1, and have he followng lo-szng problem: T Mn ( ky pw hi ) (4) 1 I w I F( x x x n ) (5) I 1 w I F( ) 2 T (6) 0 x My 1 T (7) S.T. 0 1 1 1 1 2 wi 0 (8) y {0 1} (9) In he above model, T s he forecas horzon, k s he fxed orderng cos n perod, p and h denoe un purchasng cos and un holdng cos alernavely n perod. I s he nvenory a he end of perod, y s a bnary decson varable ndcang wheher o replenshmen n perod. w denoes how much o replenshmen n perod. M s a very large number. F1( x1 x2 x n ) s he demand of he frs perod, whch s decded by he wnnow algorhm usng he clcksream x 1 x 2 x n. F( ) s he demand beyond he 6
frs perod, whch s decded by he rend-adjused exponenal algorhm wh parameer and. 3 Analyss In hs secon, we analyze how o apply our clcksream-based adjused rollng DLS decson framework n an onlne realng envronmen hrough a smulaed example. Snce he TAES algorhm s a ypcal me-seres forecasng echnque and easy o be execued n EXCEL. We can drecly use o oban he nal predced demands based on hsorcal demands. The key funcon of wnnow algorhm s o dsjunc he mos mporan varables and o make a good predcon. The onlne purchasng behavour may be correlaed wh housands of npu facors. Usng a specfc varable selecng echnque, Van den Poel and Bucknx [10] denfy nne key varables ou of 92 possble measures n predcng wheher a vsor wll purchase durng her nex vs. Whle Van den Poel and Bucknx [10] focus on he vsor level, we focus on he produc level,.e., wheher a vs o a ceran produc wll lead o a purchase of hs produc. Based on [10], we use varables shown n Table 1 for he wnnow algorhm. Table 1: Varables of wnnow algorhm Varables Defnon Descrpon x he vsor s a regsered member or no 1 yes and 0 no 1 x he cusomer vsed durng las perod or no 1 yes and 0 no 2 x he cusomer vsed before las perod or no 1 yes and 0 no 3 x he vsor clcks he personal pages or no 1 yes and 0 no 4 x he vsor clcks only hs produc or no 1 yes and 0 no 5 x he vsor supples personal nformaon or no 1 yes and 0 no 6 x wheher he cusomer purchase hs produc before 1 yes and 0 no 7 x wheher he average me per clck s hgher han he 1 yes and 0 no 8 average We hen buld a basc wnnow classfer n MATLAB o judge wheher a clck leads o purchase. The classfer works by he followng seps: Sep 1: Inalze each wegh 1 ( 1 mm ; 8) ; Sep 2: Apply he ranng se o adjus he wegh : 7
If If m 1 x, and he clck does no lead o a purchase, hen reduce he wegh of hose x 0 o p (0 p 1), ll m 1 x m 1 x ;, and he clck does leads o a purchase, hen ncrease he wegh of hose x 0 o q ( q 1), ll m 1 x ; Sep 3: Apply he updaed wegh obaned n he ranng se o he es se, calculae 9 x and compare o he hreshold, hen predc f a clck wll lead o a purchase. 1 We use MATLAB o generae a 50-perod demand vecor based on a normal dsrbuon N (20, 5). Then we randomly generae 500 clcks for each perod, each clck wh a feaure vecor ( x1 x2 x8) and a purchase or no ndcaor (1 sands for purchase and 0 no). The sum of he ndcaors n each perod s equal o he rue demand of ha perod. In our example, he average rae of converson from clck o purchase s 3.58%, whch s reasonable n e-commerce seng accordng o [6]. We dvde he 50-perod daa no wo ses, he former 25 perods as he ranng se and he laer 25 perods as he es se. By seng p 0.9, q 2 and he hreshold 0.5, we frs use he wnnow classfer n he ranng se o oban an updaed wegh vecor, and hen use hs vecor o predc wheher a clck n he es se wll lead o a purchase. Fgure 2 shows a comparson beween he performance of he clcksream-based wnnow algorhm and he TAES algorhm ( 0.8, 0.7 ) n predcng he demands n he es se. We fnd ha he clcksream-based algorhm s much beer han he TAES algorhm. 8
Fgure 2: Comparson beween TAES and clcksream-based wnnow We hen solve a rollng-horzon dynamc lo szng problem for he 25-perod es se. We assume he fxed cos k 40, purchasng cos p 0 and holdng cos h 2 for all 1,...,25. We use he rollng schedule descrbed n [11] o solve our problem. The forecas horzon T s chosen o be 2, 3, 4, 5, 6, 7, 8, separaely. The demand of he frs perod n he forecas horzon s predced by he wnnow algorhm and he res demands are predced by TAES algorhm. Only he frs decson n he opmal soluon of he forecas horzon s mplemened, hen he process rolls o he nex decson perod o solve anoher DLS problem wh a plannng horzon of T. The schedule ends when reachng he 25h perod. Table 2 shows he percenage devaon of he cos from he opmaly, whch s obaned by solvng he enre 25-perod DLS problem wh he rue demand. We fnd ha he cos of usng he clcksream-based demand s closer o opmaly han usng demand obaned by TAES only. Table 2: Percenage devaon from opmaly Forecashorzon TAESdemand only Clcksream-based demand 2 3.5% 1.3% 3 5.3% 2.9% 4 5.7% 3.3% 5 5.7% 3.3% 6 5.7% 3.3% 9
7 5.7% 3.3% 8 5.7% 3.3% 4 Concludng Remarks Ths paper presens an negraed clcksream-based operaonal decson framework for onlne realers. I explores he use of an on-lne machne learnng algorhm, wnnow algorhm, o mne clcksream bg daa and mprove he demand managemen process, nally based on radonal forecasng mehod. Applyng he updaed demand nformaon n a rollng-horzon dynamc lo szng problem, we analyze s cos advanage over radonal forecasng mehod. Our curren clcksream mnng algorhm can only predc wheher a clck wll lead o purchase or no, bu canno predc he quany a purchase conans. I s neresng o explore oher algorhms o solve he problem. Acknowledgemens Ths research s suppored by Collaborave Innovaon Cener for Modern Logscs and Busness of Hube (Culvaon), Modern Informaon Managemen Research Cener (MIMRC) of HUST and NSFC (No.70901028; 71271095). References [1] Spegel, J., McKenna, M., Lakshman, G. and Nordsrom, P., Mehod and sysem for ancpaory package shppng, US Paen, 8, 615, 473 (2013). [2] Lee, J., Podlaseck, M., Schonberg, E. and Hoch, R., Vsualzaon and Analyss of Clcksream Daa of Onlne Sores for Undersandng Web Merchandsng, In Applcaons of Daa Mnng o Elecronc Commerce, 59 84. Sprnger, US (2001). [3] Hu, S. K., Fader, P. S. and Bradlow, E. T., Pah Daa n Markeng: An Inegrave Framework and Prospecus for Model Buldng, Markeng Scence, 28, 2, 320-335 (2009). [4] Huang, T. and Van Meghem, J. A., Clcksream Daa and Invenory Managemen: Model and Emprcal Analyss, Producon and Operaons Managemen, 23, 3, 333-347 (2014). 10
[5] Agaz, N. A., Fleschmann, M. and Van Nunen, J. A., E-fulfllmen and Mulchannel Dsrbuon - A Revew, European Journal of Operaonal Research, 187, 2, 339-356 (2008). [6] Moe, W. W. and Fader, P. S., Dynamc Converson Behavor a E-commerce Ses, Managemen Scence, 50, 3, 326-335 (2004). [7] Gallego, G. and Özer, Ö., Inegrang Replenshmen Decsons wh Advance Demand Informaon, Managemen Scence, 47, 10, 1344-1360 (2001). [8] Llesone, N., Learnng quckly when rrelevan arbues abound: A new lnearhreshold algorhm, Machne learnng, 2, 4, 285-318 (1988). [9] Gunasekaran, A., Marr, H. B., McGaughey, R. E. and Nebhwan, M. D., Ecommerce and s mpac on operaons managemen, Inernaonal Journal of Producon Economcs 75, 1, 185-197 (2002). [10] Van den Poel, D. and Bucknx, W., Predcng onlne-purchasng behavor, European Journal of Operaonal Research, 166, 2, 557-575 (2005). [11] Baker, Kenneh R., An expermenal sudy of he effecveness of rollng schedules n producon plannng, Decson Scences, 8, 1, 19-27 (1977). 11