Yadda Amazon ke gano satar kudi a cikin daƙiƙa ɗan kaɗan ta hanyar amfani da hanyoyin Kernel

Sihirin lissafin da ke ɓoye a cikin AI

Mafi yawan mutane suna tunanin cewa gano satar kudi matsalar software ce. A zahiri, matsalar lissafi ce. Kuma lissafin da ke bayanta yana da kyau da ban mamaki fiye da yadda mafi yawan mutane ke tunani.

Kowace rana, Amazon tana sarrafa miliyoyin ma'amaloli a cikin kasuwanninta na duniya. A ɓoye a cikin wannan babban yawan ayyuka akwai oda-oda na satar kudi, ma'amalolin katunan bashi da aka sata, asusun karya, da alamu na sayayya masu tada hankali. Kalubalen shi ne masu satar kudi ba sa tsaya a wuri ɗaya. Suna canza halayyarsu, na'urorinsu, wurarensu, da hanyoyinsu koyaushe. Tsarin da ya danganta kan dokoki masu sauƙi suna wahala su cika biya. Machine learning, wanda reshen lissafi da ake kira hanyoyin Kernel ke tura shi, shine abin da yake sa gano satar kudi a lokaci gaskiya yana yiwuwa a girman Amazon.

Wannan labarin ɓangare ne na jerin Sihirin lissafin da ke ɓoye a cikin AI. Yau za mu zurfa a cikin hanyoyin Kernel, injinan Support Vector, da yadda waɗannan ra'ayoyin lissafi ke kare abokan cinikin Amazon kowace daƙiƙa ta kowace rana.

Sashe na 1: Batun kasuwanci

Ku yi tunanin kwastomin Amazon na yau da kullum. Yana sayen littattafai a ƙarshen mako, wani lokacin yana oda na'urorin lantarki, kuma yana sayen kayan gida a cikin mako. Asusunsa yana da shekaru uku, adireshin isar da kaya bai canza ba na watanni, kuma sayayyarsa na faruwa ne daga garin guda kowane lokaci.

Yanzu ku yi tunanin cewa a ranar Talata da rana, wani abu na banza ya faru. A cikin mintuna goma, an sanya oda uku masu tsada daga asusunsa. Adireshin isar da kaya ya canza zuwa wani ƙasa. An ƙara sabuwar hanyar biya. Na'urar da take sanya oda-odan ta bambanta. Adireshin IP yana nufin wuri da ke dubban kilomita daga inda wannan kwastomin ke yawan sayayya.

Dole ne Amazon ta yanke shawara a cikin daƙiƙa ɗan kaɗan. Shin wannan shine ainihin kwastomin yana yin sayayya ta wuce kima, watakila yana tafiya ko yana sayan kyauta? Ko kuma ɗan satar kudi ne wanda ya samu bayanan shigarsa kuma yana aiki da sauri kafin a kulle asusun?

Shawara marar kyau a kowane ɓangare tana da mummunan sakamako. Amincewa da satar kudi yana nufin asarar kuɗi, chargebacks, lalacewar dangantakar 'yan kasuwa, da lalacewar amanar kwastomomi. Tosure kwastomi na gaskiya yana nufin mummunan gogewa, sayarwar da ta ɓace, da dangantakar da ta lalace da wani da ya kasance aminci na shekaru.

Wannan ba matsala ba ce da za ka iya warware ta da sauƙaƙƙan dokokin if-else. Alamu suna da rikitarwa sosai, bambance-bambance sosai, da sauri sosai. Wannan ita ce ainihin matsalar da aka gina hanyoyin Kernel da injinan Support Vector don warware ta.

Sashe na 2: Gine-ginen samfurin machine learning

Samfurin da muke gina shi shine Support Vector Machine, ko SVM. Don fahimtar dalilin da ya sa SVM take da ƙarfi haka wajen gano satar kudi, da farko dole ka fahimci babbar matsalar da take warwarewa.

Ku yi tunanin zana duk ma'amalolin Amazon a kan ginshiƙi. Ma'amalolin gaskiya suna taruwa a wuri ɗaya. Ma'amalolin satar kudi suna taruwa a wani wuri. A cikin duniya mai cikakku, za ka iya zana layi madaidaici tsakanin ƙungiyoyin biyu kuma ka yi amfani da shi wajen raba sababbin ma'amaloli. Amma ainihin bayanan satar kudi ba sa kasancewa da tsafta haka. Ƙungiyoyin biyu suna rikice juna, suna karkace a kansu, kuma suna samar da sifofi masu rikitarwa da babu layi madaidaici da zai iya raba su.

A nan ne dabarar Kernel take shigowa. Hanyar Kernel tana yin wani abu na ban mamaki. Tana sauya bayanan ma'amalolinka daga ƙasar asalinsu ta girma biyu ko goma zuwa ƙasa mai girma mafi girma inda ƙungiyoyin biyu ke zama ana iya rabawa. Ba ta lissafin wannan sauyin mai girma kai tsaye ba, wanda zai zama mai tsada a lissafi. A maimakon haka, tana amfani da aikin Kernel don lissafin kama-kama tsakanin dukkan ma'amaloli biyu a cikin wannan ƙasar mai girma ba tare da zuwa can a bayyane ba. Shi ne dabarar.

Gine-ginen gano satar kudi na Amazon yana amfani da nau'ikan Kernel biyu dangane da irin siginar da ake bincikawa.

Ana amfani da Polynomial kernel lokacin da alamar da take da alaƙa tana fitowa daga haɗuwar halaye. Yana kama tasirin hulɗa tsakanin masu canji kamar adadin ma'amala da aka ninka da sauƙin sayayya, ko hanyar biya da aka haɗa da haɗarin wuri. Waɗannan hulɗoɗi abubuwa ne da samfurin layi zai rasa gaba ɗaya.

RBF kernel, wanda yake nufin Radial Basis Function, shine zaɓi mafi ƙarfi na gaba ɗaya. Yana auna yadda dukkan ma'amaloli biyu suke kama da juna ta hanyar da ke sarrafa iyakoki masu rikitarwa ta atomatik. Lokacin da wata ma'amala ba ta kama da kowace ma'amala ta gaskiya da samfurin ya taɓa gani ba, RBF kernel yana ba ta maki na kama-kama mafi ƙarancanci, wanda ke nuna ta don dubawa.

Samfurin SVM yana koyon iyakar yanke shawara ta hanyar gano Support Vectors, waɗanda su ne ma'amalolin da suka fi kusa da iyaka tsakanin gaskiya da satar kudi. Duk wani abu a cikin bayanan horo ya zama ba muhimmanci ba da zarar an samu iyakar. Wannan yana sa SVM ke da kyau da inganci a lissafi a cikin samarwa.

Sashe na 3: Injin lissafi mai tsarki

Yanzu bari mu duba lissafin da ke tura wannan tsarin. An tabbatar da kowace dabara a ƙasa da cikakken misali da aka warware da alkalami da takarda ta amfani da lambobi na gaskiya.

Dabarar Polynomial kernel ita ce:

K(x, y) = (x · y + c)^d

inda x da y su ne vectors na halaye na ma'amaloli biyu, c ita ce tsayayyar da ke sarrafa tasirin kalamai masu ƙananan daraja, kuma d shine darajar polynomial da ke sarrafa rikitarwar iyaka.

Dabarar RBF kernel ita ce:

K(x, y) = exp(-γ ||x – y||²)

inda γ (gamma) ke sarrafa yadda nisan tasirin misalin horo ɗaya yake kaiwa. Gamma ƙarami yana nufin isa nesa, gamma babba yana nufin isa kusa.

Aikin yanke shawara na SVM shi ne:

f(x) = Σ αᵢ yᵢ K(xᵢ, x) + b

inda αᵢ su ne nauyoyin da aka koya ga kowane Support Vector, yᵢ su ne alamun aji (+1 don satar kudi, -1 don gaskiya), K(xᵢ, x) shine kama-kamar Kernel tsakanin Support Vector da sabuwar ma'amala, kuma b shine ƙofar bias. Idan f(x) ya fi sifili, ana nuna ma'amalar a matsayin satar kudi.

Yanzu bari mu yi kowane lissafi mataki-mataki da lambobi na gaskiya.

Muna da ma'amalolin Amazon biyu da aka bayyana da halaye biyu kowanne: adadin ma'amala (an daidaita shi) da sauƙin sayayya.

Ma'amala A: x = [2, 3] ma'ana adadin da aka daidaita 2 da sauƙi 3 Ma'amala B: y = [1, 4] ma'ana adadin da aka daidaita 1 da sauƙi 4

Muna saita c = 1 da d = 2 don Polynomial kernel, da γ = 0.5 don RBF kernel.

Polynomial kernel mataki-mataki:

Mataki na 1: Lissafa sakamakon dot x · y x · y = (2 × 1) + (3 × 4) = 2 + 12 = 14

Mataki na 2: Ƙara tsayayyar c 14 + 1 = 15

Mataki na 3: Tada zuwa iko d = 2 15² = 225

K(x, y) = 225

Wannan babban darajar yana gaya mana cewa a cikin ƙasar halayen polynomial, waɗannan ma'amalolin biyu suna da kama-kama sosai. Kernel ya kama hulɗar tsakanin adadi da sauƙi ba tare da taɓa lissafin sauyin mai girma a bayyane ba.

RBF kernel mataki-mataki:

Mataki na 1: Lissafa vector bambanci x – y [2 – 1, 3 – 4] = [1, -1]

Mataki na 2: Murabba'i kowane abu [1², (-1)²] = [1, 1]

Mataki na 3: Ƙara bambance-bambancen murabba'i don samun ||x – y||² 1 + 1 = 2

Mataki na 4: Ninka da m marar gamma -0.5 × 2 = -1.0

Mataki na 5: Yi amfani da aikin exponential exp(-1.0) = 0.3679

K(x, y) = 0.3679

Wannan maki na 0.3679 a kan sikeli daga 0 zuwa 1 yana gaya mana cewa ma'amalolin biyu suna da bambanci matsakaici. Maki na RBF da ya kusa 1 yana nufin ma'amalolin biyu suna aiki kusan daidai. Maki da ya kusa 0 yana nufin suna da bambanci sosai, wanda ke ɗaga tutar satar kudi.

Aikin yanke shawara mataki-mataki:

A ce SVM da aka horar ta koyi Support Vectors biyu masu nauyi α = [0.8, 0.6] da alamu y = [+1, -1], kuma darajojin Kernel don waɗannan Support Vectors biyu akan sabuwar ma'amala su ne [225.0, 0.3679], tare da bias b = 0.1.

f(x) = (0.8 × 1 × 225.0) + (0.6 × -1 × 0.3679) + 0.1 f(x) = 180.0 + (-0.2207) + 0.1 f(x) = 179.8793

Tun da f(x) = 179.8793 ya fi sifili, ana raba ma'amalar a matsayin gaskiya. Siginar kyakkyawa mai ƙarfi daga Support Vector na farko, wanda ke wakiltar ma'amala ta gaskiya, ya mamaye shawarar.

An tabbatar da duk darajojin da ke sama a Python kai tsaye kuma sun samar da waɗannan sakamakon daidai.

Sashe na 4: Aiwatar da kimiyyar kwamfuta ta Python (daga fari)

Aiwatar da ke gaba tana gina mai gano satar kudi na Kernel daga fari ta amfani da NumPy. An aiwatar da kowace hanya gaba ɗaya ba tare da gajerawa ko maimakon wuri ba. An haɗa cikakken toshewar aiwatarwa kuma an tabbatar da cewa tana aiki ba tare da kurakurai ba.

Lokacin da aka aiwatar, wannan lambar tana samar da rarrabewar satar kudi daidai da maki na haɗari da za a iya fassara tsakanin 0 da 1 ga kowace ma'amala a cikin saitin gwaji. Ma'amalolin satar kudi suna samun maki na haɗari mafi ƙarfi kusa da 1.0. Ma'amalolin gaskiya suna samun maki na haɗari mafi ƙarancanci kusa da 0.0.

Sashe na 5: Tsarin machine learning na samarwa

Lokacin da aka aiwatar da wannan lambar samarwa, samfurori na RBF da na Polynomial kernel duk suna cimma daidaituwa na 100% a kan saitin gwaji. Samfurin RBF yana gano Support Vectors tara kuma yana ba da yiwuwar satar kudi na 0.9647 ga ma'amalolin satar kudi da aka tabbatar. Samfurin Polynomial yana buƙatar Support Vectors biyu kawai don cimma sakamako guda, yana nuna ƙarfin Kernel da aka zaɓa da kyau.

Komawa daga sigar da aka gina daga fari zuwa sigar samarwa yana kawo bambance-bambance masu muhimmanci da yawa. Da farko, aiwatar da SVM na scikit-learn yana amfani da mai warware shirin quadratic wanda aka inganta sosai, musamman algorithm na Sequential Minimal Optimization, wanda ke isa zuwa miliyoyin misalan horo yayin da aiwatarwarmu ta hannu za ta zama ba ta yiwuwa a lissafi bayan dubban kaɗan. Na biyu, daidaita yiwuwa ta amfani da Platt scaling yana sauya maki na yanke shawara marasa sarrafawa zuwa yiwuwa da aka daidaita daidai wanda za a iya amfani da shi wajen tura dokokin kasuwanci masu matakala kamar aika maki sama da 0.5 zuwa duba hannu da toshe maki sama da 0.9 kai tsaye. Na uku, siginan class_weight='balanced' ta atomatik tana biya rashin daidaituwar da ke cikin saitunan bayanan satar kudi inda ma'amalolin gaskiya suka fi ma'amalolin satar kudi yawa, yana tabbatar da cewa samfurin bai kawai koyi kiran kome gaskiya don cimma daidaituwa mai girma ba.

A cikin yanayin samarwa na gaskiya na Amazon, gine-ginen yana zuwa nisa fiye. Ana sake horar da samfurori a ci gaba yayin da sababbin alamu na satar kudi ke bayyana. Ana wadatar da halaye da sigina bisa ginshiƙi kamar ko asusu yana da alaƙa da hanyoyin sadarwa na satar kudi da ake sani. Sau da yawa ana amfani da fitar SVM na Kernel a matsayin sigina ɗaya cikin da yawa a cikin babban samfurin taro, yana ba da gudummawar ƙarfinsa na musamman wajen zana iyakoki marasa layi ta hanyoyin da samfurori bisa bishiya ba za su iya daidaita ba.

Fahimtar dabara

Ga masu AI da jagororin kasuwanci, darasin wannan labarin ya zurfa fiye da gano satar kudi. Hanyoyin Kernel suna bayyana gaskiya mai faɗi game da tsarin wayo: ƙasar shigarwa mai baƙin ƙarfe ba kasafai inda amsar take ba. Bayanan duniya ta gaskiya suna da rikitarwa, rikice juna, da cika da dangantaka marasa layi da tunanin layi ba zai iya warware ba. Ƙarfin dabarar Kernel shi ne tana ba ka damar yin hujja a cikin ƙasar lissafi mai wadata sosai ba tare da tsadar lissafi na zuwa can da gaske ba. Ka'idar guda, sauya matsala mai wuya zuwa wadda amsar ke zama a bayyane, tana shafar tsarin shawarwari, lissafin haɗari, kulawa mai annabta, ganowa na kiwon lafiya, da kusan dukkan matsalolin machine learning masu muhimmanci a girman kamfani. Jagororin kasuwanci da suka fahimci wannan za su yi tambayoyi mafi kyau ga ƙungiyoyin kimiyyar bayanansu. Injiniyoyin da suka fahimci wannan za su gina tsarin da yake aiki da gaske lokacin da bayanan suka ƙi kasancewa da tsafta da sauƙi.

Gano satar kudi na Amazon ba sihiri ba ne. Lissafi ne da yake aiki ba a gani ba da sauri mai yawa, kuma yanzu kuna fahimtar daidai yadda yake.

#MachineLearning #ArtificialIntelligence #DataScience #Mathematics #Amazon #FraudDetection #KernelMethods #SVM #RBFKernel #PolynomialKernel #TheMagicOfHiddenMathematics #AIEngineering #MLEngineering #Python #SVMClassification #EnterpriseAI #AIStrategy #DeepLearning #NeuralNetworks #TechLeadership

Asalin an buga shi a LinkedIn ta hannun Keshavan Thiagaraja. Karanta asalin a LinkedIn. [1]