{"id":30044,"date":"2019-10-31T21:33:22","date_gmt":"2019-10-31T18:33:22","guid":{"rendered":"https:\/\/prohoster.info\/blog\/kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu\/"},"modified":"2019-10-31T21:33:22","modified_gmt":"2019-10-31T18:33:22","slug":"kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu","title":{"rendered":"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<p>N\u011bkdy je t\u0159eba se na probl\u00e9m pod\u00edvat z jin\u00e9ho \u00fahlu, abychom ho vy\u0159e\u0161ili. I kdy\u017e se v posledn\u00edch deseti letech podobn\u00e9 probl\u00e9my \u0159e\u0161ily stejn\u00fdm zp\u016fsobem s r\u016fzn\u00fdmi v\u00fdsledky, nen\u00ed jist\u00e9, \u017ee tento zp\u016fsob je jedin\u00fd.<\/p>\n<p>Existuje t\u00e9ma odlivu z\u00e1kazn\u00edk\u016f. Je to nevyhnuteln\u00fd jev, proto\u017ee z\u00e1kazn\u00edci jak\u00e9koli spole\u010dnosti mohou z mnoha d\u016fvod\u016f p\u0159estat vyu\u017e\u00edvat jej\u00ed produkty nebo slu\u017eby. Samoz\u0159ejm\u011b, pro spole\u010dnost je odliv \u2014 a\u010dkoliv p\u0159irozen\u00fd \u2014 ne\u017e\u00e1douc\u00ed jev, a proto se v\u0161ichni sna\u017e\u00ed minimalizovat tento odliv. Je\u0161t\u011b l\u00e9pe \u2014 p\u0159edpov\u011bd\u011bt pravd\u011bpodobnost odlivu ur\u010dit\u00e9 kategorie u\u017eivatel\u016f nebo konkr\u00e9tn\u00edho u\u017eivatele a nab\u00eddnout n\u011bjak\u00e9 kroky k jejich udr\u017een\u00ed.<\/p>\n<p>Analyzovat a pokusit se udr\u017eet z\u00e1kazn\u00edka, pokud je to mo\u017en\u00e9, je t\u0159eba alespo\u0148 z n\u00e1sleduj\u00edc\u00edch d\u016fvod\u016f:<\/p>\n<ul>\n<li><b>z\u00edsk\u00e1n\u00ed nov\u00fdch z\u00e1kazn\u00edk\u016f je dra\u017e\u0161\u00ed ne\u017e procesy udr\u017een\u00ed<\/b>. Na z\u00edsk\u00e1n\u00ed nov\u00fdch z\u00e1kazn\u00edk\u016f je obvykle t\u0159eba investovat ur\u010dit\u00e9 pen\u00edze (reklama), zat\u00edmco st\u00e1vaj\u00edc\u00ed z\u00e1kazn\u00edky lze aktivovat speci\u00e1ln\u00ed nab\u00eddkou s mimo\u0159\u00e1dn\u00fdmi podm\u00ednkami;<\/li>\n<li><b>pochopen\u00ed p\u0159\u00ed\u010din odchodu z\u00e1kazn\u00edk\u016f je kl\u00ed\u010dem ke zlep\u0161en\u00ed produkt\u016f a slu\u017eeb<\/b>.<\/li>\n<\/ul>\n<p>\nExistuj\u00ed standardn\u00ed p\u0159\u00edstupy k p\u0159edpov\u011bdi odlivu. Na jednom z mistrovstv\u00ed v AI jsme se rozhodli vyzkou\u0161et pro tento \u00fa\u010del Weibullovo rozd\u011blen\u00ed. Nej\u010dast\u011bji se pou\u017e\u00edv\u00e1 pro anal\u00fdzu p\u0159e\u017e\u00edv\u00e1n\u00ed, p\u0159edpov\u011b\u010f po\u010das\u00ed, anal\u00fdzu p\u0159\u00edrodn\u00edch katastrof, v pr\u016fmyslov\u00e9m in\u017een\u00fdrstv\u00ed a podobn\u011b. Weibullovo rozd\u011blen\u00ed je speci\u00e1ln\u00ed funkce rozd\u011blen\u00ed, parametrizovan\u00e1 dv\u011bma parametry. <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/d30a25becc44269d51957349c6e8b4a0.png\" style=\"display:block;margin: 0 auto;\" \/> a <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/42092030604bf54a98c8379fd532bffe.png\" style=\"display:block;margin: 0 auto;\" \/>. <\/p>\n<p><noindex><a rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/rostelecom\/blog\/444694\/\"><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/62e1933c8335dd16d765f0e358952d84.png\" style=\"display:block;margin: 0 auto;\" \/><\/a><\/noindex><br \/>\n<i>Wikipedia<\/i><\/p>\n<p>Celkov\u011b je to zaj\u00edmav\u00e1 v\u011bc, ale pro p\u0159edpov\u011b\u010f odlivu, a v\u016fbec v fintechu, se to nevyu\u017e\u00edv\u00e1 p\u0159\u00edli\u0161 \u010dasto. V n\u00e1sleduj\u00edc\u00edm textu se pod\u011bl\u00edme o to, jak jsme (Laborato\u0159 inteligentn\u00ed anal\u00fdzy dat) tuto metodu aplikovali a z\u00e1rove\u0148 z\u00edskali zlato na Mistrovstv\u00ed v um\u011bl\u00e9 inteligenci v kategorii \u201eAI v bank\u00e1ch\u201c.<br \/>\n<noindex><a rel=\"nofollow\" name=\"habracut\"><\/a><\/noindex><\/p>\n<h2>O odlivu obecn\u011b<\/h2>\n<p>\nPoj\u010fme si trochu objasnit, co to je odliv z\u00e1kazn\u00edk\u016f a pro\u010d je tak d\u016fle\u017eit\u00fd. Pro podnik\u00e1n\u00ed je d\u016fle\u017eit\u00e1 z\u00e1kaznick\u00e1 z\u00e1kladna. Do t\u00e9to z\u00e1kladny p\u0159ich\u00e1zej\u00ed nov\u00ed z\u00e1kazn\u00edci, nap\u0159\u00edklad kdy\u017e se dozv\u011bd\u00ed o produktu nebo slu\u017eb\u011b prost\u0159ednictv\u00edm reklamy, n\u011bjakou dobu funguj\u00ed (aktivn\u011b vyu\u017e\u00edvaj\u00ed produkty) a po n\u011bjak\u00e9 dob\u011b p\u0159estanou. Tento obdob\u00ed naz\u00fdv\u00e1me \u201e\u017divotn\u00ed cyklus z\u00e1kazn\u00edka\u201c (angl. Customer Lifecycle) \u2014 to je term\u00edn, kter\u00fd popisuje f\u00e1ze, jimi\u017e z\u00e1kazn\u00edk proch\u00e1z\u00ed, kdy\u017e se dozv\u00ed o produktu, rozhoduje o koupi, plat\u00ed, vyu\u017e\u00edv\u00e1 a st\u00e1v\u00e1 se loaj\u00e1ln\u00edm spot\u0159ebitelem, a nakonec p\u0159est\u00e1v\u00e1 pou\u017e\u00edvat z r\u016fzn\u00fdch d\u016fvod\u016f produkty. Odliv tedy p\u0159edstavuje z\u00e1v\u011bre\u010dnou f\u00e1zi \u017eivotn\u00edho cyklu z\u00e1kazn\u00edka, kdy z\u00e1kazn\u00edk p\u0159est\u00e1v\u00e1 vyu\u017e\u00edvat slu\u017eby, co\u017e pro podnik\u00e1n\u00ed znamen\u00e1, \u017ee z\u00e1kazn\u00edk p\u0159estal p\u0159in\u00e1\u0161et zisk a jak\u00fdkoli prosp\u011bch. <\/p>\n<p>Ka\u017ed\u00fd z\u00e1kazn\u00edk banky je konkr\u00e9tn\u00ed osoba, kter\u00e1 si vyb\u00edr\u00e1 tu \u010di onu bankovn\u00ed kartu speci\u00e1ln\u011b pro sv\u00e9 pot\u0159eby. \u010casto cestuje \u2014 karta se m\u00edlemi bude u\u017eite\u010dn\u00e1. Hodn\u011b nakupuje \u2014 ahoj, karta s cashbackem. Hodn\u011b nakupuje v konkr\u00e9tn\u00edch obchodech \u2014 na to u\u017e existuj\u00ed speci\u00e1ln\u00ed partnersk\u00e9 plastov\u00e9 karty. Samoz\u0159ejm\u011b, n\u011bkdy se karta vyb\u00edr\u00e1 tak\u00e9 podle krit\u00e9ria \u201eNejlevn\u011bj\u0161\u00ed slu\u017eby\u201c. V ka\u017ed\u00e9m p\u0159\u00edpad\u011b je tu dost prom\u011bnn\u00fdch.<\/p>\n<p>A je\u0161t\u011b si \u010dlov\u011bk vyb\u00edr\u00e1 s\u00e1m banku \u2014 jak\u00fd smysl m\u00e1 vyb\u00edrat kartu banky, jej\u00edch\u017e pobo\u010dek je pouze v Moskv\u011b a okol\u00ed, kdy\u017e jsi z Chabarovska? A\u0165 je karta takov\u00e9 banky jakkoli v\u00fdhodn\u00e1, p\u0159\u00edtomnost pobo\u010dek banky v bl\u00edzkosti st\u00e1le z\u016fst\u00e1v\u00e1 d\u016fle\u017eit\u00fdm krit\u00e9riem. Ano, 2019 u\u017e je tady a digital \u2014 to je na\u0161e v\u0161echno, ale \u0159adu ot\u00e1zek n\u011bkte\u0159\u00ed banky mohou vy\u0159e\u0161it pouze na pobo\u010dce. Nav\u00edc, op\u011bt \u010d\u00e1st populace d\u016fv\u011b\u0159uje mnohem v\u00edce fyzick\u00e9 bance ne\u017e aplikaci v chytr\u00e9m telefonu, to je tak\u00e9 t\u0159eba br\u00e1t v \u00favahu.<\/p>\n<p>V d\u016fsledku toho m\u016f\u017ee m\u00edt \u010dlov\u011bk mnoho d\u016fvod\u016f k odchodu od produkt\u016f banky (nebo od samotn\u00e9 banky). Zm\u011bnil pr\u00e1ci, a sazba karty se zm\u011bnila z platbn\u00ed na \u201epro oby\u010dejn\u00e9 smrteln\u00edky\u201c, co\u017e je m\u00e9n\u011b v\u00fdhodn\u00e9. P\u0159est\u011bhoval se do jin\u00e9ho m\u011bsta, kde nejsou pobo\u010dky banky. Nel\u00edbilo se mu jedn\u00e1n\u00ed s nekvalifikovan\u00fdm oper\u00e1torem na pobo\u010dce. Tedy d\u016fvod\u016f pro uzav\u0159en\u00ed \u00fa\u010dtu m\u016f\u017ee b\u00fdt daleko v\u00edce, ne\u017e pro pou\u017e\u00edv\u00e1n\u00ed produktu.<\/p>\n<p>Klient m\u016f\u017ee nejen v\u00fdslovn\u011b vyj\u00e1d\u0159it sv\u016fj \u00famysl \u2013 p\u0159ij\u00edt do banky a napsat \u017e\u00e1dost, ale tak\u00e9 jednodu\u0161e p\u0159estat pou\u017e\u00edvat produkty, ani\u017e by zru\u0161il smlouvu. Pr\u00e1v\u011b pro porozum\u011bn\u00ed takov\u00fdm \u00faloh\u00e1m bylo rozhodnuto pou\u017e\u00edt strojov\u00e9 u\u010den\u00ed a AI.<\/p>\n<p>Co v\u00edce, odliv z\u00e1kazn\u00edk\u016f m\u016f\u017ee nastat v jak\u00e9mkoli odv\u011btv\u00ed (telekomunikace, internetov\u00ed poskytovatel\u00e9, poji\u0161\u0165ovac\u00ed spole\u010dnosti, jednodu\u0161e v\u0161ude, kde existuje z\u00e1kaznick\u00e1 z\u00e1kladna a pravideln\u00e9 transakce).<\/p>\n<h2>Co jsme ud\u011blali<\/h2>\n<p>\nNejprve bylo nutn\u00e9 popsat jasnou hranici \u2013 od kter\u00e9ho okam\u017eiku za\u010dneme po\u010d\u00edtat z\u00e1kazn\u00edka jako odch\u00e1zej\u00edc\u00edho. Z pohledu banky, kter\u00e1 n\u00e1m poskytla \u00fadaje k anal\u00fdze, byla aktivita z\u00e1kazn\u00edka bin\u00e1rn\u00ed \u2013 bu\u010f byl aktivn\u00ed, nebo ne. Existovala zna\u010dka ACTIVE_FLAG v tabulce \u201eAktivita\u201c, jej\u00ed\u017e hodnota mohla b\u00fdt bu\u010f \u201e0\u201c, nebo \u201e1\u201c (p\u0159\u00edslu\u0161n\u011b \u201eNeaktivn\u00ed\u201c a \u201eAktivn\u00ed\u201c). A to by bylo v po\u0159\u00e1dku, ale \u010dlov\u011bk je takov\u00fd, \u017ee m\u016f\u017ee n\u011bjakou dobu aktivn\u011b pou\u017e\u00edvat slu\u017ebu a pak na m\u011bs\u00edc vypadnout ze seznamu aktivn\u00edch \u2013 nap\u0159\u00edklad z d\u016fvodu nemoci, odjezdu do zahrani\u010d\u00ed na dovolenou, nebo dokonce vyzkou\u0161en\u00ed karty jin\u00e9 banky. Anebo se po dlouh\u00e9m obdob\u00ed neaktivity znovu za\u010d\u00edt bankovn\u00edmi slu\u017ebami zab\u00fdvat.<\/p>\n<p>Proto jsme se rozhodli naz\u00fdvat obdob\u00edm neaktivity ur\u010dit\u00fd nep\u0159etr\u017eit\u00fd \u010dasov\u00fd \u00fasek, b\u011bhem kter\u00e9ho se pro n\u011bj nastavila zna\u010dka jako \u201e0\u201c. <\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/c4263f11be0f122e54ed7985ba0d2c9d.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nZ\u00e1kazn\u00edci p\u0159ech\u00e1zej\u00ed z neaktivn\u00edch do aktivn\u00edch po r\u016fzn\u011b dlouh\u00fdch obdob\u00edch neaktivity. M\u00e1me mo\u017enost vypo\u010d\u00edtat empirickou veli\u010dinu \u201espolehlivost obdob\u00ed neaktivity\u201c \u2013 tedy pravd\u011bpodobnost, \u017ee \u010dlov\u011bk znovu za\u010dne pou\u017e\u00edvat produkty banky po ur\u010dit\u00e9 do\u010dasn\u00e9 neaktivit\u011b. <\/p>\n<p>Nap\u0159\u00edklad na tomto grafu je zobrazeno obnoven\u00ed aktivity (ACTIVE_FLAG=1) z\u00e1kazn\u00edk\u016f po n\u011bkolika m\u011bs\u00edc\u00edch neaktivity (ACTIVE_FLAG=0).<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/3828c06a86ca6b43e77a6de61f768804.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nZde trochu up\u0159esn\u00edme soubor dat, se kter\u00fdmi jsme za\u010dali pracovat. Banka poskytla agregovan\u00e9 informace za 19 m\u011bs\u00edc\u016f v n\u00e1sleduj\u00edc\u00edch tabulk\u00e1ch:<\/p>\n<ul>\n<li>\u201eAktivita\u201c \u2013 m\u011bs\u00ed\u010dn\u00ed transakce z\u00e1kazn\u00edk\u016f (karty, internetov\u00e9 bankovnictv\u00ed a mobiln\u00ed bankovnictv\u00ed), v\u010detn\u011b v\u00fdplat a informac\u00ed o obratech.<\/li>\n<li>\u201eKarty\u201c \u2013 \u00fadaje o v\u0161ech kart\u00e1ch, kter\u00e9 m\u00e1 klient, s podrobn\u00fdm cen\u00edkem.<\/li>\n<li>\u201eSmlouvy\u201c \u2013 informace o smlouv\u00e1ch klienta (jak otev\u0159en\u00fdch, tak uzav\u0159en\u00fdch): \u00fav\u011bry, vklady a dal\u0161\u00ed, s uveden\u00edm parametr\u016f ka\u017ed\u00e9 z nich.<\/li>\n<li>\u201eKlienti\u201c \u2013 soubor demografick\u00fdch \u00fadaj\u016f (pohlav\u00ed a v\u011bk) a dostupnost kontaktn\u00edch \u00fadaj\u016f.<\/li>\n<\/ul>\n<p>\nPro na\u0161i pr\u00e1ci byly pot\u0159eba v\u0161echny tabulky krom\u011b \u201eKaret\u201c. <\/p>\n<p>Obt\u00ed\u017enost byla tak\u00e9 v tom, \u017ee banka v t\u011bchto datech neuv\u00e1d\u011bla, jak\u00e1 konkr\u00e9tn\u00ed aktivita prob\u00edhala na kart\u00e1ch. To znamen\u00e1, \u017ee jsme mohli zjistit, zda do\u0161lo k transakc\u00edm, ale ur\u010dit jejich typ \u2013 to u\u017e ne. Proto nebylo jasn\u00e9, zda klient vyb\u00edral hotovost, zda mu byla na \u00fa\u010det zasl\u00e1na mzda, nebo zda utr\u00e1cel pen\u00edze na n\u00e1kupy. A nav\u00edc jsme nem\u011bli data o z\u016fstatc\u00edch na \u00fa\u010dtech, co\u017e by bylo u\u017eite\u010dn\u00e9. <\/p>\n<p>Samotn\u00fd v\u00fdb\u011br byl nezkreslen\u00fd \u2013 na tomto \u0159ezu za 19 m\u011bs\u00edc\u016f banka nepodnikla \u017e\u00e1dn\u00e9 pokusy o udr\u017een\u00ed klient\u016f a minimalizaci odtoku.<\/p>\n<p>Te\u010f k obdob\u00edm neaktivity.<\/p>\n<p>Pro formulaci definice odtoku je pot\u0159eba vybrat obdob\u00ed neaktivity. Abychom vytvo\u0159ili progn\u00f3zu odtoku v ur\u010dit\u00e9m \u010dase, <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/d2faf7d1a11c2770987ab6316b1db5e8.png\" style=\"display:block;margin: 0 auto;\" \/>je nutn\u00e9 m\u00edt historii klient\u016f minim\u00e1ln\u011b za 3 m\u011bs\u00edce v intervalu <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/194e5601c2f0f49866d8e8e5a9ff5993.png\" style=\"display:block;margin: 0 auto;\" \/>. Historie byla u n\u00e1s omezena na 19 m\u011bs\u00edc\u016f, proto jsme se rozhodli br\u00e1t obdob\u00ed neaktivity na 6 m\u011bs\u00edc\u016f, pokud existuje. A za minim\u00e1ln\u00ed obdob\u00ed pro kvalitn\u00ed progn\u00f3zu jsme vzali 3 m\u011bs\u00edce. \u010c\u00edsla 3 a 6 m\u011bs\u00edc\u016f jsme vzali empiricky na z\u00e1klad\u011b anal\u00fdzy chov\u00e1n\u00ed t\u011bchto klient\u016f.<\/p>\n<p>Definici odtoku jsme formulovali takto: m\u011bs\u00edc odtoku klienta <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/3d581f5378975105aa6e73cacc175aeb.png\" style=\"display:block;margin: 0 auto;\" \/> je prvn\u00ed m\u011bs\u00edc s ACTIVE_FLAG=0, kde od tohoto m\u011bs\u00edce n\u00e1sleduje alespo\u0148 \u0161est nul v poli ACTIVE_FLAG, jin\u00fdmi slovy, m\u011bs\u00edc, od kter\u00e9ho klient nebyl 6 m\u011bs\u00edc\u016f aktivn\u00ed. <\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/04ea6d4c59a9497733e4a181a4a5847c.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<i>Po\u010det odchoz\u00edch klient\u016f<\/i><\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/4d470cbe91bbc6ee7ddfd1933e90e62c.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<i>Po\u010det zb\u00fdvaj\u00edc\u00edch klient\u016f<\/i><\/p>\n<h2>Jak se obecn\u011b pova\u017euje za odtok<\/h2>\n<p>\nV takov\u00fdch sout\u011b\u017e\u00edch, a obecn\u011b v praxi, se odtok \u010dasto progn\u00f3zuje t\u00edmto zp\u016fsobem. Klient pou\u017e\u00edv\u00e1 produkty a slu\u017eby v r\u016fzn\u00fdch \u010dasov\u00fdch intervalech, data o jeho interakc\u00edch se prezentuj\u00ed jako vektor rys\u016f pevn\u00e9 d\u00e9lky n. Nej\u010dast\u011bji tato informace zahrnuje:<\/p>\n<ul>\n<li>Charakterizuj\u00edc\u00ed u\u017eivatelsk\u00e1 data (demografick\u00e9 \u00fadaje, marketingov\u00fd segment).<\/li>\n<li>Historii pou\u017e\u00edv\u00e1n\u00ed bankovn\u00edch produkt\u016f a slu\u017eeb (toto jsou akce klient\u016f, kter\u00e9 jsou v\u017edy v\u00e1z\u00e1ny na konkr\u00e9tn\u00ed \u010das nebo obdob\u00ed po\u017eadovan\u00e9ho intervalu).<\/li>\n<li>Extern\u00ed data, pokud se je poda\u0159ilo z\u00edskat \u2013 nap\u0159\u00edklad recenze ze soci\u00e1ln\u00edch s\u00edt\u00ed.<\/li>\n<\/ul>\n<p>\nA a\u017e pot\u00e9 se vyvod\u00ed definice odchodu, vlastn\u00ed pro ka\u017edou \u00falohu. Pot\u00e9 se pou\u017eije algoritmus strojov\u00e9ho u\u010den\u00ed, kter\u00fd p\u0159edpov\u00edd\u00e1 pravd\u011bpodobnost odchodu z\u00e1kazn\u00edka. <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/79ce2d7f3193a9aab0b985c9a0cd0ee1.png\" style=\"display:block;margin: 0 auto;\" \/> na z\u00e1klad\u011b vektoru faktor\u016f. <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/546791c513159e41d5950adca001d620.png\" style=\"display:block;margin: 0 auto;\" \/>Pro tr\u00e9nink algoritmu se pou\u017e\u00edv\u00e1 jeden z b\u011b\u017en\u011b zn\u00e1m\u00fdch r\u00e1mc\u016f pro vytv\u00e1\u0159en\u00ed ensemble rozhodovac\u00edch strom\u016f, <noindex><a rel=\"nofollow\" href=\"https:\/\/xgboost.readthedocs.io\/en\/latest\/\">cuML<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/Microsoft\/LightGBM\">LightGBM.<\/a><\/noindex>, <noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/catboost\/catboost\">CatBoost.<\/a><\/noindex> nebo jejich modifikace. <\/p>\n<p>Algoritmus nen\u00ed \u0161patn\u00fd, ale v oblasti predikce odchodu m\u00e1 n\u011bkolik v\u00e1\u017en\u00fdch nedostatk\u016f.<\/p>\n<ul>\n<li><b>Nem\u00e1 takzvanou \u201epam\u011b\u0165\u201c.<\/b>Do modelu vstupuje stanoven\u00fd po\u010det znak\u016f, kter\u00e9 odpov\u00eddaj\u00ed aktu\u00e1ln\u00edmu okam\u017eiku. Aby bylo mo\u017en\u00e9 zahrnout informace o historii zm\u011bny parametr\u016f, je nutn\u00e9 vypo\u010d\u00edtat speci\u00e1ln\u00ed znaky, kter\u00e9 charakterizuj\u00ed zm\u011bny parametr\u016f v \u010dase, nap\u0159\u00edklad po\u010det nebo sou\u010det bankovn\u00edch transakc\u00ed za posledn\u00ed 1, 2, 3 m\u011bs\u00edce. Tento p\u0159\u00edstup m\u016f\u017ee pouze \u010d\u00e1ste\u010dn\u011b odr\u00e1\u017eet povahu \u010dasov\u00fdch zm\u011bn. <\/li>\n<li><b>Pevn\u011b stanoven\u00fd horizont predikce.<\/b> Model dok\u00e1\u017ee p\u0159edpov\u00eddat odchod z\u00e1kazn\u00edk\u016f pouze na p\u0159edem stanoven\u00e9 \u010dasov\u00e9 obdob\u00ed, nap\u0159\u00edklad predikce na jeden m\u011bs\u00edc dop\u0159edu. Pokud je t\u0159eba p\u0159edpov\u011bd\u011bt na jin\u00e9 \u010dasov\u00e9 obdob\u00ed, nap\u0159\u00edklad na t\u0159i m\u011bs\u00edce, je nutn\u00e9 restrukturalizovat tr\u00e9ninkov\u00fd vzorek a znovu nau\u010dit nov\u00fd model. <\/li>\n<\/ul>\n<p><\/p>\n<h2>N\u00e1\u0161 p\u0159\u00edstup.<\/h2>\n<p>\nRozhodli jsme se hned na za\u010d\u00e1tku, \u017ee nebudeme pou\u017e\u00edvat standardn\u00ed p\u0159\u00edstupy. V sout\u011b\u017ei se krom\u011b n\u00e1s zaregistrovalo dal\u0161\u00edch 497 lid\u00ed, z nich\u017e ka\u017ed\u00fd m\u011bl zna\u010dn\u00e9 zku\u0161enosti. Tak\u017ee sna\u017eit se d\u011blat n\u011bco podle standardn\u00edho sch\u00e9matu za takov\u00fdch podm\u00ednek \u2013 nen\u00ed nejlep\u0161\u00ed n\u00e1pad.<\/p>\n<p>A za\u010dali jsme \u0159e\u0161it probl\u00e9my, kter\u00e9 st\u00e1li p\u0159ed modelem bin\u00e1rn\u00ed klasifikace, pomoc\u00ed predikce pravd\u011bpodobnostn\u00edho rozd\u011blen\u00ed \u010dasu odchodu z\u00e1kazn\u00edk\u016f. Takov\u00fd p\u0159\u00edstup si m\u016f\u017eete prohl\u00e9dnout, <noindex><a rel=\"nofollow\" href=\"https:\/\/ragulpr.github.io\/assets\/draft_master_thesis_martinsson_egil_wtte_rnn_2016.pdf\">zde<\/a><\/noindex>umo\u017e\u0148uje flexibiln\u011bji p\u0159edpov\u00eddat odchod a ov\u011b\u0159ovat slo\u017eit\u011bj\u0161\u00ed hypot\u00e9zy ne\u017e v klasick\u00e9m p\u0159\u00edstupu. Jako rodinu rozd\u011blen\u00ed modeluj\u00edc\u00edch \u010das odchodu jsme zvolili rozd\u011blen\u00ed <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%A0%D0%B0%D1%81%D0%BF%D1%80%D0%B5%D0%B4%D0%B5%D0%BB%D0%B5%D0%BD%D0%B8%D0%B5_%D0%92%D0%B5%D0%B9%D0%B1%D1%83%D0%BB%D0%BB%D0%B0\">Weibull,<\/a><\/noindex> pro jeho \u0161irok\u00e9 vyu\u017eit\u00ed v anal\u00fdze p\u0159e\u017eit\u00ed. Chov\u00e1n\u00ed z\u00e1kazn\u00edka lze pova\u017eovat za jak\u00fdsi druh p\u0159e\u017eit\u00ed. <\/p>\n<p>Zde jsou p\u0159\u00edklady rozd\u011blen\u00ed hustot pravd\u011bpodobnosti Weibulla v z\u00e1vislosti na parametrech <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/57c935da52faf825ceedb4389d21c354.png\" style=\"display:block;margin: 0 auto;\" \/> a <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/85d1c92d5b9dfa83d13092edefb29a09.png\" style=\"display:block;margin: 0 auto;\" \/>:<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/371e51ab8df2c10d5b139aadf290ebf8.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nToto je hustota rozd\u011blen\u00ed pravd\u011bpodobnosti odchodu z\u00e1kazn\u00edka t\u0159\u00ed r\u016fzn\u00fdch klient\u016f v pr\u016fb\u011bhu \u010dasu. \u010cas je zn\u00e1zorn\u011bn v m\u011bs\u00edc\u00edch. Jin\u00fdmi slovy, tento graf ukazuje, kdy je nejpravd\u011bpodobn\u011bj\u0161\u00ed, \u017ee dojde k odchodu z\u00e1kazn\u00edka v n\u00e1sleduj\u00edc\u00edch dvou m\u011bs\u00edc\u00edch. Jak je vid\u011bt, klient s rozd\u011blen\u00edm m\u00e1 v\u011bt\u0161\u00ed potenci\u00e1l odej\u00edt d\u0159\u00edve ne\u017e klienti s rozd\u011blen\u00edmi Weibull(2, 0.5) a Weibull(3, 1). <\/p>\n<p>V\u00fdsledkem je model, kter\u00fd pro ka\u017ed\u00e9ho klienta v jak\u00e9mkoli<br \/>\nm\u011bs\u00edci predikuje parametry rozd\u011blen\u00ed Weibulla, kter\u00e9 nejl\u00e9pe odr\u00e1\u017e\u00ed pravd\u011bpodobnost odchodu v pr\u016fb\u011bhu \u010dasu. Pokud podrobn\u011bji: <\/p>\n<ul>\n<li>C\u00edlov\u00e9 znaky na tr\u00e9novac\u00ed vzorku - \u010das zb\u00fdvaj\u00edc\u00ed do odchodu v konkr\u00e9tn\u00edm m\u011bs\u00edci pro konkr\u00e9tn\u00edho klienta.<\/li>\n<li>Pokud nem\u00e1me \u00fadaj o odchodu klienta, p\u0159edpokl\u00e1d\u00e1me, \u017ee \u010das odchodu je del\u0161\u00ed ne\u017e po\u010det m\u011bs\u00edc\u016f, po\u010d\u00ednaje aktu\u00e1ln\u00edm m\u011bs\u00edce a a\u017e do konce dostupn\u00fdch historick\u00fdch dat.<\/li>\n<li>Pou\u017eit\u00fd model: rekurentn\u00ed neuronov\u00e1 s\u00ed\u0165 s LSTM vrstvou.<\/li>\n<li>Jako ztr\u00e1tovou funkci pou\u017e\u00edv\u00e1me z\u00e1pornou logaritmickou funkci pravd\u011bpodobnosti pro rozd\u011blen\u00ed Weibulla.<\/li>\n<\/ul>\n<p>\nZde jsou v\u00fdhody t\u00e9to metody: <\/p>\n<ul>\n<li>Pravd\u011bpodobnostn\u00ed rozd\u011blen\u00ed, krom\u011b z\u0159ejm\u00e9 mo\u017enosti bin\u00e1rn\u00ed klasifikace, umo\u017e\u0148uje flexibiln\u011b predikovat r\u016fzn\u00e9 ud\u00e1losti, nap\u0159\u00edklad zda klient p\u0159estane vyu\u017e\u00edvat slu\u017eeb banky b\u011bhem 3 m\u011bs\u00edc\u016f. Pokud je pot\u0159eba, na z\u00e1klad\u011b tohoto rozd\u011blen\u00ed lze tak\u00e9 pr\u016fm\u011brovat r\u016fzn\u00e9 metriky.<\/li>\n<li>Rekurentn\u00ed neuronov\u00e1 s\u00ed\u0165 LSTM m\u00e1 pam\u011b\u0165 a efektivn\u011b vyu\u017e\u00edv\u00e1 celou dostupnou historii. S roz\u0161\u00ed\u0159en\u00edm nebo up\u0159esn\u011bn\u00edm historie se p\u0159esnost zvy\u0161uje.<\/li>\n<li>P\u0159\u00edstup lze snadno \u0161k\u00e1lovat p\u0159i rozd\u011blen\u00ed \u010dasov\u00fdch \u00fasek\u016f na men\u0161\u00ed (nap\u0159\u00edklad p\u0159i rozd\u011blen\u00ed m\u011bs\u00edc\u016f na t\u00fddny).<\/li>\n<\/ul>\n<p>\nAle nesta\u010d\u00ed vytvo\u0159it dobr\u00fd model, je tak\u00e9 nutn\u00e9 spr\u00e1vn\u011b zhodnotit jeho kvalitu.<\/p>\n<h2>Jak jsme hodnotili kvalitu<\/h2>\n<p>\nJako metrikou jsme zvolili Lift Curve. Pou\u017e\u00edv\u00e1 se v obchod\u011b pro podobn\u00e9 p\u0159\u00edpady d\u00edky jasn\u00e9 interpretaci, je dob\u0159e pops\u00e1na <noindex><a rel=\"nofollow\" href=\"http:\/\/www2.cs.uregina.ca\/~dbd\/cs831\/notes\/lift_chart\/lift_chart.html\">tady<\/a><\/noindex> a <noindex><a rel=\"nofollow\" href=\"https:\/\/www.information-management.com\/news\/why-lift\">zde<\/a><\/noindex>. Pokud bychom popsali smysl t\u00e9to metriky jednou v\u011btou, byla by to \u201eO kolik l\u00e9pe algoritmus p\u0159edpov\u00edd\u00e1 v prvn\u00edch <img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/72818f84f73c2838fe7d63b33a96fd7b.png\" style=\"display:block;margin: 0 auto;\" \/>% oproti n\u00e1hodn\u00e9mu v\u00fdb\u011bru\u201c. <\/p>\n<h2>\u0160kol\u00edme modely<\/h2>\n<p>\nPodm\u00ednky sout\u011b\u017ee neuv\u00e1d\u011bly konkr\u00e9tn\u00ed metriky kvality, podle kter\u00fdch by bylo mo\u017eno porovn\u00e1vat r\u016fzn\u00e9 modely a p\u0159\u00edstupy. Nav\u00edc definice pojmu odtoku m\u016f\u017ee b\u00fdt r\u016fzn\u00e1 a m\u016f\u017ee z\u00e1viset na formulaci \u00fakolu, kter\u00e1 je zase ur\u010dena obchodn\u00edmi c\u00edli. Proto, abychom pochopili, kter\u00fd zp\u016fsob je lep\u0161\u00ed, jsme vytr\u00e9novali dva modely:<\/p>\n<ol>\n<li>\u010casto pou\u017e\u00edvan\u00fd p\u0159\u00edstup bin\u00e1rn\u00ed klasifikace s vyu\u017eit\u00edm algoritmu strojov\u00e9ho u\u010den\u00ed na z\u00e1klad\u011b rozhodovac\u00edch strom\u016f (<noindex><a rel=\"nofollow\" href=\"https:\/\/github.com\/Microsoft\/LightGBM\">LightGBM.<\/a><\/noindex>);<\/li>\n<li> Model Weibull-LSTM<\/li>\n<\/ol>\n<p>\nTestovac\u00ed vzorek sest\u00e1val z 500 p\u0159edem vybran\u00fdch klient\u016f, kte\u0159\u00ed nebyli ve tr\u00e9ninkov\u00e9m vzorku. Pro model se prov\u00e1d\u011bl v\u00fdb\u011br hyperparametr\u016f pomoc\u00ed k\u0159\u00ed\u017eov\u00e9 validace rozd\u011blen\u00e9 podle klient\u016f. Pro tr\u00e9nink ka\u017ed\u00e9ho modelu byly pou\u017eity stejn\u00e9 sady znak\u016f. <\/p>\n<p>Vzhledem k tomu, \u017ee model nem\u00e1 pam\u011b\u0165, byly pro n\u011bj br\u00e1ny speci\u00e1ln\u00ed znaky, kter\u00e9 ukazuj\u00ed na pom\u011br zm\u011bn parametr\u016f jednoho m\u011bs\u00edce k pr\u016fm\u011brn\u00e9mu hodnocen\u00ed za posledn\u00ed t\u0159i m\u011bs\u00edce. To charakterizovalo rychlost zm\u011bny hodnot za posledn\u00ed t\u0159i m\u011bs\u00edce. Bez toho by byl model na b\u00e1zi Random Forest v p\u0159edem nev\u00fdhodn\u00e9 pozici v\u016f\u010di Weibull-LSTM.<\/p>\n<h2>Jak je LSTM s Weibullov\u00fdm rozd\u011blen\u00edm lep\u0161\u00ed ne\u017e p\u0159\u00edstup zalo\u017een\u00fd na rozhodovac\u00edch stromech <\/h2>\n<p>\nZde je v\u0161e n\u00e1zorn\u00e9 doslova p\u00e1r obr\u00e1zky.<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/0a87a27804f1b285f66a889ed798cf6b.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<i>Srovn\u00e1n\u00ed Lift Curve pro klasick\u00fd algoritmus a Weibull-LSTM<\/i><\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/6c281677260098c7993a690c923692c8.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<i>Srovn\u00e1n\u00ed metriky Lift Curve podle m\u011bs\u00edc\u016f pro klasick\u00fd algoritmus a Weibull-LSTM<\/i><\/p>\n<p>Obecn\u011b LSTM p\u0159ekon\u00e1v\u00e1 klasick\u00fd algoritmus t\u00e9m\u011b\u0159 ve v\u0161ech p\u0159\u00edpadech.<\/p>\n<h2>Predikce odtoku<\/h2>\n<p>\nModel na z\u00e1klad\u011b rekurentn\u00ed neuronov\u00e9 s\u00edt\u011b s LSTM bu\u0148kami a Weibullov\u00fdm rozd\u011blen\u00edm m\u016f\u017ee p\u0159edv\u00eddat odtok p\u0159edem, nap\u0159\u00edklad predikovat odchod klienta b\u011bhem n\u00e1sleduj\u00edc\u00edch n m\u011bs\u00edc\u016f. Zva\u017eme p\u0159\u00edpad pro n = 3. V tomto p\u0159\u00edpad\u011b mus\u00ed neuronov\u00e1 s\u00ed\u0165 spr\u00e1vn\u011b ur\u010dit: odejde-li klient od p\u0159\u00ed\u0161t\u00edho m\u011bs\u00edce a\u017e do n-t\u00e9ho m\u011bs\u00edce. Jin\u00fdmi slovy, mus\u00ed spr\u00e1vn\u011b ur\u010dit, zda klient z\u016fstane po uplynut\u00ed n m\u011bs\u00edc\u016f. To lze pova\u017eovat za p\u0159edpov\u011b\u010f dop\u0159edu: predikci okam\u017eiku, kdy klient pr\u00e1v\u011b za\u010dal uva\u017eovat o odchodu. <\/p>\n<p>Porovnejme Lift Curve pro Weibull-LSTM na 1, 2 a 3 m\u011bs\u00edce p\u0159ed odtokem:<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/4b2ed9cb21844cf5dc8f54bcfe092a9d.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nJak jsme ji\u017e zm\u00ednili v\u00fd\u0161e, d\u016fle\u017eit\u00e9 jsou tak\u00e9 ty p\u0159edpov\u011bdi, kter\u00e9 byly provedeny pro klienty, kte\u0159\u00ed jsou na n\u011bjakou dobu neaktivn\u00ed. Proto tu p\u0159id\u00e1me do v\u00fdb\u011bru takov\u00e9 p\u0159\u00edpady, kdy byl odchoz\u00ed klient neaktivn\u00ed jeden nebo dva m\u011bs\u00edce, a zkontrolujeme, zda Weibull-LSTM spr\u00e1vn\u011b klasifikuje tyto p\u0159\u00edpady jako odliv. Vzhledem k tomu, \u017ee takov\u00e9 p\u0159\u00edpady byly ve v\u00fdb\u011bru, o\u010dek\u00e1v\u00e1me, \u017ee si s\u00ed\u0165 s nimi porad\u00ed dob\u0159e:<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/4fe6e918aff2f16758c671c8d547895d.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h2>Udr\u017een\u00ed klient\u016f<\/h2>\n<p>\nVe skute\u010dnosti, to je hlavn\u00ed, co lze ud\u011blat, kdy\u017e m\u00e1te informace o tom, \u017ee tito klienti se chystaj\u00ed p\u0159estat pou\u017e\u00edvat produkt. Pokud hovo\u0159\u00edme o budov\u00e1n\u00ed modelu, kter\u00fd by mohl nab\u00edzet n\u011bco u\u017eite\u010dn\u00e9ho klient\u016fm, abychom je udr\u017eeli \u2014 toho nebude mo\u017en\u00e9 dos\u00e1hnout, pokud nem\u00e1te historick\u00e9 pokusy, kter\u00e9 by skon\u010dily dob\u0159e.<\/p>\n<p>Takovou historii jsme nem\u011bli, a proto jsme se rozhodli takto.<\/p>\n<ol>\n<li>Vytv\u00e1\u0159\u00edme model, kter\u00fd ur\u010duje zaj\u00edmav\u00e9 produkty pro ka\u017ed\u00e9ho klienta.<\/li>\n<li>Ka\u017ed\u00fd m\u011bs\u00edc spou\u0161t\u00edme klasifik\u00e1tor a ur\u010d\u00edme potenci\u00e1ln\u011b odch\u00e1zej\u00edc\u00ed klienty.<\/li>\n<li>\u010c\u00e1stem klient\u016f nab\u00edz\u00edme produkt podle modelu z bodu 1, zapamatujeme si na\u0161e akce.<\/li>\n<li>Po n\u011bkolika m\u011bs\u00edc\u00edch se pod\u00edv\u00e1me, kdo z t\u011bchto potenci\u00e1ln\u011b odch\u00e1zej\u00edc\u00edch klient\u016f ode\u0161el a kdo z\u016fstal. T\u00edmto zp\u016fsobem vytv\u00e1\u0159\u00edme tr\u00e9ninkovou sadu.<\/li>\n<li>Tr\u00e9nujeme model na historii z\u00edskan\u00e9 v bod\u011b 4.<\/li>\n<li>Voliteln\u011b opakujeme postup a nahrazujeme model z bodu 1 modelem z\u00edskan\u00fdm v bod\u011b 5.<\/li>\n<\/ol>\n<p>\nKontrolou kvality takov\u00e9ho udr\u017een\u00ed m\u016f\u017ee b\u00fdt b\u011b\u017en\u00e9 A\/B testov\u00e1n\u00ed \u2014 rozd\u011bl\u00edme klienty, kte\u0159\u00ed potenci\u00e1ln\u011b odch\u00e1zej\u00ed, do dvou skupin. Jedn\u00e9 nab\u00edz\u00edme produkty na z\u00e1klad\u011b na\u0161eho modelu udr\u017een\u00ed, druh\u00e9 nenab\u00edz\u00edme nic. Rozhodli jsme se vytr\u00e9novat model, kter\u00fd by mohl p\u0159in\u00e9st u\u017eitek ji\u017e v bod\u011b 1 na\u0161eho p\u0159\u00edkladu.<\/p>\n<p>Cht\u011bli jsme segmentaci ud\u011blat co nejv\u00edce interpretovatelnou. K tomu jsme vybrali n\u011bkolik znak\u016f, kter\u00e9 by mohly b\u00fdt snadno interpretovateln\u00e9: celkov\u00fd po\u010det transakc\u00ed, mzda, celkov\u00fd obrat na \u00fa\u010dtu, v\u011bk, pohlav\u00ed. Znaky z tabulky \u201eKarty\u201c nebyly br\u00e1ny v \u00favahu jako m\u00e1lo informativn\u00ed, a znaky z tabulky 3 \u201eSmlouvy\u201c byly vypu\u0161t\u011bny kv\u016fli slo\u017eitosti zpracov\u00e1n\u00ed, aby se p\u0159ede\u0161lo \u00faniku \u00fadaj\u016f mezi valida\u010dn\u00edm a tr\u00e9ninkov\u00fdm souborem.<\/p>\n<p>Kl\u00e1strov\u00e1n\u00ed bylo provedeno pomoc\u00ed Gaussovsk\u00fdch sm\u011bsn\u00fdch model\u016f. Krit\u00e9rium Akaike umo\u017enilo ur\u010dit 2 optim\u00e1ln\u00ed hodnoty. Prvn\u00ed optim\u00e1ln\u00ed hodnota odpov\u00edd\u00e1 1 klastru. Druh\u00e1, m\u00e9n\u011b v\u00fdrazn\u00e1 optim\u00e1ln\u00ed hodnota odpov\u00edd\u00e1 80 klastr\u016fm. Na z\u00e1klad\u011b tohoto v\u00fdsledku lze u\u010dinit n\u00e1sleduj\u00edc\u00ed z\u00e1v\u011br: data je velmi obt\u00ed\u017en\u00e9 rozd\u011blit na klastry bez apriorn\u00ed informace. Pro kvalitn\u011bj\u0161\u00ed kl\u00e1strov\u00e1n\u00ed jsou zapot\u0159eb\u00ed data, kter\u00e1 detailn\u011b popisuj\u00ed ka\u017ed\u00e9ho klienta.<\/p>\n<p>Proto byla zv\u00e1\u017eena \u00faloha u\u010den\u00ed s u\u010ditelem, aby bylo mo\u017en\u00e9 nab\u00eddnout ka\u017ed\u00e9mu jednotliv\u00e9mu klientovi jeho produkt. Byly zva\u017eov\u00e1ny n\u00e1sleduj\u00edc\u00ed produkty: \u201eTerm\u00ednovan\u00fd vklad\u201c, \u201eKreditn\u00ed karta\u201c, \u201eOvdrafter\u201c, \u201eSpot\u0159ebitelsk\u00fd \u00fav\u011br\u201c, \u201eAuto\u00fav\u011br\u201c, \u201eHypot\u00e9ka\u201c. <\/p>\n<p>V datech byl p\u0159\u00edtomen je\u0161t\u011b jeden typ produktu: \u201eB\u011b\u017en\u00fd \u00fa\u010det\u201c. Ale ten jsme nezva\u017eovali kv\u016fli n\u00edzk\u00e9 informa\u010dn\u00ed hodnot\u011b. Byla sestavena model zalo\u017een\u00fd na u\u017eivatel\u00edch, kte\u0159\u00ed jsou klienty banky, t.j. nep\u0159estali pou\u017e\u00edvat jej\u00ed produkty, kter\u00e9 p\u0159edpov\u00eddaj\u00ed, jak\u00fd produkt by je mohl zaj\u00edmat. Jako model byla vybr\u00e1na logistick\u00e1 regrese a jako metrika kvality byla pou\u017eita hodnota Lift pro prvn\u00edch 10 percentil\u016f. <\/p>\n<p>Kvalitu modelu lze posoudit na obr\u00e1zku.<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/77cab18988ab9c3b768b86e31479817f.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<i>V\u00fdsledky doporu\u010dovac\u00edho modelu produkt\u016f pro klienty <\/i> <\/p>\n<h2>V\u00fdsledek<\/h2>\n<p>\nTento p\u0159\u00edstup n\u00e1m p\u0159inesl prvn\u00ed m\u00edsto v kategorii \u201eAI v bank\u00e1ch\u201c na \u0160ampion\u00e1tu AI RAIF-Challenge 2017.<\/p>\n<p><img decoding=\"async\" alt=\"Jak jsme predikovali odliv, p\u0159istoupili jsme k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed\" src=\"\/wp-content\/uploads\/2019\/03\/0c321869dfd02323766c025ed4021bcf.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nZ\u0159ejm\u011b bylo hlavn\u00edm c\u00edlem p\u0159istoupit k probl\u00e9mu z m\u00e9n\u011b obvykl\u00e9ho \u00fahlu a pou\u017e\u00edt metodu, kter\u00e1 se b\u011b\u017en\u011b pou\u017e\u00edv\u00e1 v jin\u00fdch situac\u00edch.<\/p>\n<p>I kdy\u017e masivn\u00ed odliv u\u017eivatel\u016f m\u016f\u017ee b\u00fdt pro slu\u017eby \u017eivelnou pohromou.<\/p>\n<p>Tuto metodu lze vz\u00edt na v\u011bdom\u00ed a pro jakoukoli jinou oblast, kde je d\u016fle\u017eit\u00e9 zohlednit odliv, nejen u bank. Nap\u0159\u00edklad jsme ji pou\u017eili i k v\u00fdpo\u010dtu vlastn\u00edho odlivu \u2014 v sibi\u0159sk\u00e9m a petrohradsk\u00e9m filiale Rostelekomu.<\/p>\n<p><i>\u201eLaborato\u0159 inteligentn\u00ed anal\u00fdzy dat\u201c spole\u010dnost \u201eVyhled\u00e1vac\u00ed port\u00e1l \u201eSputnik\u201c<\/i><br \/>\n<br \/>Zdroj: <a content=\"nofollow\" rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/rostelecom\/blog\/444694\/\">habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"<p>\u0418\u043d\u043e\u0433\u0434\u0430 \u0434\u043b\u044f \u0442\u043e\u0433\u043e, \u0447\u0442\u043e\u0431\u044b \u0440\u0435\u0448\u0438\u0442\u044c \u043a\u0430\u043a\u0443\u044e-\u0442\u043e \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u0443, \u043d\u0430\u0434\u043e \u043f\u0440\u043e\u0441\u0442\u043e \u0432\u0437\u0433\u043b\u044f\u043d\u0443\u0442\u044c \u043d\u0430 \u043d\u0435\u0435 \u043f\u043e\u0434 \u0434\u0440\u0443\u0433\u0438\u043c \u0443\u0433\u043b\u043e\u043c. \u0414\u0430\u0436\u0435 \u0435\u0441\u043b\u0438 \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0438\u0435 \u043b\u0435\u0442 10 \u043f\u043e\u0434\u043e\u0431\u043d\u044b\u0435 \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u044b \u0440\u0435\u0448\u0430\u043b\u0438 \u043e\u0434\u043d\u0438\u043c \u0438 \u0442\u0435\u043c \u0436\u0435 \u0441\u043f\u043e\u0441\u043e\u0431\u043e\u043c \u0441 \u0440\u0430\u0437\u043d\u044b\u043c \u044d\u0444\u0444\u0435\u043a\u0442\u043e\u043c, \u043d\u0435 \u0444\u0430\u043a\u0442, \u0447\u0442\u043e \u044d\u0442\u043e\u0442 \u0441\u043f\u043e\u0441\u043e\u0431 \u0435\u0434\u0438\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0439. \u0415\u0441\u0442\u044c \u0442\u0430\u043a\u0430\u044f \u0442\u0435\u043c\u0430, \u043a\u0430\u043a \u043e\u0442\u0442\u043e\u043a \u043a\u043b\u0438\u0435\u043d\u0442\u043e\u0432. \u0428\u0442\u0443\u043a\u0430 \u043d\u0435\u0438\u0437\u0431\u0435\u0436\u043d\u0430\u044f, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e \u043a\u043b\u0438\u0435\u043d\u0442\u044b \u043b\u044e\u0431\u043e\u0439 \u043a\u043e\u043c\u043f\u0430\u043d\u0438\u0438 \u043c\u043e\u0433\u0443\u0442 \u043f\u043e \u043c\u043d\u043e\u0436\u0435\u0441\u0442\u0432\u0443 \u043f\u0440\u0438\u0447\u0438\u043d \u0432\u0437\u044f\u0442\u044c [&hellip;]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-30044","post","type-post","status-publish","format-standard","hentry"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 5.0.3 - aioseo.com -->\n\t<meta name=\"description\" content=\"\u0418\u043d\u043e\u0433\u0434\u0430 \u0434\u043b\u044f \u0442\u043e\u0433\u043e, \u0447\u0442\u043e\u0431\u044b \u0440\u0435\u0448\u0438\u0442\u044c \u043a\u0430\u043a\u0443\u044e-\u0442\u043e \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u0443, \u043d\u0430\u0434\u043e \u043f\u0440\u043e\u0441\u0442\u043e \u0432\u0437\u0433\u043b\u044f\u043d\u0443\u0442\u044c \u043d\u0430 \u043d\u0435\u0435 \u043f\u043e\u0434 \u0434\u0440\u0443\u0433\u0438\u043c \u0443\u0433\u043b\u043e\u043c.\" \/>\n\t<meta name=\"robots\" content=\"max-image-preview:large\" \/>\n\t<meta name=\"author\" content=\"Yuri Gagarin\"\/>\n\t<link rel=\"canonical\" href=\"https:\/\/prohoster.info\/cs\/blog\/kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu\" \/>\n\t\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.3\" \/>\n\t\t<meta property=\"og:locale\" content=\"cs_CZ\" \/>\n\t\t<meta property=\"og:site_name\" content=\"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b\" \/>\n\t\t<meta property=\"og:type\" content=\"article\" \/>\n\t\t<meta property=\"og:title\" content=\"\ud83e\udd47\u041a\u0430\u043a \u043c\u044b \u043f\u0440\u0435\u0434\u0441\u043a\u0430\u0437\u044b\u0432\u0430\u043b\u0438 \u043e\u0442\u0442\u043e\u043a, \u043f\u043e\u0434\u043e\u0439\u0434\u044f \u043a \u043d\u0435\u043c\u0443 \u043a\u0430\u043a \u043a \u0441\u0442\u0438\u0445\u0438\u0439\u043d\u043e\u043c\u0443 \u0431\u0435\u0434\u0441\u0442\u0432\u0438\u044e | ProHoster\" \/>\n\t\t<meta property=\"og:description\" content=\"\u0418\u043d\u043e\u0433\u0434\u0430 \u0434\u043b\u044f \u0442\u043e\u0433\u043e, \u0447\u0442\u043e\u0431\u044b \u0440\u0435\u0448\u0438\u0442\u044c \u043a\u0430\u043a\u0443\u044e-\u0442\u043e \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u0443, \u043d\u0430\u0434\u043e \u043f\u0440\u043e\u0441\u0442\u043e \u0432\u0437\u0433\u043b\u044f\u043d\u0443\u0442\u044c \u043d\u0430 \u043d\u0435\u0435 \u043f\u043e\u0434 \u0434\u0440\u0443\u0433\u0438\u043c \u0443\u0433\u043b\u043e\u043c.\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/prohoster.info\/cs\/blog\/kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu\" \/>\n\t\t<meta property=\"og:image\" content=\"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg\" \/>\n\t\t<meta property=\"og:image:secure_url\" content=\"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg\" \/>\n\t\t<meta property=\"og:image:width\" content=\"350\" \/>\n\t\t<meta property=\"og:image:height\" content=\"350\" \/>\n\t\t<meta property=\"article:published_time\" content=\"2019-10-31T18:33:22+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2019-10-31T18:33:22+00:00\" \/>\n\t\t<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/prohoster\" \/>\n\t\t<meta property=\"article:author\" content=\"https:\/\/www.facebook.com\/prohoster\" \/>\n\t\t<!-- All in One SEO -->\n\n","aioseo_head_json":{"title":"\ud83e\udd47Jak jsme p\u0159edpov\u011bd\u011bli odliv, p\u0159istoupiv\u0161e k n\u011bmu jako k \u017eiveln\u00e9mu ne\u0161t\u011bst\u00ed | ProHoster","description":"N\u011bkdy sta\u010d\u00ed, abychom se na probl\u00e9m pod\u00edvali z jin\u00e9ho \u00fahlu, abychom ho vy\u0159e\u0161ili.","canonical_url":"https:\/\/prohoster.info\/cs\/blog\/kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"cs_CZ","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47\u041a\u0430\u043a \u043c\u044b \u043f\u0440\u0435\u0434\u0441\u043a\u0430\u0437\u044b\u0432\u0430\u043b\u0438 \u043e\u0442\u0442\u043e\u043a, \u043f\u043e\u0434\u043e\u0439\u0434\u044f \u043a \u043d\u0435\u043c\u0443 \u043a\u0430\u043a \u043a \u0441\u0442\u0438\u0445\u0438\u0439\u043d\u043e\u043c\u0443 \u0431\u0435\u0434\u0441\u0442\u0432\u0438\u044e | ProHoster","og:description":"\u0418\u043d\u043e\u0433\u0434\u0430 \u0434\u043b\u044f \u0442\u043e\u0433\u043e, \u0447\u0442\u043e\u0431\u044b \u0440\u0435\u0448\u0438\u0442\u044c \u043a\u0430\u043a\u0443\u044e-\u0442\u043e \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u0443, \u043d\u0430\u0434\u043e \u043f\u0440\u043e\u0441\u0442\u043e \u0432\u0437\u0433\u043b\u044f\u043d\u0443\u0442\u044c \u043d\u0430 \u043d\u0435\u0435 \u043f\u043e\u0434 \u0434\u0440\u0443\u0433\u0438\u043c \u0443\u0433\u043b\u043e\u043c.","og:url":"https:\/\/prohoster.info\/cs\/blog\/kak-my-predskazyvali-ottok-podojdya-k-nemu-kak-k-stihijnomu-bedstviyu","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2019-10-31T18:33:22+00:00","article:modified_time":"2019-10-31T18:33:22+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"30044","title":null,"description":null,"keywords":null,"keyphrases":null,"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":null,"og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"Article","isEnabled":true},"graphs":[]},"schema_type":null,"schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":null,"robots_max_videopreview":null,"robots_max_imagepreview":"large","priority":null,"frequency":null,"local_seo":null,"seo_analyzer_scan_date":"2026-01-20 23:33:19","breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-03-01 03:43:15","updated":"2026-01-20 23:33:19","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/30044","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/comments?post=30044"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/30044\/revisions"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media?parent=30044"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/categories?post=30044"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/tags?post=30044"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}