{"id":53143,"date":"2019-11-24T00:00:00","date_gmt":"2019-11-23T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru"},"modified":"2020-02-18T14:01:00","modified_gmt":"2020-02-18T11:01:00","slug":"ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/administrirovanie\/ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru","title":{"rendered":"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/ee5728dccb94b849ad3bfd9cf84ad74d.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\n<i>Pasqyra e fjalimeve t\u00eb mia n\u00eb Highload++ dhe DataFest Minsk 2019.<\/i><\/p>\n<p>P\u00ebr shum\u00eb njer\u00ebz, sot posta \u00ebsht\u00eb nj\u00eb pjes\u00eb e pandashme e jet\u00ebs n\u00eb internet. Me t\u00eb, ne kryejm\u00eb korrespondenc\u00ebn biznesore, ruajm\u00eb informacione t\u00eb ndryshme t\u00eb r\u00ebnd\u00ebsishme lidhur me financat, rezervimin e hoteleve, p\u00ebrpunimin e porosive dhe shum\u00eb t\u00eb tjera. N\u00eb mes t\u00eb vitit 2018, ne formuluam strategjin\u00eb e produktit p\u00ebr zhvillimin e post\u00ebs. Si duhet t\u00eb jet\u00eb posta moderne?<\/p>\n<p>Posta duhet t\u00eb jet\u00eb <b>e zgjuar<\/b>, dometh\u00ebn\u00eb, t\u00eb ndihmoj\u00eb p\u00ebrdoruesit t\u00eb orientohen n\u00eb volumin n\u00eb rritje t\u00eb informacionit: t\u00eb filtroj\u00eb, strukturoj\u00eb dhe ta ofroj\u00eb at\u00eb n\u00eb m\u00ebnyr\u00ebn m\u00eb t\u00eb p\u00ebrshtatshme. Ajo duhet t\u00eb jet\u00eb <b>e dobishme<\/b>, duke lejuar zgjidhjen e detyrave t\u00eb ndryshme direkt n\u00eb kutin\u00eb postare, p\u00ebr shembull, pages\u00ebn e gjobave (nj\u00eb funksion q\u00eb, p\u00ebr fat t\u00eb keq, e p\u00ebrdor). Dhe, natyrisht, posta duhet t\u00eb siguroj\u00eb mbrojtje informacioni, duke bllokuar spam-in dhe duke mbrojtur nga hakimet, pra, t\u00eb jet\u00eb <b>e sigurt\u00eb<\/b>.<br \/>\n<noindex><a rel=\"nofollow\" name=\"habracut\"><\/a><\/noindex><br \/>\nK\u00ebto drejtime p\u00ebrcaktojn\u00eb nj\u00eb seri detyrash ky\u00e7e, shum\u00eb prej t\u00eb cilave mund t\u00eb zgjidhen n\u00eb m\u00ebnyr\u00eb efikase me ndihm\u00ebn e m\u00ebsimit t\u00eb makinave. Ja disa shembuj t\u00eb ve\u00e7orive q\u00eb jan\u00eb zhvilluar brenda kuadrit t\u00eb strategjis\u00eb \u2014 nj\u00eb p\u00ebr \u00e7do drejtim.<\/p>\n<ul>\n<li><b>P\u00ebrgjigjja e zgjuar<\/b>. N\u00eb post\u00eb ka nj\u00eb funksion p\u00ebrgjigjeje t\u00eb zgjuar. Rrjeti nervor analizon tekstin e letr\u00ebs, kupton kuptimin dhe q\u00ebllimin e saj, dhe si rezultat ofron tre opsione m\u00eb t\u00eb p\u00ebrshtatshme p\u00ebr p\u00ebrgjigje: pozitive, negative dhe neutrale. Kjo ndihmon t\u00eb kursehet ndjesh\u00ebm koh\u00eb kur p\u00ebrgjigjemi n\u00eb letra, si dhe shpesh p\u00ebrgjigjemi n\u00eb m\u00ebnyr\u00eb jo standarde dhe arg\u00ebtuese p\u00ebr veten.\n<\/li>\n<li><b>Grupimi i letrave<\/b>, q\u00eb i p\u00ebrkasin porosive n\u00eb dyqanet online. Ne shpesh blejm\u00eb n\u00eb internet, dhe n\u00eb p\u00ebrgjith\u00ebsi, dyqanet mund t\u00eb d\u00ebrgojn\u00eb disa letra p\u00ebr \u00e7do porosi. P\u00ebr shembull, nga AliExpress, sh\u00ebrbimi m\u00eb i madh, vijn\u00eb shum\u00eb letra p\u00ebr nj\u00eb porosi, dhe ne kemi llogaritur se n\u00eb raste ekstremet e numri i tyre mund t\u00eb arrij\u00eb deri n\u00eb 29. Prandaj, me ndihm\u00ebn e modelit t\u00eb Njohjes s\u00eb Entiteteve t\u00eb Em\u00ebruara, ne identifikojm\u00eb numrin e porosis\u00eb dhe informacionin tjet\u00ebr nga teksti dhe grupojm\u00eb t\u00eb gjitha letrat n\u00eb nj\u00eb tem\u00eb. Po ashtu, ne tregojm\u00eb informacionin kryesor rreth porosis\u00eb n\u00eb nj\u00eb panel t\u00eb ve\u00e7ant\u00eb, q\u00eb e b\u00ebn pun\u00ebn me k\u00ebt\u00eb lloj letre m\u00eb t\u00eb leht\u00eb.\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/8b37b0bf0c0d152498027bc25d476128.jpg\" style=\"display:block;margin: 0 auto;\" \/>\n<\/li>\n<li><b>Antifishikimi<\/b>P\u00ebrgjimi \u00ebsht\u00eb nj\u00eb lloj i ve\u00e7ant\u00eb mashtrimi, p\u00ebrmes t\u00eb cilit keqb\u00ebr\u00ebsit p\u00ebrpiqen t\u00eb fitojn\u00eb informacion financiar (duke p\u00ebrfshir\u00eb informacionin mbi kartat bankare t\u00eb p\u00ebrdoruesit) dhe kredencialet. K\u00ebto mesazhe imitohen nga ato reale, t\u00eb d\u00ebrguara nga sh\u00ebrbimi, p\u00ebrfshir\u00eb edhe vizualisht. Prandaj, p\u00ebrmes Computer Vision ne njohim logo dhe stilin e dizajnit t\u00eb mesazheve t\u00eb kompanive t\u00eb m\u00ebdha (p\u00ebr shembull, Mail.ru, Sber, Alfa) dhe e marrim parasysh k\u00ebt\u00eb s\u00eb bashku me tekstin dhe karakteristika t\u00eb tjera n\u00eb klasifikator\u00ebt tan\u00eb t\u00eb spamit dhe phishingut.\n<\/li>\n<\/ul>\n<p><\/p>\n<h2>M\u00ebsimi makinerik<\/h2>\n<p>\nPak p\u00ebr m\u00ebsimin makinerik n\u00eb post\u00eb n\u00eb p\u00ebrgjith\u00ebsi. Posta \u00ebsht\u00eb nj\u00eb sistem me ngarkes\u00eb t\u00eb lart\u00eb: n\u00ebp\u00ebr serverat tan\u00eb kalojn\u00eb mesatarisht 1.5 miliard mesazhe n\u00eb dit\u00eb p\u00ebr 30 milion p\u00ebrdorues DAU. Rreth 30 sisteme m\u00ebsimi makinerik k\u00ebrkojn\u00eb t\u00eb sh\u00ebrbejn\u00eb t\u00eb gjitha funksionet dhe karakteristikat e nevojshme. <\/p>\n<p>\u00c7do mesazh kalon n\u00ebp\u00ebr nj\u00eb konvej super klasifikimi. S\u00eb pari, ne filtrojm\u00eb spam-in dhe l\u00ebm\u00eb mesazhet e mira. P\u00ebrdoruesit shpesh nuk e v\u00ebrejn\u00eb pun\u00ebn e anti-spamit, sepse 95-99% e spam-it nuk arrin as n\u00eb dosjen p\u00ebrkat\u00ebse. Njohja e spam-it \u00ebsht\u00eb nj\u00eb pjes\u00eb shum\u00eb e r\u00ebnd\u00ebsishme e sistemit ton\u00eb dhe m\u00eb e komplikuara, pasi n\u00eb fush\u00ebn e anti-spamit ka nj\u00eb adaptim t\u00eb vazhduesh\u00ebm midis sistemeve mbrojt\u00ebse dhe sulmuese, q\u00eb paraqet nj\u00eb sfid\u00eb t\u00eb vazhdueshme inxhinierike p\u00ebr ekipin ton\u00eb.<\/p>\n<p>M\u00eb pas, ndarim mesazhet nga njer\u00ebzit dhe robot\u00ebt. Mesazhet nga njer\u00ebzit jan\u00eb m\u00eb t\u00eb r\u00ebnd\u00ebsishmet, prandaj p\u00ebr to ofrojm\u00eb funksione si Smart Reply. Mesazhet nga robot\u00ebt ndahen n\u00eb dy pjes\u00eb: transaksionale \u2014 ato jan\u00eb mesazhe t\u00eb r\u00ebnd\u00ebsishme nga sh\u00ebrbimet, p\u00ebr shembull, konfirmimet e blerjeve ose rezervimeve t\u00eb hotelit, financat, dhe informative \u2014 k\u00ebto jan\u00eb reklamat e biznesit, zbritjet. <\/p>\n<p>Ne mendojm\u00eb se mesazhet transaksionale jan\u00eb po aq t\u00eb r\u00ebnd\u00ebsishme sa bisedat personale. Ato duhet t\u00eb jen\u00eb n\u00eb dor\u00eb, sepse shpesh \u00ebsht\u00eb e nevojshme t\u00eb gjenden shpejt informacionet p\u00ebr nj\u00eb porosi ose rezervimin e nj\u00eb bilete avioni, dhe ne humbim koh\u00eb n\u00eb k\u00ebrkimin e k\u00ebtyre mesazheve. Prandaj, p\u00ebr leht\u00ebsi, ne automatikisht i ndajm\u00eb ato n\u00eb gjasht\u00eb kategori kryesore: udh\u00ebtime, porosi, financa, bileta, regjistrime dhe, n\u00eb fund, nd\u00ebshkime.<\/p>\n<p>Informacionet e email-it jan\u00eb grupi m\u00eb i shum\u00ebllojsh\u00ebm dhe ndoshta m\u00eb pak i r\u00ebnd\u00ebsish\u00ebm q\u00eb nuk k\u00ebrkon nj\u00eb reagim t\u00eb menj\u00ebhersh\u00ebm, pasi asgj\u00eb thelb\u00ebsore nuk do t\u00eb ndryshoj\u00eb n\u00eb jet\u00ebn e p\u00ebrdoruesit n\u00ebse ai nuk e lexon nj\u00eb email t\u00eb till\u00eb. N\u00eb nd\u00ebrfaqen ton\u00eb t\u00eb re, ne i grumbullojm\u00eb ato n\u00eb dy grupe: rrjetet sociale dhe d\u00ebrgesat, duke pastruar vizualisht kutin\u00eb dhe duke l\u00ebn\u00eb n\u00eb shikim vet\u00ebm email-et e r\u00ebnd\u00ebsishme.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/175773879daaa29d1542a98070cb7972.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>Eksplorimi<\/h3>\n<p>\nNj\u00eb num\u00ebr i madh sistemesh sjell shum\u00eb v\u00ebshtir\u00ebsi n\u00eb eksploatimin e tyre. Modelet me kalimin e koh\u00ebs degradohen, ashtu si \u00e7do software: shenjat prishen, makinat shkojn\u00eb n\u00eb defekt, Kodi err\u00ebsirave grumbullohet. Gjithashtu, t\u00eb dh\u00ebnat vazhdimisht ndryshojn\u00eb: shtohen t\u00eb reja, transformohet modeli i zakonsh\u00ebm i sjelljes s\u00eb p\u00ebrdoruesve, etj. Prandaj, nj\u00eb model pa mb\u00ebshtetje t\u00eb duhur me kalimin e koh\u00ebs do t\u00eb funksionoj\u00eb gjithnj\u00eb e m\u00eb keq. <\/p>\n<p>Nuk duhet harruar as q\u00eb sa m\u00eb thell\u00eb t\u00eb dep\u00ebrtoj\u00eb m\u00ebsimi i makinerive n\u00eb jet\u00ebn e p\u00ebrdoruesve, aq m\u00eb shum\u00eb ndikim ata ushtrojn\u00eb n\u00eb ekosistem, dhe, si rezultat, aq m\u00eb shum\u00eb humbje financiare ose fitime mund t\u00eb marrin aktor\u00ebt e tregut. Prandaj, n\u00eb nj\u00eb num\u00ebr gjithnj\u00eb e m\u00eb t\u00eb madh fushash, aktor\u00ebt adaptohen p\u00ebr t\u00eb punuar me algoritmet e ML (shembuj klasik\u00eb \u2014 reklama, k\u00ebrkimi dhe anti-spam i p\u00ebrmendur m\u00eb par\u00eb).<\/p>\n<p>Po ashtu, detyrat e m\u00ebsimit t\u00eb makinerive kan\u00eb nj\u00eb ve\u00e7ori: \u00e7do ndryshim, madje edhe m\u00eb i vogli, n\u00eb sistem mund t\u00eb prodhoj\u00eb shum\u00eb pun\u00eb me modelin: pun\u00eb me t\u00eb dh\u00ebnat, ri-m\u00ebsim, vendosje, q\u00eb mund t\u00eb zgjas\u00eb jav\u00eb ose muaj. Prandaj, sa m\u00eb shpejt t\u00eb ndryshoj\u00eb mjedisi n\u00eb t\u00eb cilin operojn\u00eb modelet tuaja, aq m\u00eb shum\u00eb p\u00ebrpjekje k\u00ebrkon mb\u00ebshtetja e tyre. Ekipa mund t\u00eb krijoj\u00eb shum\u00eb sisteme dhe t\u00eb g\u00ebzohet p\u00ebr k\u00ebt\u00eb, dhe pastaj t\u00eb shpenzoj\u00eb pothuajse t\u00eb gjitha burimet n\u00eb mb\u00ebshtetje t\u00eb tyre, pa mund\u00ebsi p\u00ebr t\u00eb b\u00ebr\u00eb di\u00e7ka t\u00eb re. Ne ndodhi nj\u00eb her\u00eb t\u00eb p\u00ebrballemi me nj\u00eb situat\u00eb t\u00eb till\u00eb n\u00eb ekipin ton\u00eb t\u00eb anti-spamit. Dhe arrit\u00ebm n\u00eb p\u00ebrfundimin e qart\u00eb se mb\u00ebshtetja duhet automatizuar.<\/p>\n<h3>Automatizimi<\/h3>\n<p>\n\u00c7far\u00eb mund t\u00eb automatizohet? N\u00eb t\u00eb v\u00ebrtet\u00eb, pothuajse gjith\u00e7ka. Kam identifikuar kat\u00ebr drejtime q\u00eb p\u00ebrkufizojn\u00eb infrastruktur\u00ebn e m\u00ebsimit t\u00eb makinerive:<\/p>\n<ul>\n<li>mbledhja e t\u00eb dh\u00ebnave;\n<\/li>\n<li>ri-m\u00ebsimi;\n<\/li>\n<li>vendosja;\n<\/li>\n<li>testimi &amp; monitorimi.\n<\/li>\n<\/ul>\n<p>\nN\u00ebse ambienti \u00ebsht\u00eb i paq\u00ebndruesh\u00ebm dhe ndryshon vazhdimisht, infrastruktura rreth modelit b\u00ebhet shum\u00eb m\u00eb e r\u00ebnd\u00ebsishme se vet\u00eb modeli. Mund t\u00eb jet\u00eb nj\u00eb klasifikues linear tradicional, por n\u00ebse e ushqejm\u00eb at\u00eb me karakteristika t\u00eb sakta dhe krijojm\u00eb nj\u00eb reagim t\u00eb mir\u00eb nga p\u00ebrdoruesit, ai do t\u00eb funksionoj\u00eb shum\u00eb m\u00eb mir\u00eb se modelet m\u00eb t\u00eb avancuara teknologjike me gjitha mund\u00ebsit\u00eb e reja.<\/p>\n<h4>Cikli i reagimit<\/h4>\n<p>\nKy cik\u00ebl p\u00ebrfshin mbledhjen e t\u00eb dh\u00ebnave, ri-trajnim dhe shp\u00ebrndarje - n\u00eb thelb, gjith\u00eb ciklin e azhurnimit t\u00eb modelit. Pse \u00ebsht\u00eb kjo e r\u00ebnd\u00ebsishme? Shikoni grafikun e regjistrimeve n\u00eb post\u00eb:<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/f3c98bd49a101754cf299821a40e2b8e.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nNj\u00eb zhvillues i m\u00ebsimit t\u00eb makinerive ka implementuar nj\u00eb model anti-bot q\u00eb nuk lejon bot\u00ebt t\u00eb regjistrohen n\u00eb post\u00eb. Grafiku bie deri n\u00eb nj\u00eb pik\u00eb ku mbeten vet\u00ebm p\u00ebrdoruesit e v\u00ebrtet\u00eb. E gjith\u00eb kjo \u00ebsht\u00eb e shk\u00eblqyer! Por kalojn\u00eb kat\u00ebr or\u00eb, dhe bot\u00ebt e p\u00ebrshtatin skenarin e tyre, dhe \u00e7do gj\u00eb kthehet n\u00eb normalitet. N\u00eb k\u00ebt\u00eb implementim, zhvilluesi kaloi nj\u00eb muaj duke shtuar karakteristika dhe duke ri-trajnuar modelin, por spammer-at arrit\u00ebn t\u00eb adaptohen brenda kat\u00ebr or\u00ebsh.<\/p>\n<p>P\u00ebr t\u00eb mos u ndjer\u00eb aq dhej\u00eb dhe p\u00ebr t\u00eb mos pasur nevoj\u00eb t\u00eb rind\u00ebrtojm\u00eb gjith\u00e7ka, duhet t\u00eb mendojm\u00eb fillimisht se si do t\u00eb duket cikli i reagimit dhe \u00e7far\u00eb do t\u00eb b\u00ebjm\u00eb n\u00ebse ambienti ndryshon. Le t\u00eb fillojm\u00eb me mbledhjen e t\u00eb dh\u00ebnave - kjo \u00ebsht\u00eb karburanti p\u00ebr algoritmet tona.<\/p>\n<h2>Grumbullimi i t\u00eb dh\u00ebnave<\/h2>\n<p>\n\u00cbsht\u00eb e qart\u00eb se p\u00ebr rrjetet neurale moderne, sa m\u00eb shum\u00eb t\u00eb dh\u00ebna, aq m\u00eb mir\u00eb, dhe ato, n\u00eb thelb, generohet nga p\u00ebrdoruesit e produktit. Ne mund t\u00eb ndihmohemi nga p\u00ebrdoruesit n\u00eb etiketimin e t\u00eb dh\u00ebnave, por nuk duhet abuzuar me k\u00ebt\u00eb, sepse n\u00eb nj\u00eb moment p\u00ebrdoruesit do t\u00eb lodhen duke trajnuar modelet tuaja dhe do t\u00eb kalojn\u00eb n\u00eb nj\u00eb produkt tjet\u00ebr. <\/p>\n<p>Nj\u00eb nga gabimet m\u00eb t\u00eb zakonshme (k\u00ebtu po referohem te Andrew Ng) \u00ebsht\u00eb orientimi i tepruar ndaj metrikave n\u00eb setin testues, dhe jo ndaj reagimit nga p\u00ebrdoruesi, q\u00eb n\u00eb t\u00eb v\u00ebrtet\u00eb \u00ebsht\u00eb mat\u00ebsi kryesor i cil\u00ebsis\u00eb s\u00eb pun\u00ebs, pasi ne krijojm\u00eb produkt p\u00ebr p\u00ebrdoruesin. N\u00ebse p\u00ebrdoruesit nuk e kuptojn\u00eb ose nuk i p\u00eblqen funksionimi i modelit, at\u00ebher\u00eb gjith\u00e7ka \u00ebsht\u00eb e kota. <\/p>\n<p>Prandaj, p\u00ebrdoruesi gjithmon\u00eb duhet t\u00eb ket\u00eb mund\u00ebsin\u00eb t\u00eb votoj\u00eb, duhet t'i japim atij nj\u00eb mjet p\u00ebr reagimin. N\u00ebse ne mendojm\u00eb se nj\u00eb mesazh q\u00eb erdhi n\u00eb kutin\u00eb postare \u00ebsht\u00eb n\u00eb lidhje me financat, duhet ta etiketojm\u00eb at\u00eb si \"financa\" dhe t\u00eb skicojm\u00eb nj\u00eb buton q\u00eb p\u00ebrdoruesi mund ta shtyp\u00eb dhe t\u00eb thot\u00eb se kjo nuk \u00ebsht\u00eb financa.<\/p>\n<h3>Cil\u00ebsia e reagimit<\/h3>\n<p>\nLe t\u00eb flasim p\u00ebr cil\u00ebsin\u00eb e feedback-ut t\u00eb p\u00ebrdoruesve. S\u00eb pari, ju mund t\u00eb keni kuptime t\u00eb ndryshme p\u00ebr t\u00eb nj\u00ebjtin koncept me p\u00ebrdoruesin. P\u00ebr shembull, ju dhe menaxher\u00ebt e produkteve mund ta konsideroni \"financ\u00ebn\" si letra nga banka, nd\u00ebrsa p\u00ebrdoruesi mendon se letra nga gjyshja p\u00ebr pensionin \u00ebsht\u00eb gjithashtu e lidhur me financat. S\u00eb dyti, ka p\u00ebrdorues q\u00eb godasin butonat pa ndonj\u00eb logjik\u00eb. S\u00eb treti, p\u00ebrdoruesi mund t\u00eb jet\u00eb thell\u00ebsisht i p\u00ebrhersh\u00ebm n\u00eb p\u00ebrfundimet e tij. Nj\u00eb shembull i shk\u00eblqyer nga praktika jon\u00eb \u00ebsht\u00eb implementimi i klasifikuesit <noindex><a rel=\"nofollow\" href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%9D%D0%B8%D0%B3%D0%B5%D1%80%D0%B8%D0%B9%D1%81%D0%BA%D0%B8%D0%B5_%D0%BF%D0%B8%D1%81%D1%8C%D0%BC%D0%B0\">t\u00eb spam-it nigerian<\/a><\/noindex>, nj\u00eb lloj shum\u00eb qesharak i spam-it, kur p\u00ebrdoruesit u propozohet t\u00eb marrin disa milion dollar\u00eb nga nj\u00eb t\u00eb af\u00ebrm t\u00eb papritur n\u00eb Afrik\u00eb. Pas implementimit t\u00eb k\u00ebtij klasifikuesi, ne kontrolluam klikimet \"Nuk \u00ebsht\u00eb spam\" n\u00eb k\u00ebto letra, dhe doli se 80% e tyre ishin spam nigerian t\u00eb shijsh\u00ebm, gj\u00eb q\u00eb tregon se p\u00ebrdoruesit mund t\u00eb jen\u00eb jasht\u00ebzakonisht besnik\u00eb.<\/p>\n<p>Dhe mos harrojm\u00eb se jo vet\u00ebm njer\u00ebzit mund t\u00eb klikojn\u00eb n\u00eb butona, por edhe disa bota q\u00eb b\u00ebjn\u00eb sikur jan\u00eb shfletues. Prandaj, feedback-u i pap\u00ebrpunuar nuk \u00ebsht\u00eb i p\u00ebrshtatsh\u00ebm p\u00ebr m\u00ebsim. \u00c7far\u00eb mund t\u00eb b\u00ebjm\u00eb me k\u00ebt\u00eb informacion?<\/p>\n<p>Ne aplikojm\u00eb dy qasje: <\/p>\n<ul>\n<li><b>Feedback nga ML i lidhur<\/b>. P\u00ebr shembull, ne kemi nj\u00eb sistem anti-bot online, i cili, si\u00e7 e p\u00ebrmenda, merr nj\u00eb vendim t\u00eb shpejt\u00eb bazuar n\u00eb nj\u00eb num\u00ebr t\u00eb kufizuar karakteristikash. Dhe ka nj\u00eb sistem t\u00eb dyt\u00eb, t\u00eb ngadalsh\u00ebm, q\u00eb punon post faktum. Ai ka m\u00eb shum\u00eb t\u00eb dh\u00ebna p\u00ebr p\u00ebrdoruesin, p\u00ebr sjelljen e tij, etj. Si pasoje, merret vendimi m\u00eb i zgjedhur, si rezultat ka sakt\u00ebsi dhe plotesi m\u00eb t\u00eb lart\u00eb. Mund t\u00eb d\u00ebrgoni diferenc\u00ebn n\u00eb pun\u00ebn e k\u00ebtyre sistemeve n\u00eb t\u00eb parin si t\u00eb dh\u00ebna p\u00ebr m\u00ebsim. K\u00ebshtu, sistemi m\u00eb i thjesht\u00eb do t\u00eb p\u00ebrpiqet gjithmon\u00eb t\u00eb afroj\u00eb performanc\u00ebn e sistemit m\u00eb t\u00eb komplikuar.\n<\/li>\n<li><b>Klasifikimi i klikimeve<\/b>. Thjesht mund t\u00eb klasifikojm\u00eb \u00e7do klikim t\u00eb p\u00ebrdoruesit, t\u00eb vler\u00ebsojm\u00eb vlefshm\u00ebrin\u00eb e tij dhe mund\u00ebsin\u00eb e p\u00ebrdorimit. K\u00ebshtu veprojm\u00eb n\u00eb anti-spamin e post\u00ebs, duke p\u00ebrdorur karakteristikat e p\u00ebrdoruesit, historin\u00eb e tij, karakteristikat e d\u00ebrguesit, tekstin e vet\u00eb mesazhit dhe rezultatet e pun\u00ebs s\u00eb klasifikator\u00ebve. N\u00eb fund, arrijm\u00eb nj\u00eb sistem automatik q\u00eb vler\u00ebson feedback-un e p\u00ebrdoruesit. Dhe, pasi duhet ta st\u00ebrvitim m\u00eb pak shpesh, puna e tij mund t\u00eb b\u00ebhet themelore p\u00ebr t\u00eb gjitha sistemet e tjera. P\u00ebrpar\u00ebsia kryesore n\u00eb k\u00ebt\u00eb model ka precision, sepse st\u00ebrvitja e modelit me t\u00eb dh\u00ebna t\u00eb pasakta ka pasoja. \n<\/li>\n<\/ul>\n<p>\nNd\u00ebrsa po pastrkojm\u00eb t\u00eb dh\u00ebnat dhe po st\u00ebrvitim sistemet tona ML, nuk duhet t\u00eb harrojm\u00eb p\u00ebr p\u00ebrdoruesit, sepse p\u00ebr ne mij\u00ebra, miliona gabime n\u00eb grafik jan\u00eb statistik\u00eb, por p\u00ebr p\u00ebrdoruesin \u00e7do defekt \u00ebsht\u00eb nj\u00eb tragjedi. P\u00ebrve\u00e7 faktit q\u00eb p\u00ebrdoruesi duhet t\u00eb jetoj\u00eb me gabimin tuaj n\u00eb produkt, ai pas feedback-ut pret q\u00eb situata e till\u00eb t\u00eb p\u00ebrjashtohet n\u00eb t\u00eb ardhmen. Prandaj, gjithmon\u00eb duhet t'u japim p\u00ebrdoruesve jo vet\u00ebm mund\u00ebsin\u00eb p\u00ebr t\u00eb votuar, por edhe t\u00eb korrigjojn\u00eb sjelljen e sistemeve ML, duke krijuar, p\u00ebr shembull, heuristika personale p\u00ebr secilin klikim, p\u00ebr rastin e post\u00ebs, kjo mund t\u00eb jet\u00eb mund\u00ebsia p\u00ebr t\u00eb filtruar mesazhe t\u00eb ngjashme nga d\u00ebrguesi dhe subjekti p\u00ebr k\u00ebt\u00eb p\u00ebrdorues.<\/p>\n<p>Gjithashtu, duhet t\u00eb bazojm\u00eb modelin n\u00eb raportet e ndonj\u00eb raporti ose k\u00ebrkesave n\u00eb mb\u00ebshtetje, n\u00eb nj\u00eb m\u00ebnyr\u00eb gjysm\u00eb automatike ose manuale, p\u00ebr t\u00eb ndihmuar q\u00eb p\u00ebrdoruesit e tjer\u00eb t\u00eb mos vuajn\u00eb nga probleme t\u00eb ngjashme.<\/p>\n<h3>Heuristika p\u00ebr st\u00ebrvitje<\/h3>\n<p>\nMe k\u00ebto heuristika dhe ndihma ka dy probleme. E para \u00ebsht\u00eb se numri q\u00eb rritet gjithmon\u00eb i ndihmave \u00ebsht\u00eb i v\u00ebshtir\u00eb p\u00ebr t'u mbajtur, pa p\u00ebrmendur cil\u00ebsin\u00eb dhe funksionimin e tyre n\u00eb afat t\u00eb gjat\u00eb. Problemi i dyt\u00eb \u00ebsht\u00eb se gabimi mund t\u00eb mos jet\u00eb i shpesht\u00eb, dhe disa klikime p\u00ebr t\u00eb st\u00ebrvitur modelin mund t\u00eb mos mjaftojn\u00eb. Si\u00e7 dukej, k\u00ebta dy efekti q\u00eb nuk lidhen mund t\u00eb eliminohen n\u00eb m\u00ebnyr\u00eb t\u00eb ndjeshme n\u00ebse aplikohet qasja e m\u00ebposhtme.<\/p>\n<ol>\n<li>Krijojm\u00eb nj\u00eb ndihm\u00eb t\u00eb p\u00ebrkohshme. \n<\/li>\n<li>D\u00ebrgojm\u00eb t\u00eb dh\u00ebnat nga ajo n\u00eb model, ai rregullisht st\u00ebrvitet, p\u00ebrfshir\u00eb t\u00eb dh\u00ebnat e marra. K\u00ebtu, sigurisht, \u00ebsht\u00eb e r\u00ebnd\u00ebsishme q\u00eb heuristika t\u00eb ket\u00eb nj\u00eb sakt\u00ebsi t\u00eb lart\u00eb, p\u00ebr t\u00eb mos ulur cil\u00ebsin\u00eb e t\u00eb dh\u00ebnave n\u00eb setin e trajnimit. \n<\/li>\n<li>Pastaj vendosim monitorimin p\u00ebr aktivizimin e kostilit, dhe n\u00ebse pas nj\u00eb kohe, kostili nuk aktivizohet m\u00eb dhe plot\u00ebsisht mbulohet nga modeli, at\u00ebher\u00eb mund ta heqim at\u00eb pa ndonj\u00eb hezitim. Tani kjo problem nuk ka gjas\u00eb t\u00eb p\u00ebrs\u00ebritet.\n<\/li>\n<\/ol>\n<p>\nPrandaj, nj\u00eb ushtri kostilesh \u00ebsht\u00eb shum\u00eb e dobishme. E r\u00ebnd\u00ebsishme \u00ebsht\u00eb q\u00eb sh\u00ebrbimi i tyre t\u00eb jet\u00eb urgjent dhe jo i p\u00ebrhersh\u00ebm. <\/p>\n<h2>Rifitimi<\/h2>\n<p>\nRifitimi \u00ebsht\u00eb nj\u00eb proces i shtimit t\u00eb t\u00eb dh\u00ebnave t\u00eb reja, t\u00eb marra si rezultat i reagimeve nga p\u00ebrdoruesit ose sisteme t\u00eb tjera, dhe trajnimi i modelit ekzistues mbi to. Ka disa probleme q\u00eb mund t\u00eb lindin gjat\u00eb rifitimit:<\/p>\n<ol>\n<li>Modeli mund thjesht t\u00eb mos mb\u00ebshtes\u00eb rifitimin dhe t\u00eb m\u00ebsoj\u00eb vet\u00ebm nga e para. \n<\/li>\n<li>Askund n\u00eb librin e natyr\u00ebs nuk \u00ebsht\u00eb shkruar se rifitimi patjet\u00ebr do t\u00eb p\u00ebrmir\u00ebsoj\u00eb cil\u00ebsin\u00eb e pun\u00ebs n\u00eb prodhim. Shpesh ndodh p\u00ebr t\u00eb kund\u00ebrt\u00ebn, q\u00eb do t\u00eb thot\u00eb se \u00ebsht\u00eb e mundur vet\u00ebm p\u00ebrkeq\u00ebsimi.\n<\/li>\n<li>Ndryshimet mund t\u00eb jen\u00eb t\u00eb paparashikueshme. Ky \u00ebsht\u00eb nj\u00eb aspekt mjaft delikat q\u00eb ne e kemi zbuluar p\u00ebr veten. Edhe n\u00ebse modeli i ri n\u00eb testin A\/B tregon rezultate t\u00eb ngjashme me ata aktual, kjo nuk do t\u00eb thot\u00eb se do t\u00eb funksionoj\u00eb identikisht. Performanca e tyre mund t\u00eb ndryshoj\u00eb n\u00eb ndonj\u00eb p\u00ebrqindje t\u00eb vog\u00ebl, e cila mund t\u00eb sjell\u00eb gabime t\u00eb reja ose t\u00eb rikthej\u00eb gabime t\u00eb vjetra q\u00eb jan\u00eb korrigjuar. Me gabimet aktuale, si ne ashtu edhe p\u00ebrdoruesit, tashm\u00eb dim\u00eb t\u00eb jetojm\u00eb, dhe kur ndodhin shum\u00eb gabime t\u00eb reja, p\u00ebrdoruesi gjithashtu mund t\u00eb mos kuptoj\u00eb se \u00e7far\u00eb po ndodh, sepse ai pret sjellje t\u00eb parashikueshme.\n<\/li>\n<\/ol>\n<p>\nPrandaj, gj\u00ebja m\u00eb e r\u00ebnd\u00ebsishme gjat\u00eb rifitimit \u00ebsht\u00eb t\u00eb garantosh p\u00ebrmir\u00ebsimin e modelit, ose t\u00eb pakt\u00ebn t\u00eb mos e p\u00ebrkeq\u00ebsosh at\u00eb. <\/p>\n<p>E para q\u00eb na vjen n\u00eb mendje kur flasim p\u00ebr rifitimin, \u00ebsht\u00eb qasja e M\u00ebsimit Aktiv. \u00c7far\u00eb do t\u00eb thot\u00eb kjo? P\u00ebr shembull, nj\u00eb klasifikues p\u00ebrcakton n\u00ebse nj\u00eb mesazh i p\u00ebrket financave, dhe p\u00ebrreth kufijve t\u00eb vendosjes s\u00eb vendimeve shtojm\u00eb nj\u00eb most\u00ebr nga shembuj t\u00eb markuar. Kjo funksionon mir\u00eb, p\u00ebr shembull, n\u00eb reklamim, ku ka shum\u00eb reagime dhe mund t\u00eb trajnohet modeli n\u00eb m\u00ebnyr\u00eb online. Por n\u00ebse ka pak reagime, at\u00ebher\u00eb ne kemi nj\u00eb most\u00ebr shum\u00eb t\u00eb shtremb\u00ebruar n\u00eb lidhje me shp\u00ebrndarjen e t\u00eb dh\u00ebnave n\u00eb prodhim, mbi t\u00eb cil\u00ebn nuk mund t\u00eb vler\u00ebsojm\u00eb sjelljen e modelit gjat\u00eb eksploatimit.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/fd1b3e07bfaf896dde3248e43537f61a.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nN\u00eb t\u00eb v\u00ebrtet\u00eb, q\u00ebllimi yn\u00eb \u00ebsht\u00eb t\u00eb ruajm\u00eb modelet e vjetra, ato t\u00eb njohura, dhe t\u00eb fitojm\u00eb t\u00eb reja. K\u00ebtu kushtet \u00ebsht\u00eb pasardh\u00ebsia. Modeli, t\u00eb cilin shpesh e lan\u00e7ojm\u00eb me shum\u00eb v\u00ebshtir\u00ebsi, tashm\u00eb funksionon, k\u00ebshtu q\u00eb mund t\u00eb orientohemi nga performanca e tij. <\/p>\n<p>N\u00eb post\u00eb p\u00ebrdoren modele t\u00eb ndryshme: pem\u00eb, lineare, rrjeta nervore. P\u00ebr secil\u00ebn, ne krijojm\u00eb algoritmin ton\u00eb t\u00eb p\u00ebrshtatjes. Gjat\u00eb procesit t\u00eb p\u00ebrshtatjes, ne fitojm\u00eb jo vet\u00ebm t\u00eb dh\u00ebna t\u00eb reja, por shpesh ndodhin edhe karakteristika t\u00eb reja q\u00eb do t'i marrim parasysh n\u00eb t\u00eb gjith\u00eb algoritmet m\u00eb posht\u00eb.<\/p>\n<h3>Modelet lineare<\/h3>\n<p>\nSupozoni se kemi regresion logjistik. Nd\u00ebrtojm\u00eb loss-in e modelit nga komponent\u00ebt e m\u00ebposht\u00ebm:<\/p>\n<ul>\n<li>LogLoss mbi t\u00eb dh\u00ebnat e reja;\n<\/li>\n<li>regularizojm\u00eb peshat e karakteristikave t\u00eb reja (mos e preki t\u00eb vjetrat);\n<\/li>\n<li>m\u00ebsojm\u00eb edhe nga t\u00eb dh\u00ebnat e vjetra, p\u00ebr t\u00eb ruajtur modelet e vjetra;\n<\/li>\n<li>dhe, ndoshta, m\u00eb e r\u00ebnd\u00ebsishme: aplikojm\u00eb Harmonic Regularization, e cila garanton q\u00eb pesha t\u00eb mos ndryshoj\u00eb shum\u00eb n\u00eb krahasim me modelin e vjet\u00ebr sipas norm\u00ebs.\n<\/li>\n<\/ul>\n<p>\nDuke marr\u00eb parasysh se \u00e7do komponent i humbjes ka koeficient\u00eb, ne mund t'i p\u00ebrshtatim vlerat optimale p\u00ebr detyr\u00ebn ton\u00eb n\u00eb kros-validim ose n\u00eb baz\u00eb t\u00eb k\u00ebrkesave t\u00eb produktit.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/00ae86d23afb780f5e260e61f833d4bd.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>Pem\u00ebt<\/h3>\n<p>\nT\u00eb kalojm\u00eb te pem\u00ebt e vendimeve. Ne kemi filluar algoritmin e m\u00ebposht\u00ebm p\u00ebr p\u00ebrshtatjen e pem\u00ebve:<\/p>\n<ol>\n<li>N\u00eb prodhim punon nj\u00eb pyll me 100\u2014300 pem\u00eb, i trajnuar n\u00eb grupin e vjet\u00ebr t\u00eb t\u00eb dh\u00ebnave.\n<\/li>\n<li>N\u00eb fund t\u00eb fshijm\u00eb M = 5 pem\u00eb dhe shtojm\u00eb 2M = 10 t\u00eb reja, t\u00eb trajnuara mbi gjith\u00eb grupin e t\u00eb dh\u00ebnave, por me pesh\u00eb t\u00eb lart\u00eb p\u00ebr t\u00eb dh\u00ebnat e reja, q\u00eb natyrsh\u00ebm garanton nj\u00eb ndryshim inkremental t\u00eb modelit.\n<\/li>\n<\/ol>\n<p>\nE qart\u00eb se me kalimin e koh\u00ebs numri i pem\u00ebve rritet ndjesh\u00ebm, dhe \u00ebsht\u00eb e nevojshme t'i pak\u00ebsojm\u00eb ato her\u00eb pas here p\u00ebr t\u00eb q\u00ebndruar brenda afateve t\u00eb caktuara. P\u00ebr k\u00ebt\u00eb, p\u00ebrdorim Knowledge Distillation (KD), q\u00eb \u00ebsht\u00eb tani gjithandej. Nj\u00eb p\u00ebrmbledhje e parimit t\u00eb funksionimit t\u00eb tij.<\/p>\n<ol>\n<li>Ne kemi modelin aktual \"t\u00eb komplikuar\". E nis\u00ebm at\u00eb n\u00eb grupin e t\u00eb dh\u00ebnave p\u00ebr trajnim dhe marrim shp\u00ebrndarjen e mund\u00ebsive t\u00eb klasave n\u00eb daljen.\n<\/li>\n<li>M\u00eb pas, e m\u00ebsojm\u00eb modelin e nx\u00ebn\u00ebsit (n\u00eb k\u00ebt\u00eb rast, nj\u00eb model me num\u00ebr m\u00eb t\u00eb vog\u00ebl pem\u00ebsh) t\u00eb p\u00ebrs\u00ebris\u00eb rezultatet e pun\u00ebs s\u00eb modelit, duke p\u00ebrdorur shp\u00ebrndarjen e klasave si variablin e synuar.\n<\/li>\n<li>\u00cbsht\u00eb e r\u00ebnd\u00ebsishme t\u00eb theksohet se ne nuk p\u00ebrdorim asnj\u00eb markup t\u00eb dataset-it, dhe prandaj mund t\u00eb p\u00ebrdorim t\u00eb dh\u00ebna t\u00eb rast\u00ebsishme. Natyrisht, ne p\u00ebrdorim nj\u00eb most\u00ebr t\u00eb dh\u00ebnash nga dega e prodhimit si nj\u00eb most\u00ebr trajnimi p\u00ebr modelin e nx\u00ebn\u00ebsit. K\u00ebshtu, seti i trajnimit na lejon t\u00eb sigurojm\u00eb sakt\u00ebsin\u00eb e modelit, nd\u00ebrsa mostra e rrjedh\u00ebs garanton nj\u00eb performanc\u00eb t\u00eb ngjashme n\u00eb shp\u00ebrndarjen e prodhimit, duke kompensuar devijimin e mostr\u00ebs s\u00eb trajnimit.\n<\/li>\n<\/ol>\n<p>\n<img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/3fdfcb9b1e5a6fa824226a429afc475a.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nKombinimi i k\u00ebtyre dy metodave (shtimi i pem\u00ebve dhe reduktimi periodik i numrit t\u00eb tyre p\u00ebrmes Knowledge Distillation) siguron hyrjen e modeleve t\u00eb reja dhe vazhdim\u00ebsin\u00eb t\u00eb plot\u00eb.<\/p>\n<p>P\u00ebrmes KD ne gjithashtu b\u00ebjm\u00eb diferencimin e operacioneve me karakteristikat e modelit, p\u00ebr shembull, duke hequr karakteristika dhe punuar me mungesa. N\u00eb rastin ton\u00eb, ne kemi nj\u00eb s\u00ebr\u00eb karakteristikash statistikore t\u00eb r\u00ebnd\u00ebsishme (sip\u00ebrmarr\u00ebsit, hash-eve tekstuale, URL-ve, etj.) q\u00eb ruhen n\u00eb nj\u00eb baz\u00eb t\u00eb dh\u00ebnash q\u00eb kan\u00eb prirjen p\u00ebr t'u ndalur. Ky zhvillim natyrisht, modeli nuk \u00ebsht\u00eb i pregatitur, pasi situatat e ndalimit nuk p\u00ebrjetohen n\u00eb setin e trajnimit. N\u00eb raste t\u00eb tilla, ne kombinojm\u00eb teknikat KD dhe augmentimin: gjat\u00eb trajnimit p\u00ebr nj\u00eb pjes\u00eb t\u00eb t\u00eb dh\u00ebnave, ne heqim ose nullojm\u00eb karakteristikat e nevojshme, nd\u00ebrsa etiketat (daljet e modelit aktual) i marrim nga origjinali, modeli-nx\u00ebn\u00ebs m\u00ebson t\u00eb p\u00ebrs\u00ebris\u00eb k\u00ebt\u00eb shp\u00ebrndarje.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/5e1f5af9a359646a49f4fe88ffed1025.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nV\u00ebrejm\u00eb se sa m\u00eb serioze t\u00eb jet\u00eb manipulimi i modeleve, aq m\u00eb shum\u00eb n\u00eb p\u00ebrqindje k\u00ebrkohet mostra e rrjedh\u00ebs.<\/p>\n<p>P\u00ebr heqjen e karakteristikave, operacioni m\u00eb i thjesht\u00eb, k\u00ebrkohet vet\u00ebm nj\u00eb pjes\u00eb e vog\u00ebl e rrjedh\u00ebs, pasi ndryshohet vet\u00ebm disa karakteristika, dhe modeli aktual \u00ebsht\u00eb m\u00ebsuar n\u00eb t\u00eb nj\u00ebjtin set \u2014 ndryshimi \u00ebsht\u00eb minimal. P\u00ebr thjeshtimin e modelit (reduktimi i numrit t\u00eb pem\u00ebve disa her\u00eb) k\u00ebrkohet tashm\u00eb 50 nga 50. Dhe p\u00ebr mungesat e karakteristikave statistikore t\u00eb r\u00ebnd\u00ebsishme, t\u00eb cilat kan\u00eb nj\u00eb ndikim t\u00eb konsideruesh\u00ebm n\u00eb performanc\u00ebn e modelit, k\u00ebrkohet edhe m\u00eb shum\u00eb rrjedh\u00eb p\u00ebr t\u00eb barazuar pun\u00ebn e modelit t\u00eb ri t\u00eb q\u00ebndruesh\u00ebm ndaj mungesave n\u00eb t\u00eb gjitha llojet e letrave. <\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/8f3729cfff43be9ecdc47d532daa8d6f.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>FastText<\/h3>\n<p>\nT\u00eb kalojm\u00eb te FastText. Kujtoj se p\u00ebrfaq\u00ebsimi (Embedding) i fjal\u00ebs p\u00ebrb\u00ebhet nga shuma e embedding-ut t\u00eb vet\u00eb fjal\u00ebs dhe t\u00eb gjitha N-gramet e saj letrare, zakonisht trigramet. Pasi q\u00eb trigramet mund t\u00eb jen\u00eb mjaft t\u00eb shumta, p\u00ebrdoret Bucket Hashing, pra transformimi i gjith\u00eb hap\u00ebsir\u00ebs n\u00eb nj\u00eb hashmap t\u00eb fikst\u00eb. Si rezultat, matrica e peshave merr dimensionet e shtres\u00ebs s\u00eb brendshme mbi numrin e fjal\u00ebve + baketet. <\/p>\n<p>Gjat\u00eb ri-trajnimit shfaqen karakteristika t\u00eb reja: fjal\u00ebt dhe trigramet. N\u00eb ri-trajnimin standard nga Facebook nuk ndodh asgj\u00eb thelb\u00ebsore. Vet\u00ebm pesha t\u00eb vjetra trajnohen me entropin\u00eb e kryq\u00ebzuar mbi t\u00eb dh\u00ebna t\u00eb reja. N\u00eb k\u00ebt\u00eb m\u00ebnyr\u00eb, karakteristikat e reja nuk p\u00ebrdoren, natyrisht, ky qasje ka t\u00eb gjitha disavantazhet e p\u00ebrmendura m\u00eb par\u00eb, t\u00eb lidhura me parashikueshm\u00ebrin\u00eb e modelit n\u00eb prodhim. Prandaj, ne e p\u00ebrmir\u00ebsuam disi FastText-in. Shtojm\u00eb t\u00eb gjitha peshat e reja (fjal\u00ebt dhe trigramet), ri-trajnojm\u00eb t\u00eb gjith\u00eb matric\u00ebn me entropi t\u00eb kryq\u00ebzuar dhe shtojm\u00eb rregullimin harmonik n\u00eb p\u00ebrputhje me modelin linear, i cili garanton nj\u00eb ndryshim t\u00eb pap\u00ebrfillsh\u00ebm t\u00eb peshave t\u00eb vjetra.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/4491639587d1c91ac91d77b587ee0111.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>CNN<\/h3>\n<p>\nMe rrjetet konvencionale \u00ebsht\u00eb pak m\u00eb e komplikuar. N\u00ebse n\u00eb CNN trajnoni shtresat e fundit, natyrisht, mund t\u00eb aplikoni rregullimin harmonik dhe t\u00eb garantoni vazhdim\u00ebsin\u00eb. Por n\u00eb rastin kur k\u00ebrkohet ri-trajnim i gjith\u00eb rrjetit, at\u00ebher\u00eb nj\u00eb rregullim t\u00eb till\u00eb nuk mund ta vendosni n\u00eb t\u00eb gjitha shtresat. Megjithat\u00eb, ka nj\u00eb opsion p\u00ebr m\u00ebsimin e embedding-\u00ebve komplementar\u00eb p\u00ebrmes Triplet Loss (<noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/abs\/1503.03832\">artikulli origjinal<\/a><\/noindex>).<\/p>\n<h4>Triplet Loss<\/h4>\n<p>\nN\u00eb shembullin e detyr\u00ebs s\u00eb anti-phishing, do t\u00eb analizojm\u00eb n\u00eb terma t\u00eb p\u00ebrgjithsh\u00ebm Triplet Loss. Marrim logon ton\u00eb, si dhe shembuj pozitiv\u00eb dhe negativ\u00eb t\u00eb logove t\u00eb kompanive t\u00eb tjera. Minimizohet distanca midis t\u00eb par\u00ebve dhe maksimumizohet distanca midis t\u00eb dyt\u00ebve, duke b\u00ebr\u00eb k\u00ebt\u00eb me nj\u00eb hap t\u00eb vog\u00ebl p\u00ebr t\u00eb siguruar nj\u00eb kompaktes\u00eb m\u00eb t\u00eb madhe t\u00eb klasave. <\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/dc8547fa042286798eb5c2f0887c4da6.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nN\u00ebse ne ri-trajnojm\u00eb rrjetin, at\u00ebher\u00eb ne e ndryshojm\u00eb plot\u00ebsisht hap\u00ebsir\u00ebn metrike, dhe ajo b\u00ebhet plot\u00ebsisht e papajtueshme me t\u00eb m\u00ebparshmen. Kjo \u00ebsht\u00eb nj\u00eb problem serioz n\u00eb detyrat q\u00eb p\u00ebrdorin vektor\u00ebt. P\u00ebr ta zgjidhur k\u00ebt\u00eb problem, ne do t\u00eb p\u00ebrziem gjat\u00eb trajnimit embedding-et e vjetra.<\/p>\n<p>Kemi shtuar t\u00eb dh\u00ebna t\u00eb reja n\u00eb setin e trajnimit dhe po e trajnojm\u00eb nga e para versionin e dyt\u00eb t\u00eb modelit. N\u00eb faz\u00ebn e dyt\u00eb ne po e p\u00ebrfundojm\u00eb rrjetin ton\u00eb (Finetuning): fillimisht p\u00ebrfundohet shtresa e fundit, dhe pastaj e gjith\u00eb rrjeti lirohet. Gjat\u00eb procesit t\u00eb formimit t\u00eb tripleteve, vet\u00ebm nj\u00eb pjes\u00eb e embedDing-eve llogaritet me modelin e trajnuar, ndryshe pjesa tjet\u00ebr llogaritet me modelin e vjet\u00ebr. K\u00ebshtu, gjat\u00eb procesit t\u00eb nd\u00ebrtrajnimit ne sigurojm\u00eb p\u00ebrshtatshm\u00ebrin\u00eb e hap\u00ebsirave metrikore v1 dhe v2. Nj\u00eb variant i ve\u00e7ant\u00eb i rregullimit harmonik.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/854d4fcc97775b24e6863cc38743cd27.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<\/p>\n<h3>Arktitektura er\u00ebs<\/h3>\n<p>\nN\u00ebse e shqyrtojm\u00eb sistemin t\u00ebr\u00ebsisht me shembullin e anti-spam-it, modelet nuk jan\u00eb t\u00eb izoluar, por t\u00eb nd\u00ebrlidhura nj\u00ebra me tjetr\u00ebn. Merrni imazhe, tekst dhe karakteristika t\u00eb tjera, me ndihm\u00ebn e CNN dhe Fast Text merrni embedDing-e. M\u00eb pas, mbi embedDing-et aplikohet klasifikuesi, i cili jep skoret p\u00ebr klasa t\u00eb ndryshme (tipet e letrave, spam, prania e logos). Skoret dhe karakteristikat tashm\u00eb kalojn\u00eb n\u00eb pyllin e pem\u00ebve p\u00ebr t\u00eb marr\u00eb vendimin p\u00ebrfundimtar. Klasifikues t\u00eb ve\u00e7ant\u00eb n\u00eb k\u00ebt\u00eb skem\u00eb lejojn\u00eb t\u00eb interpretojm\u00eb m\u00eb mir\u00eb rezultatet e pun\u00ebs s\u00eb sistemit dhe t\u00eb p\u00ebrmir\u00ebsojm\u00eb komponentet n\u00eb rast t\u00eb problemeve, sesa t\u00eb paraqesim t\u00eb dh\u00ebnat e t\u00ebra si jan\u00eb n\u00eb pem\u00ebt e vendimeve.<\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/30effc5ef7398db5f085b6dd415647d3.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nN\u00eb fund ne garantuam vazhdim\u00ebsin\u00eb n\u00eb \u00e7do nivel. N\u00eb nivelin e posht\u00ebm n\u00eb CNN dhe Fast Text p\u00ebrdorim rregullimin harmonik, p\u00ebr klasifikuesit n\u00eb mes \u2014 gjithashtu rregullim harmonik dhe kalibrim t\u00eb skoreve p\u00ebr t\u00eb siguruar p\u00ebrputhshm\u00ebrin\u00eb e shp\u00ebrndarjes s\u00eb probabilitetit. Nd\u00ebrsa mbushja e pem\u00ebve trajnohet n\u00eb m\u00ebnyr\u00eb incrementale ose me ndihm\u00ebn e Knowledge Distillation.<\/p>\n<p>N\u00eb p\u00ebrgjith\u00ebsi, mb\u00ebshtetja e nj\u00eb sistemi t\u00eb till\u00eb t\u00eb nd\u00ebrlidhur t\u00eb m\u00ebsimit t\u00eb makinerive zakonisht paraqet sfida, pasi \u00e7do komponent n\u00eb nivelin e posht\u00ebm \u00e7on n\u00eb p\u00ebrdit\u00ebsimin e gjith\u00eb sistemit lart. Por, pasi n\u00eb konfigurimin ton\u00eb \u00e7do komponent ndryshon pak dhe \u00ebsht\u00eb kompatibil me t\u00eb kaluarin, e gjith\u00eb sistemi mund t\u00eb p\u00ebrdit\u00ebsohet n\u00eb pjes\u00eb pa pasur nevoj\u00eb t\u00eb ri-trajnohet e gjith\u00eb struktura, gj\u00eb q\u00eb e lejon ta mbajm\u00eb at\u00eb pa kosto t\u00eb larta. <\/p>\n<h2>Deploy<\/h2>\n<p>\nKemi trajtuar mbledhjen e t\u00eb dh\u00ebnave dhe ri-trajnimin e tipeve t\u00eb ndryshme t\u00eb modeleve, prandaj kalojm\u00eb n\u00eb implementimin e tyre n\u00eb ambientin e prodhimit.<\/p>\n<h3>A\/B-testimi<\/h3>\n<p>\nSi e thash\u00eb m\u00eb her\u00ebt, gjat\u00eb procesit t\u00eb grumbullimit t\u00eb t\u00eb dh\u00ebnave, zakonisht marrim nj\u00eb most\u00ebr t\u00eb shtremb\u00ebruar, p\u00ebr t\u00eb cil\u00ebn nuk \u00ebsht\u00eb e mundur t\u00eb vler\u00ebsojm\u00eb performanc\u00ebn e modelit n\u00eb prodhim. Prandaj, gjat\u00eb implementimit, \u00ebsht\u00eb thelb\u00ebsore t\u00eb krahasohet modeli me versionin e m\u00ebparsh\u00ebm, p\u00ebr t\u00eb kuptuar se si po shkojn\u00eb realisht ndodhit\u00eb, duke kryer teste A\/B. N\u00eb t\u00eb v\u00ebrtet\u00eb, procesi i lansimit dhe analizimi i grafik\u00ebve \u00ebsht\u00eb mjaft rutin\u00eb dhe i p\u00ebrshtatet automatikimit shum\u00eb mir\u00eb. Ne e lan\u00e7ojm\u00eb modelet tona gradualisht p\u00ebr 5%, 30%, 50% dhe 100% t\u00eb p\u00ebrdoruesve, duke mbledhur t\u00eb gjitha metrikat e disponueshme p\u00ebr p\u00ebrgjigjet e modelit dhe feedbackun e p\u00ebrdoruesve. N\u00eb rast t\u00eb ndonj\u00eb shp\u00ebrthimi t\u00eb r\u00ebnd\u00ebsish\u00ebm, ne e kthejm\u00eb automatikisht modelin, nd\u00ebrsa p\u00ebr rastet e tjera, pasi t\u00eb kemi mbledhur nj\u00eb num\u00ebr t\u00eb mjaftuesh\u00ebm klikimesh nga p\u00ebrdoruesit, marrim vendim p\u00ebr t\u00eb rritur p\u00ebrqindjen. N\u00eb fund, ne e \u00e7ojm\u00eb modelin e ri deri n\u00eb 50% t\u00eb p\u00ebrdoruesve plot\u00ebsisht automatikisht, nd\u00ebrsa personi e aprovon lansimin p\u00ebr audienc\u00ebn e gjith\u00eb p\u00ebrmbajtjes, edhe pse ky hap gjithashtu mund t\u00eb automatizohet.<\/p>\n<p>Megjithat\u00eb, procesi i testeve A\/B ofron mund\u00ebsi p\u00ebr optimizim. E gjith\u00eb \u00e7\u00ebshtja \u00ebsht\u00eb se \u00e7do test A\/B \u00ebsht\u00eb mjaft i gjat\u00eb (n\u00eb rastin ton\u00eb zgjat nga 6 deri n\u00eb 24 or\u00eb, n\u00eb var\u00ebsi t\u00eb numrit t\u00eb feedbackut), q\u00eb e b\u00ebn at\u00eb mjaft t\u00eb shtrenjt\u00eb dhe me burime t\u00eb kufizuara. P\u00ebrve\u00e7 k\u00ebsaj, k\u00ebrkohet nj\u00eb p\u00ebrqindje e mjaftueshme e fluksit p\u00ebr testin, p\u00ebr t\u00eb p\u00ebrshpejtuar koh\u00ebn e p\u00ebrgjithshme t\u00eb testit A\/B (t\u00eb grumbullosh nj\u00eb most\u00ebr statistike t\u00eb r\u00ebnd\u00ebsishme p\u00ebr t\u00eb vler\u00ebsuar metrikat me nj\u00eb p\u00ebrqindje t\u00eb vog\u00ebl mund t\u00eb marr\u00eb shum\u00eb koh\u00eb), q\u00eb e b\u00ebn numrin e vendeve t\u00eb A\/B testeve jasht\u00ebzakonisht t\u00eb kufizuar. Sigurisht, na duhet t\u00eb sjellim n\u00eb test vet\u00ebm modelet m\u00eb premtuese, t\u00eb cilat ne i marrim mjaft gjat\u00eb procesit t\u00eb rikualifikimit.<\/p>\n<p>P\u00ebr t\u00eb zgjidhur k\u00ebt\u00eb \u00e7\u00ebshtje, ne kemi trajnuar nj\u00eb klasifikues t\u00eb ve\u00e7ant\u00eb, q\u00eb parashikon suksesin e testit A\/B. P\u00ebr k\u00ebt\u00eb, si karakteristika p\u00ebrdorim statistik\u00ebn e marrjes s\u00eb vendimeve, Precision, Recall dhe metrika t\u00eb tjera n\u00eb grupin e trajnimet, n\u00eb mostrat e rezervuara dhe n\u00eb mostrat nga fluksi. Gjithashtu, krahasojm\u00eb modelin me at\u00eb aktual n\u00eb prodhim, me heuristat, dhe marrim parasysh kompleksitetin (Complexity) e modelit. Duke p\u00ebrdorur t\u00eb gjitha k\u00ebto karakteristika, klasifikuesi i trajnuar mbi historin\u00eb e testeve vler\u00ebson modelet kandidate, n\u00eb rastin ton\u00eb k\u00ebto jan\u00eb pyjet e pem\u00ebve, dhe merr vendim se cili prej tyre t\u00eb d\u00ebrgohet n\u00eb testin A\/B. <\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/41b121987209a663075be399e83f5a50.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nN\u00eb momentin e implementimit, ky qasje lejo t\u00eb rritet disa her\u00eb numri i testeve A\/B t\u00eb suksesshme.<\/p>\n<h3>Testimi &amp; monitorimi<\/h3>\n<p>\nTestimi dhe monitorimi, \u00e7udit\u00ebrisht, nuk d\u00ebmtojn\u00eb sh\u00ebndetin ton\u00eb; p\u00ebrkundrazi, e p\u00ebrmir\u00ebsojn\u00eb dhe na ndihmojn\u00eb t\u00eb heqim qafe stresin e tep\u00ebrt. Testimi lejon parandalimin e d\u00ebshtimeve, nd\u00ebrsa monitorimi \u2013 zbardhjen e tyre n\u00eb koh\u00eb, p\u00ebr t\u00eb zvog\u00ebluar ndikimin mbi p\u00ebrdoruesit.<\/p>\n<p>K\u00ebtu \u00ebsht\u00eb e r\u00ebnd\u00ebsishme t\u00eb kuptojm\u00eb se her\u00ebt apo von\u00eb, sistemi juaj gjithmon\u00eb do t\u00eb gaboj\u00eb \u2013 kjo \u00ebsht\u00eb e lidhur me ciklin e zhvillimit t\u00eb \u00e7do programi. N\u00eb fillim t\u00eb zhvillimit t\u00eb sistemit, ka shum\u00eb gabime derisa gjith\u00e7ka t\u00eb stabilizohet dhe t\u00eb p\u00ebrfundoj\u00eb faza kryesore e inovacioneve. Por me kalimin e koh\u00ebs, entropia merr mbizot\u00ebrimin dhe gabimet shfaqen s\u00ebrish \u2013 p\u00ebr shkak t\u00eb degradimit t\u00eb komponenteve p\u00ebrreth dhe ndryshimeve n\u00eb t\u00eb dh\u00ebna, si\u00e7 thash\u00eb n\u00eb fillim.<\/p>\n<p>K\u00ebtu do t\u00eb doja t\u00eb theksoja se \u00e7do sistem m\u00ebsimi makinerik duhet t\u00eb shqyrtohet nga k\u00ebndv\u00ebshtrimi i p\u00ebrfitimit t\u00eb tij gjat\u00eb gjith\u00eb ciklit t\u00eb jet\u00ebs. M\u00eb posht\u00eb n\u00eb grafik tregohet nj\u00eb shembull i funksionimit t\u00eb nj\u00eb sistemi p\u00ebr kapjen e spamit t\u00eb rrall\u00eb (n\u00eb grafik, linja \u00ebsht\u00eb af\u00ebr zeros). Nj\u00ebher\u00eb, p\u00ebr shkak t\u00eb nj\u00eb karakteristike t\u00eb keqakumuluar, ai u \u00e7mend. \u00c7udit\u00ebrisht, nuk kishte monitorim p\u00ebr aktivizim anormal, e n\u00eb rezultat, sistemi filloi t\u00eb ruaj\u00eb e-maile n\u00eb folderin \"spam\" n\u00eb pragun e vendimmarrjes n\u00eb nj\u00eb sasi t\u00eb madhe. Pavar\u00ebsisht nga p\u00ebrmir\u00ebsimi i pasojave, sistemi tashm\u00eb kishte gabuar aq shum\u00eb her\u00eb sa nuk do ta mbulonte shpenzimin dhe p\u00ebr pes\u00eb vjet. Kjo \u00ebsht\u00eb nj\u00eb d\u00ebshtim i plot\u00eb nga pik\u00ebpamja e ciklit t\u00eb jet\u00ebs s\u00eb modelit. <\/p>\n<p><img decoding=\"async\" alt=\"Zbatimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru\" src=\"\/wp-content\/uploads\/2019\/11\/65efd799537ceea3761cabc3a764ba90.jpg\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<br \/>\nPrandaj, nj\u00eb gj\u00eb e till\u00eb e thjesht\u00eb si monitorimi mund t\u00eb b\u00ebhet ky\u00e7 n\u00eb jet\u00ebn e modelit. P\u00ebrve\u00e7 metrikave standarde dhe t\u00eb qarta, ne shqyrtojm\u00eb shp\u00ebrndarjen e p\u00ebrgjigjeve dhe skor\u00ebve t\u00eb modelit, si dhe shp\u00ebrndarjen e vlerave t\u00eb karakteristikave ky\u00e7e. Me ndihm\u00ebn e divergjenc\u00ebs KL, ne mund t\u00eb krahasojm\u00eb shp\u00ebrndarjen aktuale me at\u00eb historike ose vlerat n\u00eb testin A\/B me rrjedh\u00ebn e tjera, q\u00eb na lejon t\u00eb v\u00ebrejm\u00eb anomali n\u00eb model dhe t\u00eb rikthejm\u00eb nd\u00ebrrimet n\u00eb koh\u00eb.<\/p>\n<p>N\u00eb shumic\u00ebn e rasteve, ne e nis\u00ebm versionin ton\u00eb t\u00eb par\u00eb t\u00eb sistemeve duke p\u00ebrdorur heuristika t\u00eb thjeshta ose modele, t\u00eb cilat m\u00eb von\u00eb i p\u00ebrdorim p\u00ebr monitorim. P\u00ebr shembull, ne monitorojm\u00eb modelin NER n\u00eb krahasim me regex p\u00ebr dyqanet specifike online, dhe n\u00ebse mbulimi i klasifikuesit bie n\u00eb krahasim me to, ne shqyrtojm\u00eb arsyet. Nj\u00eb tjet\u00ebr p\u00ebrdorim i dobish\u00ebm i heuristikave!<\/p>\n<h2>P\u00ebrfundime<\/h2>\n<p>\nLe t\u00eb kalojm\u00eb p\u00ebrs\u00ebri p\u00ebrmes mendimeve ky\u00e7e t\u00eb artikullit.<\/p>\n<ul>\n<li><b>Fibdaek<\/b>. Kemi gjithmon\u00eb parasysh p\u00ebrdoruesin: si do t\u00eb jetoj\u00eb ai me gabimet tona, si do t\u00eb mund t'i raportoj\u00eb ato. Mos harro, q\u00eb p\u00ebrdoruesit nuk jan\u00eb burim i sisht\u00ebm i feedback-ut p\u00ebr t\u00eb st\u00ebrvitur modelet, dhe \u00ebsht\u00eb e nevojshme t\u00eb filtrohet p\u00ebrmes sistemeve mb\u00ebshtet\u00ebse ML. N\u00ebse nuk ka mund\u00ebsi p\u00ebr t\u00eb mbledhur sinjal nga p\u00ebrdoruesi, ne k\u00ebrkojm\u00eb burime alternative feedback-u, p\u00ebr shembull, sisteme t\u00eb lidhura. \n<\/li>\n<li><b>Rifitimi<\/b>. K\u00ebtu e r\u00ebnd\u00ebsishme \u00ebsht\u00eb vazhdim\u00ebsia, prandaj mb\u00ebshtetemi n\u00eb modelin aktual t\u00eb prodhimit. Modelet e reja i st\u00ebrvitim k\u00ebshtu q\u00eb ato t\u00eb mos shfaqin ndryshime t\u00eb m\u00ebdha nga t\u00eb m\u00ebparshmet p\u00ebrmes rregullimit harmonik dhe trukesh t\u00eb ngjashme.<\/li>\n<li><b>Deploy<\/b>. Autodeploy sipas metrikeve redukton ndjesh\u00ebm koh\u00ebn e zbatimit t\u00eb modeleve. Monitorimi i statistikave dhe shp\u00ebrndarjes s\u00eb vendimeve, numri i gabimeve nga p\u00ebrdoruesit \u00ebsht\u00eb obligativ p\u00ebr nj\u00eb gjum\u00eb t\u00eb qet\u00eb dhe fundjava produktive.\n<\/li>\n<\/ul>\n<p>\nShpresoj q\u00eb ajo q\u00eb lexuat do t'ju ndihmoj\u00eb t\u00eb p\u00ebrmir\u00ebsoni m\u00eb shpejt sistemet tuaja ML, t\u00eb p\u00ebrshpejtoni daljen e tyre n\u00eb treg dhe t'i b\u00ebni ato m\u00eb t\u00eb besueshme, duke reduktuar nivelet e stresit nga puna.<br \/>\n<br \/>Burimi: <a content=\"nofollow\" rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/company\/mailru\/blog\/476714\/\">habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"<p>\u041f\u043e \u043c\u043e\u0442\u0438\u0432\u0430\u043c \u043c\u043e\u0438\u0445 \u0432\u044b\u0441\u0442\u0443\u043f\u043b\u0435\u043d\u0438\u0439 \u043d\u0430 Highload++ \u0438 DataFest Minsk 2019 \u0433. \u0414\u043b\u044f \u043c\u043d\u043e\u0433\u0438\u0445 \u0441\u0435\u0433\u043e\u0434\u043d\u044f \u043f\u043e\u0447\u0442\u0430 \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u043d\u0435\u043e\u0442\u044a\u0435\u043c\u043b\u0435\u043c\u043e\u0439 \u0447\u0430\u0441\u0442\u044c\u044e \u0436\u0438\u0437\u043d\u0438 \u0432 \u0441\u0435\u0442\u0438. \u0421 \u0435\u0435 \u043f\u043e\u043c\u043e\u0449\u044c\u044e \u043c\u044b \u0432\u0435\u0434\u0435\u043c \u0431\u0438\u0437\u043d\u0435\u0441-\u043f\u0435\u0440\u0435\u043f\u0438\u0441\u043a\u0443, \u0445\u0440\u0430\u043d\u0438\u043c \u0432\u0441\u0435\u0432\u043e\u0437\u043c\u043e\u0436\u043d\u0443\u044e \u0432\u0430\u0436\u043d\u0443\u044e \u0438\u043d\u0444\u043e\u0440\u043c\u0430\u0446\u0438\u044e, \u0441\u0432\u044f\u0437\u0430\u043d\u043d\u0443\u044e \u0441 \u0444\u0438\u043d\u0430\u043d\u0441\u0430\u043c\u0438, \u0431\u0440\u043e\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435\u043c \u043e\u0442\u0435\u043b\u0435\u0439, \u043e\u0444\u043e\u0440\u043c\u043b\u0435\u043d\u0438\u0435\u043c \u0437\u0430\u043a\u0430\u0437\u043e\u0432 \u0438 \u043c\u043d\u043e\u0433\u0438\u043c \u0434\u0440\u0443\u0433\u0438\u043c. \u0412 \u0441\u0435\u0440\u0435\u0434\u0438\u043d\u0435 2018 \u0433\u043e\u0434\u0430 \u043c\u044b \u0441\u0444\u043e\u0440\u043c\u0443\u043b\u0438\u0440\u043e\u0432\u0430\u043b\u0438 \u043f\u0440\u043e\u0434\u0443\u043a\u0442\u043e\u0432\u0443\u044e \u0441\u0442\u0440\u0430\u0442\u0435\u0433\u0438\u044e \u0440\u0430\u0437\u0432\u0438\u0442\u0438\u044f \u043f\u043e\u0447\u0442\u044b. \u041a\u0430\u043a\u043e\u0439 \u0436\u0435 \u0434\u043e\u043b\u0436\u043d\u0430 \u0431\u044b\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":[688],"tags":[],"class_list":["post-53143","post","type-post","status-publish","format-standard","hentry","category-administrirovanie"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 5.0.2 - aioseo.com -->\n\t<meta name=\"description\" content=\"\u041f\u043e \u043c\u043e\u0442\u0438\u0432\u0430\u043c \u043c\u043e\u0438\u0445 \u0432\u044b\u0441\u0442\u0443\u043f\u043b\u0435\u043d\u0438\u0439 \u043d\u0430.\" \/>\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\/sq\/blog\/administrirovanie\/ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru\" \/>\n\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.2\" \/>\n\t\t<meta property=\"og:locale\" content=\"sq_AL\" \/>\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\u042d\u043a\u0441\u043f\u043b\u0443\u0430\u0442\u0430\u0446\u0438\u044f \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u0433\u043e \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f \u0432 \u041f\u043e\u0447\u0442\u0435 Mail.ru | ProHoster\" \/>\n\t\t<meta property=\"og:description\" content=\"\u041f\u043e \u043c\u043e\u0442\u0438\u0432\u0430\u043c \u043c\u043e\u0438\u0445 \u0432\u044b\u0441\u0442\u0443\u043f\u043b\u0435\u043d\u0438\u0439 \u043d\u0430.\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/prohoster.info\/sq\/blog\/administrirovanie\/ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru\" \/>\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=\"2019-11-23T21:00:00+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2020-02-18T11:01:00+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\udd47 Shfryt\u00ebzimi i m\u00ebsimit t\u00eb makinerive n\u00eb Mail.ru | ProHoster","description":"Nga motivet e paraqitjeve t\u00eb mia n\u00eb.","canonical_url":"https:\/\/prohoster.info\/sq\/blog\/administrirovanie\/ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"sq_AL","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\u042d\u043a\u0441\u043f\u043b\u0443\u0430\u0442\u0430\u0446\u0438\u044f \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u0433\u043e \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f \u0432 \u041f\u043e\u0447\u0442\u0435 Mail.ru | ProHoster","og:description":"\u041f\u043e \u043c\u043e\u0442\u0438\u0432\u0430\u043c \u043c\u043e\u0438\u0445 \u0432\u044b\u0441\u0442\u0443\u043f\u043b\u0435\u043d\u0438\u0439 \u043d\u0430.","og:url":"https:\/\/prohoster.info\/sq\/blog\/administrirovanie\/ekspluatatsiya-mashinnogo-obucheniya-v-pochte-mail-ru","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":"2019-11-23T21:00:00+00:00","article:modified_time":"2020-02-18T11:01:00+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"53143","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":"","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":"2026-01-24 06:14:20","breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-02-28 20:30:25","updated":"2026-01-24 06:14:20","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/posts\/53143","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/comments?post=53143"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/posts\/53143\/revisions"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/media?parent=53143"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/categories?post=53143"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/sq\/wp-json\/wp\/v2\/tags?post=53143"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}