Zgjedhja e Ve\u00e7orive n\u00eb M\u00ebsimin e Makinerive<\/a><\/noindex>. Shpresojm\u00eb se do t\u00eb jet\u00eb e dobishme p\u00ebr t\u00eb gjith\u00eb ata q\u00eb jan\u00eb t\u00eb interesuar n\u00eb k\u00ebt\u00eb tem\u00eb.<\/p>\nN\u00eb bot\u00ebn reale, t\u00eb dh\u00ebnat nuk jan\u00eb gjithmon\u00eb aq t\u00eb pastra sa ndonj\u00ebher\u00eb mendojn\u00eb porosit\u00ebsit e bizneseve. Pik\u00ebrisht p\u00ebr k\u00ebt\u00eb arsye analiza inteligjente e t\u00eb dh\u00ebnave (data mining dhe data wrangling) \u00ebsht\u00eb n\u00eb k\u00ebrkes\u00eb. Ajo ndihmon n\u00eb zbulojn\u00eb vlerat e munguar dhe modelet n\u00eb t\u00eb dh\u00ebna t\u00eb strukturuara duke p\u00ebrdorur k\u00ebrkesa q\u00eb nuk mund t\u00eb p\u00ebrcaktohet nga njeriu. P\u00ebr t\u00eb gjetur dhe p\u00ebrdorur k\u00ebto modele p\u00ebr t\u00eb parashikuar rezultatet p\u00ebrmes lidhjeve t\u00eb zbuluara n\u00eb t\u00eb dh\u00ebna do t\u00eb ndihmoj\u00eb m\u00ebsimi i makin\u00ebs (Machine Learning).<\/p>\n
P\u00ebr t\u00eb kuptuar \u00e7do algorit\u00ebm, \u00ebsht\u00eb e nevojshme t\u00eb shqyrtohen t\u00eb gjitha variablat n\u00eb t\u00eb dh\u00ebna dhe t\u00eb zbulohet se \u00e7far\u00eb p\u00ebrfaq\u00ebsojn\u00eb k\u00ebto variabla. Kjo \u00ebsht\u00eb jasht\u00ebzakonisht e r\u00ebnd\u00ebsishme, pasi arsyetimi i rezultateve bazohet n\u00eb kuptimin e t\u00eb dh\u00ebnave. N\u00ebse t\u00eb dh\u00ebnat p\u00ebrmbajn\u00eb 5 apo edhe 50 variabla, mund t\u00eb studiohen t\u00eb gjith\u00eb ato. Por \u00e7far\u00eb n\u00ebse ka 200? At\u00ebher\u00eb thjesht nuk do t\u00eb mjaftojn\u00eb burimet p\u00ebr t\u00eb studiuar \u00e7do variab\u00ebl t\u00eb ve\u00e7ant\u00eb. P\u00ebr m\u00eb tep\u00ebr, disa algoritme nuk funksionojn\u00eb p\u00ebr t\u00eb dh\u00ebnat kategorike, dhe at\u00ebher\u00eb do t\u00eb duhet t\u00eb kthehen t\u00eb gjitha kolonat kategorike n\u00eb variabla sasior\u00eb (ata mund t\u00eb duken sasior\u00eb, por metrikat do t\u00eb tregojn\u00eb se jan\u00eb kategorike), p\u00ebr t'i shtuar n\u00eb model. N\u00eb k\u00ebt\u00eb m\u00ebnyr\u00eb, numri i variablave rritet dhe arrin n\u00eb rreth 500. \u00c7far\u00eb duhet t\u00eb b\u00ebjm\u00eb tani? Mund t\u00eb mendohet se p\u00ebrgjigja do t\u00eb ishte reduktimi i dimensioneve. Algoritmet e reduktimit t\u00eb dimensioneve zvog\u00eblojn\u00eb numrin e parametrave, por ndikojn\u00eb negativisht n\u00eb interpretuarshm\u00ebrin\u00eb. \u00c7far\u00eb n\u00ebse ekzistojn\u00eb teknika t\u00eb tjera q\u00eb p\u00ebrjashtojn\u00eb ve\u00e7orit\u00eb dhe p\u00ebr m\u00eb tep\u00ebr lejojn\u00eb t\u00eb kuptohen dhe interpretohen leht\u00ebsisht ato q\u00eb mbeten?<\/p>\n
N\u00eb var\u00ebsi t\u00eb asaj se a bazohet analiza n\u00eb regresion apo klasifikim, algoritmet e selektimit t\u00eb ve\u00e7orive mund t\u00eb jen\u00eb t\u00eb ndryshme, por ideja kryesore e realizimit t\u00eb tyre mbetet e nj\u00ebjt\u00eb.<\/p>\n
Variablat me korrelacion t\u00eb lart\u00eb<\/h3>\n Variablat e forta t\u00eb korrelacionit mes nj\u00ebra-tjetr\u00ebs ofrojn\u00eb t\u00eb nj\u00ebjtin informacion p\u00ebr modelin, prandaj nuk \u00ebsht\u00eb e nevojshme t\u00eb p\u00ebrdoren t\u00eb gjitha ato p\u00ebr analiz\u00eb. P\u00ebr shembull, n\u00ebse grupi i t\u00eb dh\u00ebnave (dataset) p\u00ebrmban karakteristikat \"Koha n\u00eb rrjet\" dhe \"Trafiku i p\u00ebrdorur\", mund t\u00eb supozojm\u00eb se ato do t\u00eb jen\u00eb n\u00eb nj\u00eb far\u00eb m\u00ebnyre t\u00eb korrelacionuara, dhe do t\u00eb shohim nj\u00eb korrelacion t\u00eb fort\u00eb, pavar\u00ebsisht se zgjidhim nj\u00eb most\u00ebr t\u00eb pandikshme t\u00eb t\u00eb dh\u00ebnave. N\u00eb k\u00ebt\u00eb rast, n\u00eb model \u00ebsht\u00eb e nevojshme vet\u00ebm nj\u00eb nga k\u00ebto variabla. N\u00ebse p\u00ebrdoren t\u00eb dyja, modelin do ta d\u00ebmtoj\u00eb (overfit) dhe do t\u00eb jet\u00eb i kaubur n\u00eb lidhje me nj\u00eb karakteristik\u00eb t\u00eb ve\u00e7ant\u00eb.<\/p>\n
Vlerat P<\/h3>\n N\u00eb algoritme t\u00eb tilla si regresioni linear, nd\u00ebrtimi i nj\u00eb modeli fillestar statistik \u00ebsht\u00eb gjithmon\u00eb nj\u00eb ide e mir\u00eb. Ai ndihmon p\u00ebr t\u00eb ilustruar r\u00ebnd\u00ebsin\u00eb e karakteristikave me an\u00eb t\u00eb vlerave t\u00eb tyre p, t\u00eb cilat jan\u00eb marr\u00eb nga ky model. Duke vendosur nivelin e r\u00ebnd\u00ebsis\u00eb, ne kontrollojm\u00eb vlerat e marra p, dhe n\u00ebse ndonj\u00eb vler\u00eb \u00ebsht\u00eb m\u00eb e ul\u00ebt se niveli i caktuar i r\u00ebnd\u00ebsis\u00eb, at\u00ebher\u00eb kjo karakteristik\u00eb shpallet e r\u00ebnd\u00ebsishme, dometh\u00ebn\u00eb ndryshimi i vler\u00ebs s\u00eb saj, me sa duke mund t\u00eb \u00e7oj\u00eb n\u00eb ndryshimin e vler\u00ebs s\u00eb objektivit (target).<\/p>\n
Zgjedhja direkte<\/h3>\n Zgjedhja direkte \u00ebsht\u00eb nj\u00eb teknik\u00eb q\u00eb p\u00ebrfshin p\u00ebrdorimin e regresionit hap pas hapi. Nd\u00ebrtimi i modelit fillon nga zero, dmth nga nj\u00eb model bosh, dhe pastaj \u00e7do iteracion shton nj\u00eb variab\u00ebl q\u00eb sjell p\u00ebrmir\u00ebsim n\u00eb modelin n\u00eb nd\u00ebrtim. Cila variab\u00ebl do t\u00eb shtohet n\u00eb model, p\u00ebrcaktohet nga r\u00ebnd\u00ebsia e saj. Kjo mund t\u00eb llogaritet duke p\u00ebrdorur metrika t\u00eb ndryshme. M\u00eb e zakonshmja \u00ebsht\u00eb p\u00ebrdorimi i vlerave p, t\u00eb marra n\u00eb modelin fillestar statistik me p\u00ebrfshirjen e t\u00eb gjitha variablave. Ndonj\u00ebher\u00eb, zgjedhja direkte mund t\u00eb \u00e7oj\u00eb n\u00eb d\u00ebmshp\u00ebrblim t\u00eb modelit, sepse n\u00eb model mund t\u00eb p\u00ebrfshihen variablat e forta t\u00eb korrelacionit, ndon\u00ebse ato ofrojn\u00eb t\u00eb nj\u00ebjtin informacion p\u00ebr modelin (por modeli mund t\u00eb tregoj\u00eb p\u00ebrmir\u00ebsim).<\/p>\n
Zgjedhja e kund\u00ebrt<\/h3>\n Selekcioni i p\u00ebrs\u00ebritur p\u00ebrfshin p\u00ebrjashtimin hap pas hapi t\u00eb karakteristikave, megjithat\u00eb n\u00eb drejtimin e kund\u00ebrt krahasuar me at\u00eb t\u00eb drejtp\u00ebrdrejt\u00eb. N\u00eb k\u00ebt\u00eb rast, modeli fillestar p\u00ebrfshin t\u00eb gjitha variablat e pavarur. M\u00eb pas, variablat p\u00ebrjashtohen (nj\u00eb p\u00ebr nj\u00eb n\u00eb \u00e7do iteracion), n\u00ebse ata nuk kontribuojn\u00eb asnj\u00eb vler\u00eb p\u00ebr modelin e ri regresion n\u00eb \u00e7do iteracion. Baza e p\u00ebrjashtimit t\u00eb karakteristikave jan\u00eb treguesit e vlerave-p t\u00eb modelit fillestar. N\u00eb k\u00ebt\u00eb metod\u00eb gjithashtu ekziston paqart\u00ebsia gjat\u00eb heqjes s\u00eb variablave me korrelacion t\u00eb lart\u00eb.<\/p>\n
P\u00ebrjashtimi i p\u00ebrs\u00ebritur i karakteristikave<\/h3>\n RFE \u00ebsht\u00eb nj\u00eb teknik\u00eb\/algorit\u00ebm i p\u00ebrdorur gjer\u00ebsisht p\u00ebr t\u00eb zgjedhur numrin e sakt\u00eb t\u00eb karakteristikave mjaft t\u00eb r\u00ebnd\u00ebsishme. Ndonj\u00ebher\u00eb metoda p\u00ebrdoret p\u00ebr t\u00eb shpjeguar disa nga \"karakteristikat m\u00eb t\u00eb r\u00ebnd\u00ebsishme\" q\u00eb ndikojn\u00eb n\u00eb rezultatet; dhe ndonj\u00ebher\u00eb p\u00ebr t\u00eb zvog\u00ebluar nj\u00eb num\u00ebr shum\u00eb t\u00eb madh variablash (rreth 200-400), dhe t\u00eb lihen vet\u00ebm ata q\u00eb japin ndonj\u00eb kontribut n\u00eb model, nd\u00ebrsa t\u00eb tjer\u00ebt p\u00ebrjashtohen. RFE p\u00ebrdor nj\u00eb sistem rangimi. Karakteristikave n\u00eb grupin e t\u00eb dh\u00ebnave u jepen rangu. M\u00eb pas k\u00ebto rangje p\u00ebrdoren p\u00ebr p\u00ebrjashtimin e p\u00ebrs\u00ebritur t\u00eb karakteristikave n\u00eb var\u00ebsi t\u00eb kolinearitetit midis tyre dhe r\u00ebnd\u00ebsis\u00eb s\u00eb k\u00ebtyre karakteristikave n\u00eb model. P\u00ebrve\u00e7 rangimit t\u00eb karakteristikave, RFE mund t\u00eb tregoj\u00eb n\u00ebse k\u00ebto karakteristika jan\u00eb t\u00eb r\u00ebnd\u00ebsishme apo jo, edhe p\u00ebr nj\u00eb num\u00ebr t\u00eb caktuar karakteristikash (sepse \u00ebsht\u00eb shum\u00eb e mundshme q\u00eb numri i zgjedhur i karakteristikave mund t\u00eb mos jet\u00eb optimal, dhe numri optimal i karakteristikave mund t\u00eb jet\u00eb m\u00eb i madh ose m\u00eb i vog\u00ebl se ai i zgjedhuri).<\/p>\n
Diagrami i r\u00ebnd\u00ebsis\u00eb s\u00eb karakteristikave<\/h3>\n Kur flitet p\u00ebr interpretimin e algoritmeve t\u00eb m\u00ebsimit t\u00eb makinerive, zakonisht diskutohet rreth regresioneve lineare (t\u00eb cilat lejojn\u00eb analizimin e r\u00ebnd\u00ebsis\u00eb s\u00eb karakteristikave p\u00ebrmes vlerave-p) dhe pem\u00ebve t\u00eb vendimeve (t\u00eb cilat tregojn\u00eb pik\u00ebrisht r\u00ebnd\u00ebsin\u00eb e karakteristikave n\u00eb form\u00ebn e nj\u00eb peme, si dhe hierarkin\u00eb e tyre). Nga ana tjet\u00ebr, n\u00eb algoritme t\u00eb tilla si Random Forest, LightGBM dhe XG Boost, shpesh p\u00ebrdoret diagrami i r\u00ebnd\u00ebsis\u00eb s\u00eb karakteristikave, dmth. krijohet nj\u00eb diagram i variablave dhe \"sasis\u00eb s\u00eb r\u00ebnd\u00ebsis\u00eb s\u00eb tyre\". Kjo \u00ebsht\u00eb ve\u00e7an\u00ebrisht e dobishme kur nevojitet t\u00eb ofrohet nj\u00eb arsyetim i strukturuar p\u00ebr r\u00ebnd\u00ebsin\u00eb e karakteristikave nga pik\u00ebpamja e ndikimit t\u00eb tyre n\u00eb biznes.<\/p>\n
Rregullimi<\/h3>\n Rregullimi b\u00ebhet p\u00ebr t\u00eb kontrolluar balanc\u00ebn midis paragjykimit (bias) dhe devijimit (variance). Paragjykimi tregon se sa mir\u00eb \u00ebsht\u00eb p\u00ebrshtatur modeli n\u00eb setin e t\u00eb dh\u00ebnave t\u00eb st\u00ebrvitjes. Devijimi tregon se sa t\u00eb ndryshme ishin parashikimet midis seteve t\u00eb trajnimit dhe testimit. N\u00eb m\u00ebnyr\u00eb ideale, si paragjykimi ashtu edhe devijimi duhet t\u00eb jen\u00eb t\u00eb vogla. K\u00ebtu ndihmon rregullimi! Ekzistojn\u00eb dy teknik\u00eb kryesore:<\/p>\n
Rregullimi L1 \u2013 Lasso: Lasso nd\u00ebshkon pesha e coeficient\u00ebve t\u00eb modelit p\u00ebr t\u00eb ndryshuar r\u00ebnd\u00ebsin\u00eb e tyre p\u00ebr modelin dhe madje mund t'i zeroj\u00eb ato (dmth. t\u00eb heq\u00eb k\u00ebta variabla nga modeli p\u00ebrfundimtar). Zakonisht, Lasso p\u00ebrdoret n\u00ebse seti i t\u00eb dh\u00ebnave ka nj\u00eb num\u00ebr t\u00eb madh variablash dhe nevojitet p\u00ebrjashtimi i disa prej tyre p\u00ebr t\u00eb kuptuar m\u00eb mir\u00eb se si faktor\u00ebt e r\u00ebnd\u00ebsish\u00ebm ndikojn\u00eb n\u00eb model (dmth. ato karakteristika q\u00eb jan\u00eb zgjedhur nga Lasso dhe t\u00eb cilat kan\u00eb r\u00ebnd\u00ebsi t\u00eb caktuar).<\/p>\n
Rregullimi L2 \u2013 metoda Ridge: Q\u00ebllimi i Ridge \u00ebsht\u00eb t\u00eb ruaj\u00eb t\u00eb gjith\u00eb variablat dhe n\u00eb t\u00eb nj\u00ebjt\u00ebn koh\u00eb t'u jap\u00eb atyre r\u00ebnd\u00ebsi bazuar n\u00eb kontributin n\u00eb efikasitetin e modelit. Ridge do t\u00eb jet\u00eb nj\u00eb zgjedhje e mir\u00eb n\u00ebse seti i t\u00eb dh\u00ebnave ka nj\u00eb num\u00ebr t\u00eb vog\u00ebl variablash dhe t\u00eb gjith\u00eb ata jan\u00eb t\u00eb nevojsh\u00ebm p\u00ebr interpretime dhe rezultatet e arritura.<\/p>\n
Duke qen\u00eb se Ridge l\u00eb t\u00eb gjith\u00eb variablat, nd\u00ebrsa Lasso p\u00ebrcakton m\u00eb mir\u00eb r\u00ebnd\u00ebsin\u00eb e tyre, \u00ebsht\u00eb zhvilluar nj\u00eb algorit\u00ebm q\u00eb kombinon karakteristikat m\u00eb t\u00eb mira t\u00eb t\u00eb dyja rregullimeve dhe \u00ebsht\u00eb i njohur si Elastic-Net.<\/p>\n
Ka shum\u00eb m\u00ebnyra t\u00eb tjera p\u00ebr p\u00ebrzgjedhjen e karakteristikave p\u00ebr m\u00ebsimin e automatik, por ideja kryesore mbetet gjithmon\u00eb e nj\u00ebjta: t\u00eb demonstrohet r\u00ebnd\u00ebsia e variablave dhe m\u00eb pas t\u00eb p\u00ebrjashtohen disa nga ata bazuar n\u00eb r\u00ebnd\u00ebsin\u00eb e marr\u00eb. R\u00ebnd\u00ebsia \u00ebsht\u00eb nj\u00eb term shum\u00eb subjektiv, pasi nuk \u00ebsht\u00eb nj\u00eb nivel, por nj\u00eb grup i t\u00eb dh\u00ebnave dhe diagrameve q\u00eb mund t\u00eb p\u00ebrdoren p\u00ebr t\u00eb gjetur karakteristikat ky\u00e7e.<\/p>\n
Faleminderit p\u00ebr leximin! G\u00ebzime n\u00eb m\u00ebsim!<\/p>\n
Burimi: habr.com<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"\u041f\u0440\u0438\u0432\u0435\u0442, \u0425\u0430\u0431\u0440! \u041c\u044b \u0432 \u00ab\u0420\u0435\u043a\u0441\u043e\u0444\u0442\u00bb \u043f\u0435\u0440\u0435\u0432\u0435\u043b\u0438 \u043d\u0430 \u0440\u0443\u0441\u0441\u043a\u0438\u0439 \u044f\u0437\u044b\u043a \u0441\u0442\u0430\u0442\u044c\u044e Feature Selection in Machine Learning. \u041d\u0430\u0434\u0435\u0435\u043c\u0441\u044f, \u0431\u0443\u0434\u0435\u0442 \u043f\u043e\u043b\u0435\u0437\u043d\u043e \u0432\u0441\u0435\u043c, \u043a\u0442\u043e \u043d\u0435\u0440\u0430\u0432\u043d\u043e\u0434\u0443\u0448\u0435\u043d \u043a \u0442\u0435\u043c\u0435. \u0412 \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u043c \u043c\u0438\u0440\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u043d\u0435 \u0432\u0441\u0435\u0433\u0434\u0430 \u0442\u0430\u043a\u0438\u0435 \u0447\u0438\u0441\u0442\u044b\u0435, \u043a\u0430\u043a \u043f\u043e\u0440\u043e\u0439 \u0434\u0443\u043c\u0430\u044e\u0442 \u0431\u0438\u0437\u043d\u0435\u0441-\u0437\u0430\u043a\u0430\u0437\u0447\u0438\u043a\u0438. \u0418\u043c\u0435\u043d\u043d\u043e \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u0432\u043e\u0441\u0442\u0440\u0435\u0431\u043e\u0432\u0430\u043d \u0438\u043d\u0442\u0435\u043b\u043b\u0435\u043a\u0442\u0443\u0430\u043b\u044c\u043d\u044b\u0439 \u0430\u043d\u0430\u043b\u0438\u0437 \u0434\u0430\u043d\u043d\u044b\u0445 (data mining \u0438 data wrangling). \u041e\u043d \u043f\u043e\u043c\u043e\u0433\u0430\u0435\u0442 \u0432\u044b\u044f\u0432\u043b\u044f\u0442\u044c \u043d\u0435\u0434\u043e\u0441\u0442\u0430\u044e\u0449\u0438\u0435 \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044f \u0438 \u043f\u0430\u0442\u0442\u0435\u0440\u043d\u044b \u0432 \u0441\u0442\u0440\u0443\u043a\u0442\u0443\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0445 […]<\/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-93155","post","type-post","status-publish","format-standard","hentry","category-administrirovanie"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t