Riba:
- Ikon kafa samfurin layi na gaba ɗaya (GLM), abubuwan haɗari, jadawalin kuɗin fito da ra'ayoyi masu ƙima tare da tallafin fasaha na wucin gadi da fassara ƙididdiga zuwa harshen kasuwanci
- Ability don tattauna ma'auni na zaɓi mai canzawa, hulɗa, wuce gona da iri, da fassarar ƙirar koyon injin (GBM) tare da hankali na wucin gadi.
- Kasancewa iya fahimtar cewa ƙirar farashi ba ta da 'yanci daga sauye-sauyen da aka haramta/wariya kuma an bar bin jadawalin kuɗin fito na ƙarshe tare da doka da gasa ga mai aiki.
Ba a ƙayyade farashin tsarin inshora ba da gangan. Hadarin hatsarin ababen hawa ga dan shekara 22, sabon direba mai lasisi da ke zaune a babban birni ya fi na direba mai shekaru 45 da gogewar shekaru 20; A cikin tsarin gaskiya, ƙimar kuɗi kuma yakamata ya bambanta. Aikin ɗan wasan kwaikwayo na auna wannan bambanci da kuma nuna shi akan farashi ana kiransa farashi da rarrabuwar haɗari. A cikin wannan rukunin, za mu tattauna tsarin ƙirar gabaɗaya (GLM), ƙashin bayan farashi, da hanyoyin koyon injin sa, kuma mu ga yadda ake amfani da AI cikin aminci cikin wannan tsari.
Kalmomin asali kaɗan. Halin haɗari shine mai canzawa wanda ke shafar haɗarin kuma ana amfani dashi cikin farashi (shekaru, shekarun abin hawa, yanki, amfani da aka yi niyya). Kudaden kuɗin fito wani tsari ne na ƙa'idodi waɗanda ke ƙayyadaddun yadda za a ƙididdige ƙimar ƙimar bisa ga abubuwan haɗari. Ƙididdigar haɗari shine tsantsar asarar asarar da ake tsammani; Lokacin da aka ƙara kashe kuɗi, kwamitocin, ribar riba da ƙimar babban birnin, an kafa ƙimar kasuwanci (farashin tallace-tallace). Bari mu tunãtar daga farko: AI yana ba da shawara ga masu canji, gina ƙira, bayyana ƙididdiga; Amma wanne mabambanta ne na doka da ɗabi'a kuma ko jadawalin kuɗin fito na ƙarshe ya bi doka da gasa ya kasance na ƙungiyar aiki da sashin yarda.
Me yasa GLM shine ma'auni a cikin aikin
Hanyar da aka fi amfani da ita a cikin farashi shine GLM. GLM yana nufin “samfurin linzamin kwamfuta na gabaɗaya”: yana tsawaita koma bayan layin na gargajiya don dacewa da maƙasudan da ba a saba rarrabawa ba, kamar bayanan lalacewa. Yana da sassa biyu na asali. Iyalin Rarraba: An zaɓi Poisson don mita, an zaɓi gamma don ƙarfi (hankali a cikin naúrar 2). Ayyukan haɗin gwiwa (haɗin kai): yawanci logarithmic, wanda ke ba da damar abubuwan su sami tasiri mai yawa. Tsarin nau'ikan nau'ikan yana da ƙima sosai a zahiri saboda haka ne daidai yadda aka saita jadawalin kuɗin fito: tushen ƙimar ƙimar × shekaru × yanki factor factor × abin hawa.
Babban fa'idar GLM shine fassarar. Ƙididdigar ƙididdiga ta gaya muku kai tsaye: "Ƙungiyar direbobin matasa suna ƙara mitar da sau 1.6 idan aka kwatanta da ƙungiyar tunani." Wannan gaskiyar tana da mahimmanci ta hanyoyi uku: (1) mai gudanarwa da bincike na ciki na iya fahimta da amincewa da tsarin; (2) haramtacciyar tasiri / nuna bambanci yana bayyane; (3) farashin dabaru za a iya bayyana wa tallace-tallace da kuma gudanarwa tawagar. Samfuran koyon injina masu rikitarwa (a ƙasa) wani lokacin suna ba da daidaito mafi girma, amma ta hanyar nuna gaskiya.
Alamomi: Ƙididdigar GLM tana kan sikelin log. Tambayi AI don canza ƙididdiga zuwa "haɓaka haɓaka" tare da exp(coefficient); Zai fi sauƙi a fassara zuwa harshen kasuwanci (misali coefficient 0.47 → factor ≈ 1.60 → "60% riskier").
Zaɓuɓɓuka daban-daban, overfitting da koyon inji
Mafi mahimmancin yanke shawara a cikin farashi shine waɗanne masu canji zasu kasance a cikin ƙirar. Akwai tarko guda biyu. Ƙananan masu canji sun makantar da ƙirar: idan an rasa mahimman direban haɗari, farashin zai yi kuskure. Maɓalli mai yawa / wuce gona da iri yana sa ƙirar ta haddace bayanan da suka gabata: ƙirar tana kuskure amo don sigina kuma ta gaza tare da sabbin manufofi. An kafa ma'auni ta hanyar tabbatarwa - rarraba bayanai zuwa horo da gwajin gwaji da gwada shi akan bayanan da samfurin ba ya gani, sauƙi da kuma dalili na ainihi.
Ma'amala yana da mahimmanci: wani lokacin haɗakar tasirin abubuwa biyu ya bambanta da jimlar tasirinsu daban-daban (motar matasa + na motsa jiki na iya zama haɗari fiye da jimlar tasirinsu). A cikin GLM, ana ƙara mu'amala a bayyane.
A cikin 'yan shekarun nan, samfura irin su GBM (na'ura mai haɓaka gradient - hanya mai ƙarfi na koyon inji wanda ke haɗa yawancin ƙananan bishiyoyi masu yanke shawara) sun zama gama gari a farashin. Waɗannan suna ɗaukar hadaddun tsari da hulɗa ta atomatik, galibi suna ba da ƙarin ingantattun tsinkaya fiye da GLM. Amma ya zo a farashi: dabi'ar zama "akwatin baki". Ayyukan gama gari a yau shine a yi amfani da GBM don ganowa da haɓaka ra'ayi da GLM don ƙimar kuɗin fito na ƙarshe; ko fassara fitowar GBM tare da kayan aikin bayyanawa kamar SHAP.
Teburin da ke gaba ya kwatanta hanyoyin biyu:
ma'auni
GLM
GBM (ilimin inji)
Fassara
High (ƙididdigar a kan)
Ƙananan (yana buƙatar kayan aikin rubutu)
Kama mu'amala
Ƙara da hannu
atomatik
Hadarin wuce gona da iri
ƙananan matsakaici
Babban (ana buƙatar daidaitawa a hankali)
Amintacce mai gudanarwa / dubawa
mai sauki
Wuya, yana buƙatar ƙarin takaddun bayanai
Yawan amfani
jadawalin kuɗin fito na ƙarshe
gano, kwatanta
Yadda ake amfani da AI a cikin farashi
1) Fassara ƙididdigar GLM zuwa harshen kasuwanci:
Na saita mitar GLM (Poisson, log link). Wasu ƙididdiga (ma'auni): age_group_18_25: 0.47; yanki_yawan: 0.22; arac_yasi_10 da: -0.15. Maida kowane ɗayan zuwa ma'auni mai yawa tare da exp() kuma bayyana shi a cikin jumla ɗaya wanda manajan zai iya fahimta. Hakanan kuma tunatar da menene ƙungiyar tunani.
2) Canje-canje / Tattaunawar hulɗa:
Matsayinku: mataimakin farashi. Canje-canjen ɗan takara na don ƙirar mitar zirga-zirga sune: shekaru, jinsi, yanki, shekarun abin hawa, ƙarfin injin, kilomita na shekara, sana'a. - Waɗanne sauye-sauye ne za su iya zama masu haɗari ta fuskar tsari/ɗa'a (misali nuna bambanci kai tsaye)? - Wadanne mu'amala tsakanin bangarorin biyu ne zai yi ma'ana don gwadawa? - Wadanne matakan tabbatarwa zan bi don gujewa wuce gona da iri? Yanke shawara; Ka ba ni tsarin sarrafawa da tattaunawa.
3) Lambar sarrafawa ta wuce gona da iri:
Rubuta lambar sharhi wanda ke kimanta GLM tare da tabbatar da k-fold ta amfani da Python + scikit-Learn / statsmodels. Yi amfani da karkatar da Poisson azaman ma'auni. Kwatanta horon da gwajin maki kuma bayyana a cikin layin sharhi yadda ake karanta alamar wuce gona da iri. Laburare karya ne.
4) Nuna wariya/masu maye:
'Postcode' ya zama mai ƙarfi mai ƙarfi a ƙirar farashi na. Bayyana haɗarin wannan a kaikaice yana wakiltar sifar da aka haramta (misali ƙabila, samun kudin shiga) a matsayin madaidaicin wakili. - Wane bincike zan yi don gwada wannan hadarin? - Wadanne hanyoyi ne akwai don rage haɗarin? Ba da shawarar shari'a; zana tsarin fasaha da ɗabi'a.
Rauni mai ƙarfi / Ƙarfi mai ƙarfi
Rawanin faɗakarwa:
Ƙirƙiri mafi kyawun ƙirar farashi don inshorar zirga-zirga.
"Mafi kyau" ba a bayyana shi ba; Babu bayanai, babu hani, babu doka. AI yana ba da cikakkiyar amsa, mara amfani.
Ƙarfi mai ƙarfi:
Matsayinku: mataimaki ga aikin farashin farashi. Magana: Ina gina samfurin mitar reshen zirga-zirga, GLM (Poisson, hanyar haɗin shiga). Matsalolin da nake da su: ƙungiyar shekaru, shekarun abin hawa, yanki (lardi), kilomita na shekara, amfanin da aka yi niyya. Ƙuntatawa: ba za a iya amfani da jinsi ta hanyar doka ba; Ina bukata in guje wa nuna bambanci kai tsaye.Aiki:1) Ba da shawarar ƙayyadaddun ƙirar ƙirar farko tare da waɗannan masu canji (ciki har da ƙungiyoyin tunani).2) Ba da dalilai na hulɗar 2 yakamata in gwada.3) Ba da jerin matakai don bincika wuce gona da iri.4) Ƙara bayanin kula na musamman ga 'yanki' dangane da nuna bambanci kai tsaye. Zan yanke shawara; Kuna ba da tsari da hujja.
uku mini lokuta
Case 1 - Darajar gaskiya. Wani kamfani ya gabatar da ingantaccen samfurin farashin da ya gina tare da GBM ga mai gudanarwa, amma ba zai iya bayyana ma'auni ba; amincewa ya jinkirta. The actuary ya gina GLM wanda ya ɗauki sigina iri ɗaya; Daidaito ya ragu da kashi 2 cikin ɗari, amma saboda kowane abu ana iya ƙididdige shi, samfurin ya inganta cikin wata guda. An ajiye GBM don bincike, GLM ya zama jadawalin kuɗin fito. YZ cikin sauri ya samar da lambar da bayanin mai gudanarwa yana kwatanta aikin samfuran biyu.
Case na 2 - Tarko mai wuce gona da iri. Wani mataimaki ya sami cikakkiyar maki akan bayanan horo lokacin da ya ƙara fiye da 40 masu canji da kuma hulɗa da yawa zuwa samfurin. Amma akan tabbatarwa, ƙimar gwajin ta rushe: ƙirar ta haddace amo. Ayyukan ainihin duniya ya ƙaru lokacin da aka rage adadin masu canji zuwa 12 kuma aka sauƙaƙe. Darasi: duba makin bayanan da ba a taɓa yin irinsa ba, ba makin horo ba.
Harka 3 - Wariya ta kai tsaye. A cikin samfuri ɗaya, madaidaicin "unguwa" yayi bayanin ƙimar ƙima sosai. Binciken ya nuna cewa wannan madaidaicin ya fi yawa tare da maida hankali na kabilanci, watau, wakili ne don halayen da aka haramta. Mai kunnawa ya maye gurbin wannan canji tare da ƙarin alamun haɗari mai tsaka tsaki (nau'in hanya, yanayin filin ajiye motoci); Haɗarin ɗabi'a da na shari'a duka sun ragu. AI ta samar da bincike mai ma'auni mai ma'auni da madaidaicin shawarwari; Ƙungiyar aiki da yarda ta yanke shawara.
Kuskuren gama gari
- Yin samfurin da ba a iya fassara shi ya zama girke-girke na ƙarshe. Farashin da ba za a iya bayyana shi ga mai gudanarwa, dubawa da abokin ciniki ba shi da dorewa; Bayyana gaskiya yana da mahimmanci.
- Dogara akan maki ilimi. Ana gano wuce gona da iri ne kawai a cikin bayanan da ba a gani (tabbatar da giciye).
- Amfani da haramtattun masu canji/masu maye ba tare da dubawa ba. Mabambanta kamar zip code da sana'a na iya ɗaukar wariya kai tsaye; Tabbatar gwada shi.
- Ƙididdigar haɗarin haɗari tare da ƙimar kasuwanci. Farashin lalacewa kawai ba shine farashin siyarwa ba; an kara kashe kudi, riba da kudin jari.
- Barin ma'auni akan ma'aunin log da kuskuren fassara shi. Yin sharhi ba tare da canza shi zuwa ma'auni mai yawa tare da exp() zai haifar da kuskure ba.
Tsanaki: Shawarar AI na ƙirar da "ba da mafi kyawun daidaito" ba ta atomatik samfurin da "ya kamata a yi amfani da shi ba". Bayyana gaskiya, gaskiya da bin doka sharuɗɗa ne na tilas da kuma daidaito a farashin aiki.
a takaice
Farashi shine tsarin auna haɗari tare da abubuwan haɗari da canza shi zuwa ƙimar kuɗi mai kyau. GLM shine ma'auni na farashin aiki tare da nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'in nau'ino) ratings na fassara; Ana canza ƙididdiga zuwa abubuwa tare da exp(). Samfuran koyon inji kamar GBM na iya zama daidai, amma suna rage bayyana gaskiya kuma galibi ana amfani dasu don ganowa. Ya kamata a kiyaye zaɓin maɓalli daga wuce gona da iri ta hanyar tabbatarwa, kuma ya kamata a kula da wariya a kaikaice (mai canzawa) a hankali. AI yana gina samfurin, ya rubuta lambar kuma ya bayyana ma'auni; Amma bin doka-da'a da jadawalin kuɗin fito na ƙarshe na cikin sashin aiki da yarda.
Aikin aikace-aikace
Jerin abubuwan haɗari 5-6 don manyan. Tambayi AI don (a) ƙayyadaddun farko na GLM tare da waɗannan abubuwan, (b) hulɗar 2 don gwadawa da dalilansu, (c) tantance masu canji waɗanda za su iya zama cikin haɗari don nuna bambanci kai tsaye. Sa'an nan, ba da misali na 3 log-coefficients, tambayi AI don canza shi zuwa ma'auni mai yawa tare da exp() kuma canza shi zuwa harshen kasuwanci, kuma tabbatar da abubuwan da hannu.
jerin abubuwan dubawa
- [ ] Za a iya fassara samfurina; Zan iya fassara tasirin kowane abu zuwa harshen kasuwanci?
- [ ] Na duba don wuce gona da iri ta hanyar tabbatarwa?
- [ ] Shin na bincika don nuna bambanci kai tsaye don haramtattun masu canji da wakili?
- [ ] Shin a sane na raba ƙimar haɗari da ƙimar kasuwanci?
- [ ] Shin na canza ma'auni zuwa masu haɓakawa tare da exp() kuma na fassara su daidai?
- [ ] Shin na yanke shawarar jadawalin kuɗin fito na ƙarshe tare da sashin doka da bin doka?