Nigute washyira mubikorwa amanota ahuza: Intambwe ku yindi

Mperutse gukora ubushakashatsi bushingiye ku guhuza amanota ya Propensity kandi mugihe nkora ubushakashatsi ku makuru nahuye no kubura ibikoresho kuriyi ngingo. Inyinshi mu ngingo nasanze ziri ku mikorere yuburyo kandi ntabwo zirambuye neza mubijyanye na théorie. Kubwibyo, Nahisemo gusangira nawe umurongo ngenderwaho wuzuye kubijyanye no guhuza amanota hamwe nintambwe zayo

Guhuza amanota ya Propensity ni iki kandi kuki uyashyira mubikorwa?

“Amanota yerekana neza akubiyemo gukora ibice bihuye nibivurwa kandi bitavuwe bisangiye agaciro kamwe k'amanota. Icyitegererezo gihuye kimaze gukorwa, ingaruka zo kuvura zishobora kugereranywa ugereranije n'ibizagerwaho. ”

Igisobanuro cyatanzwe bwa mbere na Rosenbaum PR, Rubin DB mu kiganiro "Gusuzuma ibyiyumvo kuri covariate idakurikiranwa mu bushakashatsi bwakozwe hamwe n'ibisubizo byombi" byo mu 1983.

Kubivuga mu buryo bworoshye, ubu ni tekinike yinyongera ya A / B ikoreshwa mugihe sample randomisation idakora . Amanota yerekana (amahirwe yo guhabwa itsinda ryibizamini) yitsinda rivura abarwa kuri buri mukoresha hanyuma uyikoresha agahuzwa nundi mukoresha ashingiye kumateka yamateka yo gukoresha ibicuruzwa akora itsinda rishinzwe kugenzura. Nyuma, ibisubizo byamatsinda abiri bigereranywa ukoresheje ikizamini cyibarurishamibare kandi hapimwa ingaruka zigerageza.

Ariko kuki ukoresha tekinike igoye yo gushakisha itsinda rishinzwe kugenzura niba A / B urubuga rushobora kubikora aho? Rimwe na rimwe ntibishoboka gukoresha urubuga A / B rufite ibikorwa byubatswe. Dore imanza zishoboka:

1. Birahenze kandi bitwara igihe kugirango ubucuruzi bushyire mu bikorwa ikizamini cya A / B kubera imirimo itandukanye ikora kandi itumanaho (yego, birashoboka).
2. Isosiyete irashobora kubura inyungu mugihe igerageza amafaranga cyangwa ibikorwa bijyanye no kwamamaza / ingamba mugihe cyo gushyira mubikorwa igeragezwa rya kera.
3. Ntabwo akenshi bidashoboka kugira icyitegererezo cyateganijwe mubushakashatsi bwa interineti busabwa kugirango ikizamini cya A / B.
4. Ntabwo kandi ari byiza gukoresha ikizamini cya A / B mubushakashatsi bwo kumurongo mugihe ikintu cyageragejwe kidatanzwe kubushake.

Nagize ikibazo cya kane mubikorwa byanjye kandi byabaye mugihe nkorana nibicuruzwa bya e-bucuruzi. Itsinda ryibicuruzwa ryiteguraga kugerageza imikorere yo gutanga ibihembo kubakoresha nyuma yo gutumiza bwa mbere. Ikibazo nuko imikorere yakoraga itari kubakoresha bose bashyira gahunda yambere. Ibisabwa bimwe, nkigiciro cyurutonde, nibindi, byagombaga kubahirizwa. Muri iki kibazo, birenze imipaka yikizamini cya A / B kugirango ugabanye urujya n'uruza hagati yikizamini nitsinda rishinzwe kugenzura. Dore impanvu Guhuza amanota ya Propensity byari amahitamo.

Urwego rwo guhuza amanota

Urwego c rwuzuye rushingiye cyane ku ngingo “ Amanota yerekana amanota ahuye na R: uburyo busanzwe nibintu bishya ” kandi bigizwe n'intambwe eshanu (Ishusho 2).

Intambwe yambere ni ugukusanya amakuru agereranijwe amanota yagereranijwe hamwe nu mukoresha uhuye.

Intambwe ya kabiri ni ukugereranya amanota yerekana ukoresheje uburyo, nko gusubira inyuma, no guhugura kuri dataset kugirango umenye niba umukoresha azahabwa itsinda ryikizamini. Kuri buri mukoresha, icyitegererezo cyahuguwe gitanga amahirwe yo kuba mumatsinda yikizamini.

Intambwe ya gatatu ivuga guhuza hashingiwe ku manota yerekana, aho uburyo butandukanye bwo guhuza bwageragejwe, nkumuturanyi wegereye.

Intambwe ya kane, impuzandengo ya covariates hagati yo kuvura no kugenzura amatsinda igenzurwa no kubara imibare iringaniye no kubyara ibibanza. Impirimbanyi mbi yerekana ko icyitegererezo kigereranya amanota akeneye kubahwa.

Mu ntambwe ya gatanu yanyuma, ingaruka zigeragezwa zigereranijwe ukoresheje amakuru ahuye kandi ikizamini cyibarurishamibare kirakorwa.

Ikusanyamakuru

Iki cyiciro kijyanye no gukusanya impinduka zisabwa, covariates hamwe nuruvange. Covariate (X) nimpinduka yigenga ishobora guhindura ibisubizo byubushakashatsi (Y), ariko bidafite inyungu zitaziguye. Urujijo ni ikindi kintu kitari icyigwa gifitanye isano no kugabana itsinda ryikizamini (W) hamwe nibisubizo byubushakashatsi (Y).

Igishushanyo gikurikira kirerekana isano ihinduka. X ni covariate, W ni ikimenyetso cyerekana inshingano zo kuvura, na Y ni ibisubizo. Igishushanyo kiri ibumoso cyerekana umubano wuwitiranya kandi iburyo bwerekana isano yigenga ya covariate kubisubizo byubushakashatsi (Y) no kugerageza kugabana amatsinda (W).

Urashobora kubaza impinduka zingana iki nkeneye guhitamo? Igisubizo kiroroshye - byinshi, nibyiza kugirango ubone igereranyo kinini cyibisubizo no kugabanya kubogama kubogamye . Kandi hano ndavuga imibare minini nka 20-50 cyangwa irenga.

Kugereranya amanota

Kujya munzira ikurikira, birasabwa gukusanya amakuru no gushyiraho ibendera ryitsinda ryitsinda. Abandi bakoresha bose barashobora gushiraho itsinda rishinzwe kugenzura. Nyuma, amanota yerekana ko agereranijwe ukoresheje uburyo butandukanye, nko gusubira inyuma cyangwa amashyamba adasanzwe.

Inyinshi mu ngingo nasomye zerekana gukomera ku gusubira inyuma no kudakoresha izindi ngero zigoye kuko ubunyangamugayo buhanitse ntabwo ari umusaraba l. Nyamara, gutsindira guhuza tekinike byibanda kubwukuri.

Nyuma yo guhitamo uburyo, icyitegererezo cyo guhanura cyahuguwe kumibare ukoresheje covariates zatoranijwe kugirango hamenyekane niba umukoresha ari mubitsinda. Ubwanyuma, icyitegererezo gitanga ubuhanuzi kuri buri mukoresha, kandi amanota yerekana, amahirwe yo kuba mumatsinda yikizamini, arabaze. Kubijyanye na software ya software, muri Python urashobora gukoresha isomero iryo ari ryo ryose riteganya guhera kuri scikit shingiro-wige no kwimukira kuri leprophete.

Guhuza amakuru

Igikorwa gikurikira nugushira mubikorwa tekinike yo guhuza kugirango ushakishe umukoresha uhuje umukoresha kuva mumatsinda yikizamini. Kubwibyo, hashyizweho itsinda rishinzwe kugenzura.

Hariho uburyo butandukanye bwo guhuza guhitamo, kurugero rwose guhuza cyangwa Mahalanobis intera ihuye. Muri iki kiganiro ngiye kuganira cyane cyane kubuhanga busanzwe bwumuturanyi wegereye guhuza hamwe nuburyo butandukanye.

Umuturanyi wegereye guhuza (NNM) agizwe nibice bibiri. Ubwa mbere, algorithm itoranya abakoresha, umwe umwe mumatsinda yo kuvura, muburyo bwihariye. Ibikurikira, kuri buri mukoresha witsinda ryikizamini, algorithm isanga umukoresha mumatsinda yo kugenzura hamwe n amanota yegeranye. Izi ntambwe zisubirwamo kugeza nta mukoresha usigaye mu kizamini cyangwa kugenzura amatsinda. Muri Python, hari amasomero yihariye ya PSM nka PyTorch, Psmpy , causallib . Cyangwa burigihe ushobora kwizirika kumasomero yose ya kera hamwe na algorithms.

Hariho kandi uburyo bwo guhitamo moderi ya NNM hamwe na caliper. Caliper ishyiraho imipaka yo hejuru yintera yerekana amanota akomatanye. Rero, buri mukoresha arashobora guhuzwa gusa nabakoresha amanota yerekana ibintu mugihe gito. Niba abakoresha bujuje ibisabwa badashobora guhuzwa, uyikoresha azajugunywa.

Kuki nakoresha caliper? Nibyiza kubishyira mubikorwa mugihe intera yamanota ya tekinike ihuye ishobora kuba nini. Mugihe uhisemo ubunini bwa caliper, tekereza kuri ibi bikurikira: niba imikorere ihuza idashimishije, guhuza birashobora gukorwa hamwe na caliper ikarishye kandi niba guhuza bigenda neza ariko umubare wibihuje ni muto, Caliper irashobora kwagurwa ( https: / /www.ncbi.nlm.nih.gov/pmc/articles/PMC8246231/ ).

Gushyira mu gaciro

Muri iki cyiciro harasuzumwa niba covariates yikizamini hamwe nitsinda rishinzwe kugenzura iringaniye, bityo, ivuga niba umukino ari ukuri.

Hariho uburyo butatu bwo gupima uburimbane:

- imibare isobanura: ibipimo bisobanura itandukaniro (SMD) cyangwa igipimo gitandukanye (VR)

- ibizamini by'imibare

- visualisation: qq-umugambi, histogramu cyangwa umugambi wurukundo

Mu kiganiro ndibanda cyane kumahitamo ya mbere n'iya gatatu.

Ubwa mbere, reka tuganire kubisanzwe bisobanura itandukaniro no kugereranya. Ni izihe ndangagaciro zerekana ko covariate iringaniye? Ndasaba ko agaciro ka SMD kari munsi ya 0.1 Kubijyanye na VR, indangagaciro zigera kuri 1.0 zerekana impirimbanyi .

Umwanya wa kabiri, kubyerekeranye nuburyo bwo kubona amashusho, imwe mumibare yavuzwe haruguru ibarwa kuri buri covariate kandi yerekanwe mubishushanyo. Njye kubwanjye nkunda umugambi wurukundo kuko covariates zose zishobora gushyirwa mubishushanyo kimwe na covariates mbere na nyuma yo guhuza bishobora kugereranywa byoroshye. Ntanze urugero rwishusho hepfo.

Byagenda bite se niba covariates ikomeje kutaringaniza nyuma yo guhuza? Kugereranya, itandukaniro risobanura itandukaniro (SMD) ya covariates inshuro yo kugura na AOV ni 0.5, biri hejuru bisabwa 0.1. Bisobanura ko covariates itaringanijwe kandi hakenewe gusubiramo.

Hariho uburyo bune bwo gusubiramo guhuza:

1. Ongeramo covariates nshya

2. Hindura gusa uburyo bwo guhuza kuko haribenshi

3. Huza amanota yo guhuza hamwe nuburyo bwiza bwo guhuza

4. Ongera ingano yicyitegererezo

Kugereranya Ingaruka zo Kuvura

Hanyuma, twegereje icyiciro cyanyuma mugihe igeragezwa ryagereranijwe. Hariho ubwoko butatu bwo kugereranya ingaruka: impuzandengo yo kuvura (ATE), impuzandengo yo kuvura ku bavuwe (ATT), hamwe ningaruka yo kuvura igenzura (ATC). Muri rusange, ATE ni itandukaniro ryabazwe mubipimo byingenzi hagati yikizamini no kugenzura amatsinda (bisa no gupima ibipimo nyamukuru mubizamini bya A / B). Irabarwa nkuburyo bwo kuvura, ATE = avg (Y1 - Y1) nkuko bigaragara hano mumashusho.

Mugihe ATT na ATC ari impuzandengo yo kuvura yikizamini no kugenzura itsinda. Byose ni uburyo bworoshye kandi bwumvikana bwo kugereranya.

ATE nubwoko busanzwe kandi bukoreshwa mugihe kugenzura no gupima amatsinda yingenzi ibipimo bigereranijwe kandi ingaruka zapimwe zirapimwa. Mugihe ATT na ATC bikunzwe mugihe ibipimo byuzuye bisabwa kuri buri tsinda. Kurangiza, ikizamini cyibarurishamibare gikwiye gukorwa kugirango harebwe akamaro k'ibisubizo by'ibisubizo.

Imipaka yo guhuza amanota

Nyuma yubusobanuro burambuye bwuburyo bwo guhuza amanota , birashobora kuba igihe cyo gutangira kubishyira mubikorwa mubikorwa byawe, ariko hariho imbogamizi zigomba gutekerezwa.

1. Bootstrap ntabwo isabwa gukoreshwa hamwe no guhuza amanota ya Propensity kuko byongera itandukaniro. ( https://ubukungu.mit .

 2. Stable unit treatment value assumption (SUTVA) principle must be met. 3. Propensity Score Matching implies using two machine learning algorithms (one for propensity score calculations and the second one for matching), which can be a pricy method to use for a company. On that account, it's advisable to negotiate with your team on A/B test conduction. 4. Finally, as discussed above, a big number of covariates are suggested to be used in the models. Thus, it requires a high-powered machine(-s) to calculate the results of the models. Again, it's a costly method to implement.

Ariko, niba bishoboka gushyira mubikorwa guhuza amanota , kora kandi ntutindiganye kuzamura uburambe bwawe nubumenyi bufatika. Amahirwe hamwe nubushakashatsi bwawe buzaza hamwe no kuvumbura imashini