Kushanda kwemichina yekudzidza muMail.ru Mail

Kubva pakutaura kwangu paHighload ++ uye DataFest Minsk 2019.

Kune vakawanda nhasi, tsamba chinhu chinokosha muupenyu hwepaIndaneti. Nerubatsiro rwayo, tinoitisa bhizinesi tsamba, chengetedza marudzi ese eruzivo rwakakosha zvine chekuita nemari, kubhukidzwa kwehotera, kuisa maodha nezvimwe zvakawanda. Pakati pe2018, takagadzira zano rechigadzirwa rekuvandudza tsamba. Tsamba dzemazuva ano dzinofanira kunge dzakaita sei?

Mail inofanira kuva smart, ndiko kuti, kubatsira vashandisi kufamba nekuwedzera huwandu hweruzivo: sefa, dhizaini uye ipe iyo nenzira iri nyore. Anofanira kudaro zvinobatsira, kukubvumira kugadzirisa mabasa akasiyana-siyana mubhokisi rako retsamba, somuenzaniso, kubhadhara faindi (basa iro, zvinosuruvarisa, ndinoshandisa). Uye panguva imwecheteyo, hongu, tsamba dzinofanirwa kupa ruzivo rwekudzivirira, kucheka spam uye kudzivirira kubva pakubira, ndiko kuti, kuve. safe.

Idzi nzvimbo dzinotsanangura akati wandei ematambudziko akakosha, mazhinji ayo anogona kugadziriswa nemazvo uchishandisa muchina kudzidza. Heino mienzaniso yezvagara zviripo zvakagadziridzwa sechikamu chechirongwa - chimwe chegwara rega rega.

• Smart Reply. Mail ine smart reply feature. Iyo neural network inoongorora zvinyorwa zvetsamba, inonzwisisa zvazvinoreva uye chinangwa, uye semhedzisiro inopa matatu akakodzera mhinduro sarudzo: yakanaka, yakaipa uye isina kwayakarerekera. Izvi zvinobatsira kuchengetedza zvakanyanya nguva pakupindura mavara, uye zvakare kazhinji kupindura nenzira isiri-yakajairwa uye inosetsa.
• Kubatanidza maemailzvinoenderana nemirairo muzvitoro zvepa online. Isu tinowanzo tenga online, uye, sekutonga, zvitoro zvinogona kutumira maemail akati wandei kune yega yega. Semuenzaniso, kubva kuAliExpress, iyo yakakura sevhisi, akawanda mabhii anouya kune imwe odha, uye isu takaverenga kuti mune yekupedzisira kesi nhamba yavo inogona kusvika kusvika 29. Naizvozvo, tichishandisa Iyo Yakanyorwa Entity Recognition modhi, tinobvisa nhamba yekuraira. uye mamwe mashoko kubva pachinyorwa uye unganidza mavara ose mushinda imwe. Isu tinoratidzawo ruzivo rwekutanga nezve kurongeka mubhokisi rakasiyana, izvo zvinoita kuti zvive nyore kushanda nerudzi urwu rweemail.

  

• Anti-phishing. Phishing imhando ine ngozi yekunyepedzera yeemail, nerubatsiro rwevanorwisa vanoedza kuwana ruzivo rwemari (kusanganisira makadhi ekubhangi emushandisi) uye logins. Mavara akadaro anotevedzera chaiwo anotumirwa nesevhisi, kusanganisira nekuona. Naizvozvo, nerubatsiro rweComputer Vision, tinoona marogo uye dhizaini yemabhii kubva kumakambani makuru (semuenzaniso, Mail.ru, Sber, Alfa) uye funga izvi pamwe chete nezvinyorwa uye zvimwe zvinhu mune yedu spam uye phishing classifiers. .

Machine kudzidza

Zvishoma nezve muchina kudzidza mune email mune zvakajairika. Tsamba inzira inoremerwa zvakanyanya: avhareji yemabhii bhiriyoni 1,5 pazuva anopfuura nemasevha edu evashandisi veDAU mamirioni makumi matatu. Anosvika makumi matatu ekudzidzira masisitimu emuchina anotsigira ese anodiwa mabasa uye maficha.

Tsamba yega yega inoenda nepaipi yemhando yese. Kutanga tinocheka spam uye tinosiya maemail akanaka. Vashandisi kazhinji havacherechedzi basa reantispam, nekuti 95-99% ye spam haina kana kuguma mune yakakodzera folda. Kuzivikanwa kweSpam chikamu chakakosha zvikuru chehurongwa hwedu, uye yakanyanya kuoma, sezvo mumunda we-anti-spam pane nguva dzose inogadziriswa pakati pekudzivirira uye maitiro ekurwisa, izvo zvinopa kuenderera mberi kweinjiniya dambudziko reboka redu.

Zvadaro, tinoparadzanisa tsamba kubva kuvanhu uye marobhoti. Maemail anobva kuvanhu ndiwo akanyanya kukosha, saka tinovapa maficha akaita seSmart Reply kwavari. Tsamba dzemarobhoti dzakakamurwa kuita zvikamu zviviri: transactional - aya mabhii akakosha kubva kumasevhisi, semuenzaniso, kusimbiswa kwekutenga kana kuchengetwa kwehotera, mari, uye ruzivo - izvi kushambadzira kwebhizinesi, kuderedzwa.

Isu tinotenda kuti maemail ekutengeserana akaenzana mukukosha kune tsamba yemunhu. Vanofanira kunge vari pedyo, nekuti isu kazhinji tinoda kukurumidza kutsvaga ruzivo nezve odha kana kutakura tikiti remhepo, uye tinopedza nguva tichitsvaga mavara aya. Naizvozvo, kuti zvive nyore, tinozvipatsanura muzvikamu zvitanhatu zvikuru: kufamba, maodha, mari, matikiti, kunyoreswa uye, pakupedzisira, faindi.

Mavara eruzivo iboka rakakura uye pamwe risingakoshi, iro risingade mhinduro yekukurumidza, nekuti hapana chakakosha chichachinja muhupenyu hwemushandisi kana akasaverenga tsamba yakadaro. Muchiratidziro chedu chitsva, tinozviputira kuita tambo mbiri: masocial network uye matsamba enhau, nekudaro tichiona kuchenesa inbox uye nekusiya akakosha meseji achionekwa.

Operation

Nhamba huru yehurongwa inokonzera matambudziko akawanda mukushanda. Mushure mezvose, mamodheru anodzikisira nekufamba kwenguva, senge chero software: maficha anoputsika, michina inotadza, kodhi inove yakatsveyama. Mukuwedzera, data inogara ichichinja: zvitsva zvinowedzerwa, maitiro emushandisi anoshandurwa, nezvimwewo, saka muenzaniso usina rubatsiro rwakakodzera uchashanda zvakanyanya uye zvakanyanya kupfuura nguva.

Hatifanire kukanganwa kuti iyo yakadzika muchina kudzidza inopinda muhupenyu hwevashandisi, iyo yakakura kukanganisa kwavanoita pane ecosystem, uye, semhedzisiro, yakawanda kurasikirwa kwemari kana purofiti vatambi vemusika vanogona kugamuchira. Naizvozvo, munhamba iri kuwedzera yenzvimbo, vatambi vari kujairana nebasa reML algorithms (yekirasi mienzaniso ndeyekushambadza, kutsvaga uye yakambotaurwa antispam).

Zvakare, mabasa emuchina ekudzidza ane peculiarity: chero, kunyangwe idiki, shanduko muhurongwa inogona kuburitsa basa rakawanda nemuenzaniso: kushanda nedata, kudzidzisazve, kutumira, izvo zvinogona kutora mavhiki kana mwedzi. Naizvozvo, iyo inokurumidza nharaunda umo mamodheru anoshanda anochinja, ndipo painoda kuwedzera kushanda nesimba kuti azvichengetedze. Chikwata chinogona kugadzira akawanda masisitimu uye kufara nazvo, asi chobva chashandisa zvinenge zvese zviwanikwa zvaro chichizvichengeta, pasina mukana wekuita chero chinhu chitsva. Takambosangana nemamiriro ezvinhu akadaro muboka re antispam. Uye ivo vakaita mhedziso iri pachena yekuti rutsigiro runofanirwa kuve otomatiki.

Kuzvishandura

Chii chinogona kuitwa otomatiki? Zvinenge zvese, chaizvo. Ndaona nzvimbo ina dzinotsanangura michina yekudzidza michina:

• kuunganidza data;
• kuwedzera kudzidziswa;
• deploy;
• kuyedza & monitoring.

Kana iyo nharaunda isina kugadzikana uye inogara ichichinja, ipapo iyo yese midziyo yakatenderedza modhi inoshanduka kuve yakakosha kupfuura iyo modhi pachayo. Inogona kunge iri yakanaka yekare mutsara classifier, asi kana iwe ukaidyisa iyo chaiyo maficha uye kuwana yakanaka mhinduro kubva kune vashandisi, inoshanda zvirinani kupfuura State-Of-The-Art modhi nemabhero uye muridzo.

Feedback Loop

Kutenderera uku kunosanganisa kuunganidza data, kumwe kudzidziswa uye kutumira - kutaura zvazviri, iyo yese modhi yekuvandudza kutenderera. Nei zvichikosha? Tarisa purogiramu yekunyoresa mutsamba:

Muchina wekudzidza muchina ashandisa anti-bot modhi inodzivirira bots kubva kunyoresa muemail. Girafu inodonha kune kukosha uko chete vashandisi chaivo vanosara. Zvese zvakanaka! Asi maawa mana anopfuura, mabhoti anogadzirisa zvinyorwa zvavo, uye zvese zvinodzokera kune zvakajairika. Mukuita uku, mugadziri akapedza mwedzi achiwedzera maitiro uye kudzidzisazve modhi, asi spammer yakakwanisa kuchinjika mumaawa mana.

Kuti urege kurwadza zvakanyanya uye usazofanirwa kugadzirisa zvese gare gare, isu tinofanira kutanga tafunga kuti iyo loop yemhinduro ichataridzika sei uye zvatichaita kana zvakatipoteredza zvachinja. Ngatitangei nekuunganidza data - iyi ndiyo mafuta emaalgorithms edu.

Data collection

Zviri pachena kuti kune yemazuva ano neural network, iyo data yakawanda, iri nani, uye ivo, chaizvoizvo, inogadzirwa nevashandisi vechigadzirwa. Vashandisi vanogona kutibatsira nekumaka data, asi isu hatigone kushandisa zvisina kunaka izvi, nekuti pane imwe nguva vashandisi vachaneta nekupedzisa mamodheru ako uye vochinjira kune chimwe chigadzirwa.

Imwe yezvikanganiso zvinowanzoitika (pano ini ndinotaura nezvaAndrew Ng) yakanyanya kutarisisa mametrics pane test dataset, uye kwete pamhinduro kubva kumushandisi, inova ndiyo chiyero chikuru chemhando yebasa, sezvo isu tichigadzira. chigadzirwa chemushandisi. Kana mushandisi asinganzwisisi kana asingadi basa remuenzaniso, saka zvinhu zvose zvinoparara.

Naizvozvo, mushandisi anofanira kugara achikwanisa kuvhota uye anofanira kupihwa chishandiso chemhinduro. Kana isu tichifunga kuti tsamba ine chekuita nezvemari yasvika mubhokisi retsamba, tinoda kumaka "mari" uye dhirowa bhatani iro mushandisi anogona kudzvanya uye kutaura kuti iyi haisi mari.

Feedback quality

Ngatitaure nezve mhando yemhinduro yemushandisi. Chekutanga, iwe nemushandisi munogona kuisa zvirevo zvakasiyana mune imwe pfungwa. Semuenzaniso, iwe nevakuru vezvigadzirwa zvako munofunga kuti "mari" zvinoreva tsamba kubva kubhangi, uye mushandisi anofunga kuti tsamba kubva kuna mbuya nezve penjeni yake inorevawo mari. Chechipiri, kune vashandisi vanoda zvisina musoro kudzvanya mabhatani pasina chero pfungwa. Chechitatu, mushandisi anogona kunge akakanganisa zvakanyanya mumhedzisiro yake. Muenzaniso unokatyamadza kubva mukuita kwedu ndeyekuitwa kweiyo classifier Nigerian spam, rudzi runosetsa zvikuru rwespam apo mushandisi anokumbirwa kutora mamiriyoni anoverengeka emadhora kubva kuhama yakangoerekana yawanikwa iri kure muAfrica. Mushure mekuita kirasi iyi, takatarisa kudzvanya kwe "Kwete Spam" pamaemail aya, uye zvakazoitika kuti makumi masere muzana awo aive muto wekuNigeria spam, izvo zvinopa kuti vashandisi vanogona kunyengedzwa zvakanyanya.

Uye ngatisakanganwe kuti mabhatani anogona kudzvanywa kwete chete nevanhu, asiwo nemhando dzese dzebhoti dzinonyepedzera kuva browser. Saka mhinduro mbishi haina kunaka pakudzidza. Chii chaungaita neruzivo urwu?

Isu tinoshandisa nzira mbiri:

• Mhinduro kubva kuML yakabatana. Semuenzaniso, isu tine online anti-bot system, iyo, sezvandambotaura, inoita sarudzo nekukurumidza zvichienderana nenhamba shoma yezviratidzo. Uye kune yechipiri, inononoka sisitimu inoshanda mushure mechokwadi. Iyo ine data rakawanda nezve mushandisi, maitiro ake, nezvimwe. Nekuda kweizvozvo, sarudzo ine ruzivo rwakanyanya inoitwa; zvinoenderana, ine yakanyanya kurongeka uye kukwana. Iwe unogona kutungamira mutsauko mukushanda kweaya masisitimu kune yekutanga se data rekudzidzisa. Nokudaro, hurongwa huri nyore hunogara huchiedza kusvika kune kushanda kweimwe yakaoma.
• Click classification. Iwe unogona kungorongedza yega yega mushandisi kudzvanya, ongorora iko kutendeka kwayo uye kushandiswa kwayo. Isu tinoita izvi mune antispam mail, tichishandisa hunhu hwemushandisi, nhoroondo yake, hunhu hwekutumira, iwo mameseji pachawo uye mhedzisiro yevagadziri. Nekuda kweizvozvo, tinowana otomatiki sisitimu inosimbisa mhinduro yemushandisi. Uye sezvo ichida kudzidziswa zvakare zvishoma kazhinji, basa rayo rinogona kuve hwaro hwemamwe masisitimu ese. Chinonyanya kukosha mumuenzaniso uyu ndeyechokwadi, nokuti kudzidzisa modhi pane data isina kururama yakazara nemigumisiro.

Tichiri kuchenesa iyo data uye tichienderera mberi nekudzidzisa maML masisitimu edu, hatifanirwe kukanganwa nezvevashandisi, nekuti kwatiri, zviuru, mamirioni ezvikanganiso pagirafu nhamba, uye kumushandisi, bug yega yega inhamo. Pamusoro pekuti mushandisi anofanira neimwe nzira kurarama nekukanganisa kwako muchigadzirwa, mushure mekugamuchira mhinduro, anotarisira kuti mamiriro akafanana achabviswa mune ramangwana. Naizvozvo, zvakafanira nguva dzose kupa vashandisi kwete chete mukana wekuvhota, asiwo kugadzirisa maitiro eML masisitimu, kugadzira, semuenzaniso, heuristics yega yega yega mhinduro yekudzvanya; kana iri tsamba, uku kunogona kuve kugona kusefa. mabhii akadai nemutumiri uye zita remushandisi uyu.

Iwe zvakare unofanirwa kuvaka modhi zvichienderana nemimwe mishumo kana zvikumbiro zvekutsigira mune semi-otomatiki kana yemanyorero modhi kuitira kuti vamwe vashandisi vasatambure nematambudziko akafanana.

Heuristics yekudzidza

Pane matambudziko maviri neaya heuristics uye madondoro. Yokutanga ndeyokuti nhamba inoramba ichiwedzera yemadondoro yakaoma kuchengetedza, rega kunaka kwavo uye kushanda kwavo kwenguva refu. Dambudziko rechipiri nderekuti kukanganisa kunogona kusawanzoitika, uye kudzvanya kushoma kuti uwedzere kudzidzisa modhi hakuzokwane. Zvingaita sekuti izvi zviviri zvisina hukama mhedzisiro inogona kuderedzwa zvakanyanya kana inotevera nzira ikashandiswa.

1. Isu tinogadzira crutch yenguva pfupi.
2. Isu tinotumira data kubva kwairi kune modhi, inogara ichizvivandudza pachayo, kusanganisira pane yakagamuchirwa data. Pano, hongu, zvakakosha kuti heuristics ine hupamhi hwepamusoro kuitira kuti isaderedze kunaka kwe data mugadziriro yekudzidzira.
3. Zvadaro tinogadzirisa kutarisa kuti titange crutch, uye kana mushure meimwe nguva crutch isingachashandi uye yakafukidzwa zvachose nemuenzaniso, iwe unogona kuibvisa zvakachengeteka. Zvino dambudziko iri harisi kuzoitika zvakare.

Saka uto remadondoro rinobatsira zvikuru. Chinhu chikuru ndechokuti basa ravo nderokukurumidza uye kwete zvachose.

Kuwedzera kudzidziswa

Kudzidzirazve inzira yekuwedzera data nyowani inowanikwa semhedzisiro yemhinduro kubva kune vashandisi kana mamwe masisitimu, uye kudzidzisa iyo iripo modhi pairi. Panogona kunge paine matambudziko akati wandei nekuwedzera kudzidziswa:

1. Iyo modhi inogona kungosatsigira yekuwedzera kudzidziswa, asi dzidza kubva mukutanga.
2. Hapana papi mubhuku rezvakasikwa pakanyorwa kuti kumwe kudzidziswa kuchavandudza kunaka kwebasa mukugadzirwa. Kazhinji zvinopesana zvinoitika, ndiko kuti, kuparara chete kunogoneka.
3. Kuchinja kunogona kunge kusingatarisirwi. Iyi ipfungwa yakapusa yatakazvionera pachedu. Kunyangwe kana modhi nyowani muyedzo yeA/B inoratidza mibairo yakafanana kana ichienzaniswa neyazvino, izvi hazvireve kuti ichashanda zvakafanana. Basa ravo rinogona kusiyana muchikamu chimwe chete kubva muzana, izvo zvingaunza zvikanganiso zvitsva kana kudzorera zvekare zvakagadziriswa. Isu isu nevashandisi tatoziva kuti tingararama sei nezvikanganiso zvazvino, uye kana nhamba huru yezvikanganiso zvitsva ikamuka, mushandisi anogonawo kusanzwisisa zviri kuitika, nekuti anotarisira maitiro anofanotaurwa.

Naizvozvo, chinhu chakanyanya kukosha mukuwedzera kudzidziswa kuve nechokwadi chekuti modhi inovandudzwa, kana kuti kwete zvakanyanya.

Chinhu chekutanga chinouya mupfungwa kana tichitaura nezve kudzidziswa kwekuwedzera inzira yeKudzidza Inoshingaira. Izvi zvinorevei? Semuenzaniso, mugadziri anoona kana email ine chekuita nemari, uye kutenderedza muganhu wayo wesarudzo tinowedzera muenzaniso wemienzaniso yakanyorwa. Izvi zvinoshanda zvakanaka, semuenzaniso, mukushambadzira, uko kune mhinduro yakawanda uye iwe unogona kudzidzisa muenzaniso online. Uye kana paine mhinduro shoma, saka isu tinowana yakanyanya kurerekera sampuli maererano nekugadzira data kugovera, pahwaro hwacho hazvibviri kuongorora maitiro emuenzaniso panguva yekushanda.

Muchokwadi, chinangwa chedu ndechekuchengetedza mapatani ekare, mamodheru anozivikanwa, uye kuwana matsva. Kuenderera mberi kwakakosha pano. Iyo modhi, iyo yataiwanzotora marwadzo makuru kuburitsa, yave kutoshanda, saka tinogona kutarisa pakuita kwayo.

Mhando dzakasiyana dzinoshandiswa mutsamba: miti, mutsara, neural network. Kune yega yega tinoita yedu yekuwedzera yekudzidzira algorithm. Mukuita kwekuwedzera kudzidziswa, isu hatigamuchire chete data nyowani, asiwo kazhinji maficha matsva, atinozofunga nezvawo mune ese algorithms pazasi.

Linear modhi

Ngatitii tine logistic regression. Isu tinogadzira yekurasikirwa modhi kubva kune zvinotevera zvikamu:

• LogLoss pane itsva data;
• isu tinogadzirisa huremu hwezvinhu zvitsva (hatibati ekare);
• isu tinodzidzawo kubva kune yekare data kuitira kuchengetedza maitiro ekare;
• uye, zvichida, chinhu chinonyanya kukosha: tinowedzera Harmonic Regularization, iyo inovimbisa kuti zviyero hazvizoshanduki zvakanyanya maererano nemuenzaniso wekare maererano nemaitiro.

Sezvo chimwe nechimwe chikamu chekurasika chine coefficients, isu tinogona kusarudza iwo akanyanya kukosha ebasa redu kuburikidza nemuchinjiko-kusimbisa kana zvichibva pane zvigadzirwa zvinodiwa.

Miti

Ngatiendei kumiti yesarudzo. Isu takanyora inotevera algorithm yekuwedzera kudzidziswa kwemiti:

1. Iyo yakagadzirwa inomhanya sango re100-300 miti, iyo inodzidziswa pane yekare data set.
2. Pakupedzisira tinobvisa M = 5 zvimedu uye kuwedzera 2M = 10 zvitsva, zvakadzidziswa pane yese data set, asi nehuremu hwepamusoro hwe data itsva, iyo inovimbisa kuchinja kwekuwedzera mumuenzaniso.

Zviri pachena, nekufamba kwenguva, nhamba yemiti inowedzera zvikuru, uye inofanira kuderedzwa nguva nenguva kuitira kuti isangane nenguva. Kuti tiite izvi, tinoshandisa ikozvino ubiquitous Knowledge Distillation (KD). Muchidimbu nezve musimboti wekushanda kwayo.

1. Isu tine yazvino "complex" modhi. Isu tinomhanyisa pane yekudzidziswa data seti uye tinowana iyo kirasi mukana wekugovera pane zvakabuda.
2. Zvadaro, tinodzidzisa muenzaniso wevadzidzi (muenzaniso une miti mishoma munyaya iyi) kudzokorora migumisiro yemuenzaniso tichishandisa kugoverwa kwekirasi sechinangwa chekuchinja.
3. Izvo zvakakosha kuti ticherechedze pano kuti isu hatishandise iyo data set markup munzira ipi neipi, uye saka isu tinogona kushandisa zvisina tsarukano data. Ehe, isu tinoshandisa dhata data kubva murukova rwekurwa semuenzaniso wekudzidzira wemuenzaniso wemudzidzi. Nokudaro, seti yekudzidzira inotibvumira kuti tive nechokwadi chechokwadi chemuenzaniso, uye sampuli yerukova inovimbisa kushanda kwakafanana pakugoverwa kwekugadzirwa, kuripa kukanganisa kwegadziriro yekudzidzira.

Kubatanidzwa kwemaitiro maviri aya (kuwedzera miti uye nguva nenguva kuderedza nhamba yavo vachishandisa Knowledge Distillation) inovimbisa kuiswa kwemaitiro matsva uye kuenderera mberi kwakakwana.

Nerubatsiro rweKD, tinoitawo mashandiro akasiyana pamhando dzemhando, sekubvisa maficha uye kushanda pamagapu. Muchiitiko chedu, tine nhamba yezviverengero zvakakosha zvenhamba (nevanotumira, zvinyorwa zvinyorwa, URL, nezvimwewo) zvakachengetwa mudhatabhesi, izvo zvinowanzotadza. Muenzaniso, hongu, hauna kugadzirira kukura kwakadaro kwezviitiko, sezvo kukundikana kwemamiriro ezvinhu hakuitike mugadziriro yekudzidzira. Mumamiriro ezvinhu akadaro, isu tinosanganisa KD uye nzira dzekuwedzera: kana tichidzidzira chikamu che data, tinobvisa kana kugadzirisa zvakare maficha anodiwa, uye tinotora iwo ekutanga mavara (zvinobuda zvemuenzaniso wezvino), uye muenzaniso wevadzidzi unodzidza kudzokorora kugovera uku. .

Isu takaona kuti iyo yakakomba modhi manipulation inoitika, iyo yakakura muzana yekuyerera kwemuenzaniso inodiwa.

Kubvisa chimiro, iyo iri nyore kushanda, inoda chete chikamu chidiki chekuyerera, sezvo zvishoma chete zvezvinhu zvinoshanduka, uye iyo yazvino modhi yakadzidziswa pane imwechete seti - mutsauko uri mudiki. Kurerutsa modhi (kuderedza kuwanda kwemiti kakawanda), 50 kusvika 50 yatodiwa. Uye kune kusiiwa kwezvinhu zvakakosha zvehuwandu izvo zvinokanganisa zvakanyanya kuita kwemuenzaniso, kuyerera kwakanyanya kunodiwa kuenzanisa basa reiyo. modhi itsva inodzivirira kusiiwa pamhando dzese dzemavara.

FastText

Ngatienderei kuFastText. Rega ndikuyeuchidze kuti iyo inomiririra (Embedding) yezwi ine huwandu hwekumisikidzwa kweshoko pacharo uye nemavara aro N-grams, kazhinji matrigram. Sezvo panogona kunge paine akawanda matrigrams, Bucket Hashing inoshandiswa, kureva, kushandura nzvimbo yese kuita imwe yakagadziriswa hashmap. Nekuda kweizvozvo, huremu hwematrix hunowanikwa nehukuru hwemukati wemukati pahuwandu hwemashoko + mabhaketi.

Nekuwedzera kudzidziswa, zviratidzo zvitsva zvinoonekwa: mazwi uye trigrams. Hapana chakakosha chinoitika mukutevera kudzidziswa kubva paFacebook. Chete zviremu zvekare zvine muchinjiko-entropy zvinodzidziswa patsva data. Nokudaro, zvitsva hazvishandiswi; hongu, nzira iyi ine zvose zvakatsanangurwa pamusoro pezvipingamupinyi zvine chokuita nekusazivikanwa kwemuenzaniso mukugadzirwa. Ndosaka takagadzirisa FastText zvishoma. Isu tinowedzera huremu hutsva hutsva (mazwi uye trigrams), wedzera matrix yese nemuchinjikwa-entropy uye towedzera kurongeka kweharmonic nekuenzanisa neiyo mutsara modhi, iyo inovimbisa shanduko isingakoshi muhuremu hwekare.

CNN

Convolutional network yakati wandei yakaoma. Kana iwo ekupedzisira maseru apedzwa muCNN, saka, hongu, iwe unogona kushandisa harmonic kurongeka uye kuenderera mberi kwevimbiso. Asi kana kuwedzera kudzidziswa kwetiweki yese kuchidikanwa, saka kudzoreredza kwakadaro hakuchagone kushandiswa kune ese akaturikidzana. Nekudaro, pane sarudzo yekudzidzisa kuwedzeredza embeddings kuburikidza neTriplet Loss (chinyorwa chekutanga).

Kurasikirwa katatu

Tichishandisa anti-phishing basa semuenzaniso, ngatitarisei Triplet Kurasika mune zvakajairika. Isu tinotora logo yedu, pamwe nemienzaniso yakanaka uye yakaipa yelogos yemamwe makambani. Isu tinoderedza chinhambwe pakati pekutanga uye kuwedzera chinhambwe pakati pechipiri, tinoita izvi nekapu kadiki kuti tive nechokwadi chekubatana kukuru kwemakirasi.

Kana tikawedzera kudzidzisa network, ipapo metric nzvimbo yedu inoshanduka zvachose, uye inove isingaenderane zvachose neyekare. Iri idambudziko rakakura mumatambudziko anoshandisa mavheji. Kuti titenderere nedambudziko iri, isu tichasanganisa mune zvekare embeddings panguva yekudzidziswa.

Isu takawedzera data nyowani kune seti yekudzidziswa uye tiri kudzidzisa yechipiri vhezheni yemuenzaniso kubva pakutanga. Pachikamu chechipiri, isu tinowedzera kudzidzisa network yedu (Finetuning): kutanga chikamu chekupedzisira chinopedzwa, uyezve network yese haina chando. Mukuita kwekunyora katatu, tinoverenga chete chikamu chekumisikidza tichishandisa yakadzidziswa modhi, yasara - tichishandisa yekare. Nekudaro, mukuita kwekuwedzera kudzidziswa, tinoona kuenderana kwemetric nzvimbo v1 uye v2. Iyo yakasarudzika vhezheni yeharmonic dhizaini.

Zvose zvivakwa

Kana isu tikafunga iyo yese sisitimu tichishandisa antispam semuenzaniso, saka mamodheru haasi ega, asi akaiswa mukati meumwe neumwe. Isu tinotora mifananidzo, zvinyorwa uye zvimwe zvinhu, tichishandisa CNN uye Fast Text tinowana embeddings. Tevere, maclassifiers anoiswa pamusoro pezvakaiswa, izvo zvinopa zvibodzwa zvemakirasi akasiyana (mhando dzemavara, spam, kuvapo kwechiratidzo). Zviratidzo nezviratidzo zviri kutopinda musango remiti kuti sarudzo yekupedzisira iitwe. Vagadziri vega vega muchirongwa ichi vanoita kuti zvikwanise kududzira zvirinani mhedzisiro yehurongwa uye kunyanya kudzidzisazve zvikamu kana paine matambudziko, pane kudyisa data rese mumiti yesarudzo mune fomu mbishi.

Nekuda kweizvozvo, tinovimbisa kuenderera padanho rega rega. Pazasi nhanho muCNN uye Fast Text tinoshandisa harmonic kugarisa, kune vanoisa mukirasi pakati isu zvakare tinoshandisa harmonic kugarisa uye chiyero calibration yekuenderana kwekugovera kungangoita. Zvakanaka, kuwedzera kwemuti kunodzidziswa zvakanyanya kana kushandisa Ruzivo Distillation.

Kazhinji, kuchengetedza yakadaro nested muchina kudzidza sisitimu kazhinji kurwadza, sezvo chero chikamu chiri pazasi chinhanho chinotungamira kune inogadziridza kune yese system iri pamusoro. Asi sezvo mukugadzirisa kwedu chikamu chimwe nechimwe chinoshanduka zvishoma uye chinopindirana nechapfuura, iyo yose system inogona kuvandudzwa chidimbu nechidimbu pasina chikonzero chekudzokorora chimiro chose, chinobvumira kuti chitsigirwe pasina chakakomba pamusoro.

Deploy

Takakurukura kuunganidzwa kwedata uye kumwe kudzidziswa kwemhando dzakasiyana dzemhando, saka tiri kuenderera mberi nekuiswa kwavo munzvimbo yekugadzira.

A/B bvunzo

Sezvandambotaura, mukugadzirisa kuunganidza data, tinowanzowana sampuli yakarerekera, kubva kwairi zvisingabviri kuongorora kugadzirwa kwemuenzaniso. Naizvozvo, kana uchiendesa, modhi inofanirwa kuenzaniswa neyekare vhezheni kuti unzwisise kuti zvinhu zviri kunyatsofamba sei, ndiko kuti, kuita bvunzo dzeA/B. Muchokwadi, maitiro ekuburitsa uye kuongorora machati akajairika uye anogona kuve nyore otomatiki. Isu tinoburitsa mamodheru edu zvishoma nezvishoma kusvika ku5%, 30%, 50% uye 100% yevashandisi, tichiunganidza ese aripo metrics pamhinduro dzemodhi uye mhinduro yemushandisi. Panyaya yezvimwe zvakakomba zvekunze, isu tinongodzosera modhi, uye kune dzimwe nyaya, taunganidza huwandu hwakakwana hwekudzvanya kwevashandisi, tinosarudza kuwedzera muzana. Nekuda kweizvozvo, tinounza iyo modhi nyowani ku50% yevashandisi zvachose otomatiki, uye kuburitsa kune vese vateereri kuchatenderwa nemunhu, kunyangwe danho iri rinogona kuve otomatiki.

Nekudaro, iyo A / B yekuyedza maitiro inopa nzvimbo yekugadzirisa. Chokwadi ndechekuti chero bvunzo yeA/B yakarebesa (munyaya yedu inotora kubva pa6 kusvika kumaawa makumi maviri nemana zvichienderana nehuwandu hwemhinduro), izvo zvinoita kuti idhure uye iine zviwanikwa zvishoma. Pamusoro pezvo, chikamu chepamusoro chakakwana chekuyerera kwebvunzo chinodiwa kuti ikurumidze kumhanyisa nguva yese yebvunzo yeA/B (kutora sampuli yakakosha yekuongorora mametric padiki muzana inogona kutora nguva yakareba), izvo nhamba yeA/B slots yakanyanya kushomeka. Zviripachena, isu tinofanirwa kuyedza chete mamodheru anovimbisa, ayo atinogamuchira zvakanyanya panguva yekuwedzera kudzidziswa maitiro.

Kuti tigadzirise dambudziko iri, takadzidzisa kirasi yakasiyana inofanotaura kubudirira kwebvunzo yeA/B. Kuti tiite izvi, tinotora nhamba dzekuita sarudzo, Precision, Recall uye mamwe ma metrics pane yekudzidziswa seti, pane yakamisikidzwa, uye pane sampuli kubva murukova sezvikamu. Isu tinofananidzawo modhi neyazvino mukugadzirwa, neheuristics, uye tinofunga nezve Kuoma kweiyo modhi. Uchishandisa ese maficha aya, mukirasi akadzidziswa nezvenhoroondo yebvunzo anoongorora mhando dzevamiriri, kwatiri isu aya masango emiti, uye anosarudza kuti nderipi rekushandisa muyedzo yeA/B.

Panguva yekushandiswa, nzira iyi yakatibvumira kuwedzera nhamba yekubudirira kweA / B bvunzo kakawanda.

Testing & Monitoring

Kuedza uye kutarisa, zvisingaite, hazvikuvadze hutano hwedu; asi, pane kudaro, vanoivandudza uye vanotibvisa kushushikana kusingakoshi. Kuedza kunobvumidza iwe kudzivirira kutadza, uye kutarisa kunobvumidza iwe kuti uzvione nenguva kuti uderedze kukanganisa kune vashandisi.

Izvo zvakakosha kuti unzwisise pano kuti nekukurumidza kana gare gare system yako inogara ichiita zvikanganiso - izvi zvinokonzerwa nekutenderera kutenderera kwechero software. Pakutanga kwekuvandudzwa kwehurongwa kune nguva dzose mabhugi akawanda kusvikira zvinhu zvose zvagadzikana uye nhanho huru yekuvandudza yapera. Asi nekufamba kwenguva, entropy inotora mutero wayo, uye zvikanganiso zvinoonekwa zvakare - nekuda kwekushatiswa kwezvikamu zvakatenderedza uye shanduko yedata, yandakataura nezvayo pakutanga.

Pano ndinoda kuona kuti chero muchina wekudzidza sisitimu inofanirwa kutariswa kubva pakuona purofiti yayo mukati mehupenyu hwayo hwese. Girafu iri pazasi inoratidza muenzaniso wekuti sisitimu inoshanda sei kubata isingawanzo mhando ye spam (mutsara uri mugirafu uri pedyo ne zero). Rimwe zuva, nekuda kwehunhu hwakachengetwa zvisirizvo, akapenga. Sezvineiwo rombo rakanaka, pakanga pasina kutariswa kwekukonzeresa kusingaite; semhedzisiro, sisitimu yakatanga kuchengetedza mabhii akawanda kune "spam" folda pamuganho wekuita sarudzo. Kunyangwe kugadzirisa mhedzisiro, sisitimu yakatoita zvikanganiso kakawanda zvekuti haizozvibhadharira pachayo kunyangwe mumakore mashanu. Uye uku ndiko kukanganisa kwakakwana kubva pakuona kwehupenyu hwemuenzaniso.

Naizvozvo, chinhu chakareruka chakadai sekutarisa chinogona kuve kiyi muhupenyu hwemuenzaniso. Kuwedzera kune akajairwa uye akajeka metrics, isu tinofunga kugoverwa kwemuenzaniso mhinduro uye zvibodzwa, pamwe nekugoverwa kweakakosha maficha maitiro. Tichishandisa KL divergence, tinogona kufananidza kugovera kwazvino neiyo yenhoroondo kana kukosha muyedzo yeA/B pamwe nerukova rwasara, izvo zvinotibvumira kuona zvisizvo mumuenzaniso uye kudzoreredza shanduko nenguva.

Kazhinji, tinotangisa mavhezheni edu ekutanga emasisitimu tichishandisa zviri nyore heuristics kana modhi dzatinoshandisa sekutarisa mune ramangwana. Semuenzaniso, isu tinotarisa iyo NER modhi tichienzanisa neyakajairika kune chaiyo zvitoro zvepamhepo, uye kana iyo classifier yekuvhara inodonha mukuenzanisa navo, isu tinonzwisisa zvikonzero. Kumwe kushandiswa kunobatsira kweheuristics!

Migumisiro

Ngatiendei pamusoro pepfungwa huru dzechinyorwa zvakare.

• Fibdeck. Isu tinogara tichifunga nezvemushandisi: kuti achararama sei nekukanganisa kwedu, kuti achakwanisa sei kuzvizivisa. Usakanganwa kuti vashandisi havasi sosi yemhinduro dzakachena dzemhando dzekudzidzisa, uye inoda kucheneswa nerubatsiro rweanobatsira ML masisitimu. Kana zvisingabviri kuunganidza chiratidzo kubva kumushandisi, saka isu tinotarisa kune mamwe masosi emhinduro, semuenzaniso, akabatana masisitimu.
• Kuwedzera kudzidziswa. Chinhu chikuru pano ndechekuenderera mberi, saka tinovimba nemuenzaniso wezvino wekugadzira. Isu tinodzidzisa mhando nyowani kuitira kuti dzisasiyane zvakanyanya kubva kune yapfuura nekuda kweharmonic dhizaini uye akafanana matipi.
• Deploy. Auto-deployment yakavakirwa pamametrics inoderedza zvakanyanya nguva yekushandisa modhi. Kuongorora nhamba uye kugovera kwekuita sarudzo, iyo nhamba yekudonha kubva kuvashandisi inosungirwa kurara kwako kwakadzikama uye kunobudirira kwevhiki.

Zvakanaka, ndinovimba izvi zvinokubatsira kuvandudza masisitimu ako eML nekukurumidza, kuita kuti vatengese nekukurumidza, uye kuita kuti vawedzere kuvimbika uye kusanyanya kunetseka.

Source: www.habr.com