Yadda muka yi annabta churn ta hanyar tunkararta kamar bala'i

Wani lokaci, don magance matsala, kawai kuna buƙatar kallon ta ta wani kusurwa daban. Ko da a cikin shekaru 10 da suka gabata an magance irin wadannan matsaloli ta hanya guda tare da tasiri daban-daban, ba gaskiya ba ne cewa wannan hanya ita ce kadai.

Akwai irin wannan batu kamar abokin ciniki churn. Abun ba makawa ne, saboda abokan ciniki na kowane kamfani na iya, saboda dalilai da yawa, daina amfani da samfuransa ko sabis. Tabbas, ga kamfani, churn abu ne na halitta, amma ba aikin da ya fi kyawawa ba, don haka kowa yana ƙoƙari ya rage girman wannan churn. Mafi kyau kuma, kintace yuwuwar churn ga wani nau'in masu amfani, ko takamaiman mai amfani, da ba da shawarar wasu matakai don riƙe su.

Wajibi ne a yi nazari da ƙoƙarin riƙe abokin ciniki, idan zai yiwu, don akalla dalilai masu zuwa:

• jawo sababbin abokan ciniki ya fi tsada fiye da hanyoyin riƙewa. Don jawo hankalin sababbin abokan ciniki, a matsayin mai mulkin, kuna buƙatar kashe kuɗi (talla), yayin da abokan ciniki na yanzu za a iya kunna su tare da tayin na musamman tare da yanayi na musamman;
• Fahimtar dalilan da yasa abokan ciniki ke barin shine mabuɗin haɓaka samfura da sabis.

Akwai daidaitattun hanyoyi don tsinkayar tsinkewa. Amma a daya daga cikin gasar zakarun AI, mun yanke shawarar gwada rarraba Weibull don wannan. Yawancin lokaci ana amfani da shi don nazarin rayuwa, hasashen yanayi, nazarin bala'i, injiniyan masana'antu da makamantansu. Rarraba Weibull aiki ne na rarrabawa na musamman wanda aka daidaita ta sigogi biyu и .

Wikipedia

Gabaɗaya, abu ne mai ban sha'awa, amma don tsinkayar fitarwa, kuma a cikin fintech gabaɗaya, ba a amfani da shi sau da yawa. A ƙasa da yanke za mu gaya muku yadda mu (Data Mining Laboratory) muka yi wannan, tare da lashe zinare a lokaci guda a gasar Intelligence ta Artificial Intelligence a cikin "AI a Banks".

Game da churn gabaɗaya

Bari mu ɗan fahimci abin da abokin ciniki churn yake da kuma dalilin da yasa yake da mahimmanci. Tushen abokin ciniki yana da mahimmanci ga kasuwanci. Sabbin abokan ciniki suna zuwa wannan tushe, misali, sun koyi game da samfur ko sabis daga talla, suna rayuwa na ɗan lokaci (amfani da samfuran sosai) kuma bayan ɗan lokaci sun daina amfani da shi. Wannan lokacin ana kiransa “Customer Lifecycle” - kalmar da ke bayyana matakan da abokin ciniki ke bi lokacin da ya koyi samfur, ya yanke shawarar siyan, biya, amfani kuma ya zama mabukaci mai aminci, kuma a ƙarshe ya daina amfani da samfuran. saboda wani dalili ko wani. Saboda haka, churn shine mataki na ƙarshe na rayuwar abokin ciniki, lokacin da abokin ciniki ya daina amfani da ayyukan, kuma ga kasuwanci wannan yana nufin cewa abokin ciniki ya daina kawo riba ko wata fa'ida.

Kowane abokin ciniki na banki wani mutum ne na musamman wanda ya zaɓi ɗaya ko wani katin banki musamman don bukatunsa. Idan kuna tafiya akai-akai, kati mai mil zai zo da amfani. Sayi da yawa - sannu, katin cashback. Ya saya da yawa a cikin takamaiman shaguna - kuma akwai rigar filastik abokin tarayya na musamman don wannan. Tabbas, wani lokacin ana zaɓar katin bisa ma'aunin "Sabis mafi arha". Gabaɗaya, akwai isassun masu canji a nan.

Kuma mutum ma ya zaɓi bankin da kansa - shin akwai wata ma'ana a zabar katin daga banki wanda rassansa ke kawai a Moscow da yankin, lokacin da kuke daga Khabarovsk? Ko da katin daga irin wannan banki ya fi riba aƙalla sau 2, kasancewar rassan banki a kusa har yanzu yana da mahimmancin ma'auni. Haka ne, 2019 ya riga ya zo kuma dijital shine komai namu, amma yawancin batutuwa tare da wasu bankuna za a iya warware su kawai a cikin reshe. Bugu da ƙari, kuma, wasu ɓangaren jama'a sun amince da banki na zahiri fiye da aikace-aikacen kan wayar hannu, wannan kuma yana buƙatar la'akari.

A sakamakon haka, mutum na iya samun dalilai da yawa na ƙin samfuran banki (ko bankin kansa). Na canza ayyuka, kuma kuɗin kuɗin katin ya canza daga albashi zuwa "Don kawai mutane," wanda ba shi da riba. Na ƙaura zuwa wani birni inda babu rassan banki. Ban ji daɗin hulɗa da ma'aikacin da bai cancanta ba a reshen. Wato, ƙila a sami ƙarin dalilai na rufe asusu fiye da amfani da samfur.

Kuma abokin ciniki ba zai iya bayyana niyyarsa kawai ba - zo banki kuma ya rubuta sanarwa, amma kawai dakatar da amfani da samfuran ba tare da ƙare kwangilar ba. An yanke shawarar yin amfani da na'ura koyo da AI don fahimtar irin waɗannan matsalolin.

Haka kuma, churn abokin ciniki na iya faruwa a kowace masana'antu (telecom, masu samar da Intanet, kamfanonin inshora, gabaɗaya, duk inda akwai tushen abokin ciniki da ma'amala na lokaci-lokaci).

Me muka yi

Da farko, ya zama dole don bayyana iyaka mai haske - daga wane lokaci za mu fara la'akari da abokin ciniki ya bar. Daga ra'ayi na bankin da ya ba mu bayanai don aikinmu, matsayin aikin abokin ciniki ya kasance binary - yana aiki ko a'a. Akwai tuta ACTIVE_FLAG a cikin "Ayyukan" tebur, wanda darajarsa zai iya zama ko dai "0" ko "1" ("Ba aiki" da "Active" bi da bi). Kuma duk abin da zai zama lafiya, amma mutum ne irin wannan cewa zai iya rayayye amfani da shi na wani lokaci, sa'an nan kuma fada daga cikin aiki jerin ga wata daya - ya yi rashin lafiya, ya tafi wata kasa a kan hutu, ko ma ya tafi gwada wani. katin daga wani banki. Ko watakila bayan dogon lokaci na rashin aiki, fara amfani da sabis na banki kuma

Saboda haka, mun yanke shawarar kiran lokacin rashin aiki wani lokaci mai ci gaba wanda aka saita tutarsa ​​zuwa "0".

Abokan ciniki suna motsawa daga rashin aiki zuwa aiki bayan lokutan rashin aiki na tsayi daban-daban. Muna da damar da za mu ƙididdige ƙimar ƙimar ƙimar "amintacce na lokutan rashin aiki" - wato, yiwuwar mutum zai sake amfani da samfuran banki bayan rashin aiki na ɗan lokaci.

Misali, wannan jadawali yana nuna ci gaban ayyuka (ACTIVE_FLAG=1) na abokan ciniki bayan watanni da yawa na rashin aiki (ACTIVE_FLAG=0).

Anan za mu ɗan fayyace saitin bayanan da muka fara aiki da su. Don haka, bankin ya ba da cikakkun bayanai na tsawon watanni 19 a cikin tebur masu zuwa:

• "Ayyukan" - mu'amalar abokin ciniki na wata-wata (ta katunan, a cikin banki na Intanet da banki ta wayar hannu), gami da biyan kuɗi da bayanai kan canji.
• "Katuna" - bayanai game da duk katunan da abokin ciniki ke da shi, tare da cikakken jadawalin jadawalin kuɗin fito.
• "Yarjejeniyar" - bayanai game da yarjejeniyar abokin ciniki (duka bude da rufe): rance, adibas, da dai sauransu, nuna sigogi na kowane.
• "Abokan ciniki" - saitin bayanan alƙaluma (jinsi da shekaru) da kuma kasancewar bayanan lamba.

Don aiki muna buƙatar duk teburin ban da "Taswirar".

Akwai wata wahala a nan - a cikin wannan bayanan bankin bai nuna irin ayyukan da aka yi a kan katunan ba. Wato, za mu iya fahimtar ko akwai ciniki ko a'a, amma ba za mu iya tantance nau'in su ba. Saboda haka, ba a sani ba ko abokin ciniki yana cire tsabar kudi, yana karɓar albashi, ko kuma yana kashe kuɗin akan sayayya. Hakanan ba mu da bayanai kan ma'auni na asusu, wanda zai yi amfani.

Samfurin da kansa bai nuna son kai ba - a cikin wannan samfurin, sama da watanni 19, bankin bai yi wani yunƙuri na riƙe abokan ciniki da rage fitar da su ba.

Don haka, game da lokutan rashin aiki.

Don tsara ma'anar chun, dole ne a zaɓi lokacin rashin aiki. Don ƙirƙirar hasashen tsinkaya a lokaci guda , Dole ne ku sami tarihin abokin ciniki na akalla watanni 3 a tazara . Tarihinmu ya iyakance ga watanni 19, don haka mun yanke shawarar ɗaukar lokacin rashin aiki na watanni 6, idan akwai. Kuma don mafi ƙarancin lokaci don tsinkaya mai inganci, mun ɗauki watanni 3. Mun dauki alkalumman na tsawon watanni 3 da 6 bisa ka'ida bisa nazarin halayen bayanan abokin ciniki.

Mun tsara ma'anar churn kamar haka: watan abokin ciniki wannan shine wata na farko da ACTIVE_FLAG=0, inda daga wannan watan akwai akalla sifili shida a jere a cikin filin ACTIVE_FLAG, wato watan da abokin ciniki bai yi aiki ba har tsawon watanni 6.

Adadin abokan cinikin da suka tafi

Yawan sauran abokan ciniki

Ta yaya ake lissafin churn?

A irin wannan gasa, da kuma a aikace a gaba ɗaya, ana hasashen fitowar ta ta wannan hanya. Abokin ciniki yana amfani da samfurori da ayyuka a lokuta daban-daban, bayanai akan hulɗa tare da shi ana wakilta shi azaman vector na fasali na tsayayyen tsayi n. Yawancin wannan bayanin ya haɗa da:

• Bayanan da ke nuna mai amfani (bayanan alƙaluma, ɓangaren tallace-tallace).
• Tarihin amfani da samfurori da sabis na banki (waɗannan ayyuka ne na abokin ciniki waɗanda koyaushe ke ɗaure da takamaiman lokaci ko lokacin tazarar da muke buƙata).
• Bayanan waje, idan yana yiwuwa a samu shi - alal misali, sake dubawa daga cibiyoyin sadarwar jama'a.

Kuma bayan haka, suna samun ma'anar churn, daban-daban ga kowane aiki. Sannan suna amfani da na'urar koyon na'ura, wanda ke hasashen yiwuwar abokin ciniki ya fita bisa ga vector na dalilai . Don horar da algorithm, ana amfani da ɗaya daga cikin sanannun ka'idoji don gina gunkin bishiyar yanke shawara, XGBoost, Haske, KatBoost ko gyare-gyarensa.

Algorithm din kanta ba shi da kyau, amma yana da babban lahani idan ya zo ga tsinkayar churn.

• Ba shi da abin da ake kira "memory". Shigar da samfurin ƙayyadadden adadin fasalulluka ne waɗanda suka dace da yanayin yanzu cikin lokaci. Don adana bayanai game da tarihin canje-canje a cikin sigogi, yana da mahimmanci don ƙididdige fasalulluka na musamman waɗanda ke nuna canje-canje a cikin sigogi akan lokaci, misali, lamba ko adadin ma'amalolin banki a cikin watanni 1,2,3, XNUMX, XNUMX na ƙarshe. Wannan tsarin zai iya nuna ɗan lokaci kaɗan kawai yana nuna yanayin canje-canje na ɗan lokaci.
• Kafaffen hasashen hasashen. Samfurin yana iya yin hasashen ɓarnar abokin ciniki kawai na wani ƙayyadadden lokaci, misali, hasashen wata ɗaya gaba. Idan ana buƙatar tsinkaya don wani lokaci daban-daban, alal misali, watanni uku, to kuna buƙatar sake gina tsarin horo da sake horar da sabon samfurin.

Hanyarmu

Mun yanke shawarar nan da nan cewa ba za mu yi amfani da daidaitattun hanyoyin ba. Ban da mu, an samu karin mutane 497 da suka yi rajista a gasar, kowannensu yana da kwarewa sosai a bayansu. Don haka ƙoƙarin yin wani abu bisa ga daidaitaccen tsari a irin waɗannan yanayi ba kyakkyawan ra'ayi ba ne.

Kuma mun fara magance matsalolin da ke fuskantar samfurin rabe-raben binary ta hanyar tsinkayar yiwuwar rarraba lokutan chun abokin ciniki. Ana iya ganin irin wannan hanya a nan, yana ba ku damar yin hasashen churn da sassauƙa kuma ku gwada ƙarin hadaddun hasashe fiye da tsarin al'ada. A matsayin dangin rabawa da ke yin ƙirar lokacin fita, mun zaɓi rarraba Weibull don yawan amfani da shi wajen nazarin rayuwa. Ana iya kallon halin abokin ciniki azaman nau'in tsira.

Anan akwai misalan rabe-raben yuwuwar Weibull dangane da sigogi и :

Wannan shine aikin yuwuwar yawa na abokan ciniki daban-daban guda uku suna jinkirin lokaci. Ana gabatar da lokaci a cikin watanni. A wasu kalmomi, wannan jadawali yana nuna lokacin da abokin ciniki zai fi jin dadi a cikin watanni biyu masu zuwa. Kamar yadda kake gani, abokin ciniki tare da rarraba yana da babban damar barin baya fiye da abokan ciniki tare da Weibull(2, 0.5) da Weibull. (3,1) rabawa.

Sakamakon shine samfurin wanda, ga kowane abokin ciniki, ga kowane
watan yayi hasashen ma'auni na rarrabawar Weibull, wanda mafi kyawun nuna faruwar yuwuwar fita daga lokaci. A cikin ƙarin bayani:

• Abubuwan da aka yi niyya akan saitin horon shine lokacin da ya rage har sai lokacin da aka yanke a cikin takamaiman wata don takamaiman abokin ciniki.
• Idan babu ƙimar churn ga abokin ciniki, muna ɗauka cewa lokacin churn ya fi adadin watanni daga watan yanzu zuwa ƙarshen tarihin da muke da shi.
• Samfurin da aka yi amfani da shi: cibiyar sadarwa na yau da kullun tare da Layer LSTM.
• A matsayin aikin asara, muna amfani da aikin yuwuwar log mara kyau don rarraba Weibull.

Ga fa'idodin wannan hanyar:

• Rarraba yuwuwar, ban da yuwuwar rabe-raben binary, yana ba da damar hasashen sassauƙa na abubuwan da suka faru daban-daban, alal misali, ko abokin ciniki zai daina amfani da sabis na banki a cikin watanni 3. Hakanan, idan ya cancanta, ana iya daidaita ma'auni daban-daban akan wannan rarraba.
• Cibiyar sadarwa ta LSTM na yau da kullun tana da ƙwaƙwalwar ajiya kuma tana amfani da duk tarihin da ke akwai yadda ya kamata. Yayin da aka faɗaɗa ko kuma inganta labarin, daidaito yana ƙaruwa.
• Za'a iya daidaita tsarin cikin sauƙi yayin rarraba lokutan lokaci zuwa ƙananan (misali, lokacin raba watanni zuwa makonni).

Amma bai isa ba don ƙirƙirar samfuri mai kyau; Hakanan kuna buƙatar kimanta ingancinsa yadda yakamata.

Ta yaya aka tantance inganci?

Mun zaɓi Lift Curve azaman ma'auni. Ana amfani da shi a cikin kasuwanci don irin waɗannan lokuta saboda fassarar fassararsa, an kwatanta shi da kyau a nan и a nan. Idan kun bayyana ma'anar wannan awo a cikin jumla ɗaya, zai zama "Sau nawa algorithm ke yin hasashen mafi kyau a farkon. % fiye da bazuwar."

Samfuran horo

Yanayin gasa bai kafa takamaiman ma'auni mai inganci wanda za'a iya kwatanta samfura da hanyoyin daban-daban ba. Bugu da ƙari, ma'anar churn na iya zama daban-daban kuma yana iya dogara ne akan bayanin matsalar, wanda, bi da bi, an ƙaddara shi ta hanyar manufofin kasuwanci. Don haka, don fahimtar wace hanya ce mafi kyau, mun horar da samfuran guda biyu:

1. Hanyar rarraba binaryar da aka saba amfani da ita ta amfani da gunkin yanke shawara bishiyar koyo algorithm (Haske);
2. Weibull-LSTM samfurin

Saitin gwajin ya ƙunshi abokan ciniki 500 da aka zaɓa waɗanda ba sa cikin tsarin horo. An zaɓi madaidaicin madaidaicin don ƙirar ta amfani da ingantaccen giciye, wanda abokin ciniki ya rushe. An yi amfani da sifofi iri ɗaya don horar da kowane samfuri.

Saboda gaskiyar cewa samfurin ba shi da ƙwaƙwalwar ajiya, an ɗauki siffofi na musamman don shi, yana nuna rabon canje-canje a cikin sigogi na wata daya zuwa matsakaicin darajar ga sigogi a cikin watanni uku na ƙarshe. Abin da ke nuna ƙimar canjin ƙima a cikin tsawon watanni uku na ƙarshe. Idan ba tare da wannan ba, ƙirar tushen dajin na Random zai kasance cikin asara dangane da Weibull-LSTM.

Me yasa LSTM tare da rarraba Weibull ya fi tsarin tsarin yanke shawara mai tarin yawa

Komai ya bayyana a nan a cikin hotuna biyu kawai.

Kwatanta Lift Curve don algorithm na gargajiya da Weibull-LSTM

Kwatanta ma'aunin Lift Curve ta wata-wata don algorithm na gargajiya da Weibull-LSTM

Gabaɗaya, LSTM ya fi na gargajiya algorithm a kusan dukkan lokuta.

Tsinkaya tsinkaya

Samfuran da ke kan hanyar sadarwa mai maimaitawa tare da sel LSTM tare da rarraba Weibull na iya yin hasashen ɓarna a gaba, alal misali, tsinkayar ɓarnar abokin ciniki a cikin watanni n masu zuwa. Yi la'akari da yanayin n = 3. A wannan yanayin, kowane wata, cibiyar sadarwar jijiyoyi dole ne ta ƙayyade daidai ko abokin ciniki zai tafi, farawa daga wata mai zuwa kuma har zuwa watan n. A wasu kalmomi, dole ne ya ƙayyade daidai ko abokin ciniki zai kasance bayan n watanni. Ana iya ɗaukar wannan a matsayin tsinkaya a gaba: tsinkaya lokacin da abokin ciniki ke fara tunanin barin.

Bari mu kwatanta Lift Curve don Weibull-LSTM 1, 2 da watanni 3 kafin fitowar:

Mun riga mun rubuta a sama cewa hasashen da aka yi don abokan ciniki waɗanda ba su da aiki na ɗan lokaci suna da mahimmanci. Saboda haka, a nan za mu ƙara zuwa samfurin irin waɗannan lokuta lokacin da abokin ciniki da ya tafi ya riga ya yi aiki na wata ɗaya ko biyu, kuma duba cewa Weibull-LSTM yana rarraba irin waɗannan lokuta daidai. Tun da irin waɗannan shari'o'in sun kasance a cikin samfurin, muna tsammanin cibiyar sadarwar za ta kula da su da kyau:

Riƙewar abokin ciniki

A zahiri, wannan shine babban abin da za'a iya yi, samun bayanan hannu cewa irin waɗannan abokan ciniki suna shirin dakatar da amfani da samfurin. Da yake magana game da gina samfurin da zai iya ba da wani abu mai amfani ga abokan ciniki don riƙe su, ba za a iya yin wannan ba idan ba ku da tarihin irin wannan ƙoƙarin da zai ƙare da kyau.

Ba mu da irin wannan labarin, don haka muka yanke shawarar haka.

1. Muna gina samfurin da ke gano samfurori masu ban sha'awa ga kowane abokin ciniki.
2. Kowane wata muna gudanar da classifier da gano yiwuwar barin abokan ciniki.
3. Muna ba wa wasu abokan ciniki samfurin, bisa ga samfurin daga aya 1, kuma ku tuna ayyukanmu.
4. Bayan 'yan watanni, muna duban wanene daga cikin waɗannan masu yuwuwar barin abokan ciniki ya bari kuma waɗanda suka rage. Don haka, muna samar da samfurin horo.
5. Muna horar da ƙirar ta amfani da tarihin da aka samu a mataki na 4.
6. Optionally, mu maimaita hanya, musanya model daga mataki 1 da model samu a mataki na 5.

Ana iya yin gwajin ingancin irin wannan riƙewa ta gwajin A/B na yau da kullun - muna raba abokan ciniki waɗanda ke yuwuwar barin gida biyu. Muna ba da samfurori ga ɗaya bisa ga tsarin mu, kuma ga ɗayan ba mu bayar da kome ba. Mun yanke shawarar horar da samfurin da zai iya zama da amfani riga a aya ta 1 na misalinmu.

Mun so mu sanya sashi a matsayin mai fassara kamar yadda zai yiwu. Don yin wannan, mun zaɓi fasaloli da yawa waɗanda za a iya fassara su cikin sauƙi: jimlar adadin ma'amaloli, albashi, jimlar juyar da asusu, shekaru, jinsi. Ba a la'akari da fasali daga tebur na "Taswirori" a matsayin marasa fahimta, kuma ba a la'akari da fasali daga tebur na 3 "Kwangiyoyi" ba saboda rikitarwa na aiki don kauce wa ɗigon bayanai tsakanin saitunan tabbatarwa da tsarin horo.

An gudanar da tari ta amfani da samfuran Gaussian cakuda. Ma'aunin bayanin Akaike ya ba mu damar tantance 2 optima. Mafi kyawu na farko yayi daidai da tari 1. Mafi kyawu na biyu, ƙasan magana, yayi daidai da tari 80. Dangane da wannan sakamakon, zamu iya zana ƙarshe mai zuwa: yana da matukar wahala a raba bayanai zuwa gungu ba tare da an ba da bayanin fifiko ba. Don ingantacciyar tari, kuna buƙatar bayanai waɗanda ke bayyana kowane abokin ciniki daki-daki.

Don haka, an yi la'akari da matsalar ilmantarwa mai kulawa don baiwa kowane abokin ciniki samfurin daban. An yi la'akari da waɗannan samfuran: "Aikace-aikacen ajiya", "Katin Credit", "Overdraft", "Lamunin Mabukaci", "Lamunin Mota", "Mortgage".

Bayanan sun haɗa da ƙarin nau'in samfur guda ɗaya: "Asusun yanzu". Amma ba mu yi la'akari da shi ba saboda ƙarancin bayanansa. Ga masu amfani waɗanda abokan cinikin banki ne, i.e. bai daina amfani da samfuransa ba, an gina ƙirar don hasashen wane samfurin zai iya sha'awar su. An zaɓi koma bayan dabaru a matsayin abin ƙira, kuma an yi amfani da ƙimar ɗagawa na kashi 10 na farko azaman ma'aunin ƙima mai inganci.

Ana iya kimanta ingancin samfurin a cikin adadi.

Sakamakon samfurin shawarwarin samfur ga abokan ciniki

Sakamakon

Wannan tsarin ya kawo mu wuri na farko a cikin "AI a Banks" a rukunin RAIF-Challenge 2017 AI Championship.

A bayyane yake, babban abu shine a tunkari matsalar daga kusurwar da ba ta dace ba kuma a yi amfani da hanyar da aka saba amfani da ita don wasu yanayi.

Ko da yake ɗimbin fitowar masu amfani na iya zama bala'i ga ayyuka.

Ana iya la'akari da wannan hanyar don kowane yanki inda yake da mahimmanci a la'akari da fitar da kaya, ba kawai bankuna ba. Alal misali, mun yi amfani da shi don ƙididdige fitar da kanmu - a cikin sassan Siberian da St. Petersburg na Rostelecom.

"Data Mining Laboratory" kamfanin "Search portal"Sputnik"

source: www.habr.com