$\"Yagiz$ <\/div>\n<\/div>\n

<\/a><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n

Dengesiz veri problemi olarak bilinen s\u0131n\u0131f dengesizli\u011fi, veri bilimi projeleri \u00e7er\u00e7evesinde dikkat edilmesi gereken bir sorundur. S\u0131n\u0131fland\u0131rma algoritmalar\u0131n\u0131n \u00e7o\u011fu, e\u011fitim setlerinin iyi dengelendi\u011fini varsayar. Algoritmalar\u0131n amac\u0131, genellikle, do\u011fru tahmin oran\u0131n\u0131 maksimize etmektir. Ancak varsay\u0131lan bu dengeli da\u011f\u0131l\u0131m \u00e7o\u011fu zaman ger\u00e7ek hayattaki veri k\u00fcmelerinde bulunmaz. S\u0131n\u0131flardan biri \u00e7ok az \u00f6rnekle temsil edilebilirken, di\u011fer s\u0131n\u0131f \u00e7ok say\u0131da \u00f6rnekle temsil edilir. Bu durumda s\u0131n\u0131fland\u0131rmada sorunlar ortaya \u00e7\u0131kabilir. Yani dengesiz veri seti, s\u0131n\u0131fland\u0131rma problemlerinde farkl\u0131 s\u0131n\u0131flar i\u00e7in e\u015fit say\u0131da g\u00f6zlem olmamas\u0131 durumunu tarif eder.<\/p>\n

\u00d6rne\u011fin bir veri seti d\u00fc\u015f\u00fcnelim:<\/p>\n

15 bin g\u00f6zlem i\u00e7ersin.<\/li>\n
Kullan\u0131c\u0131lar\u0131n sahte hesap olup olmad\u0131\u011f\u0131n\u0131 tahmin etmeye \u00e7al\u0131\u015f\u0131yor olal\u0131m.<\/li>\n
Hedef de\u011fi\u015fkendeki \u20181\u2019ler sahte hesaplar\u0131, \u20180\u2019lar ger\u00e7ek kullan\u0131c\u0131lar\u0131 temsil etsin.<\/li>\n<\/ul>\n
E\u011fer burada sahte hesaplar ile ger\u00e7ek kullan\u0131c\u0131lar\u0131n say\u0131lar\u0131 birbirine e\u015fit de\u011filse (hedef de\u011fi\u015fkende 0\u20131 dengesi e\u015fit de\u011filse) dengesiz veri seti ile kar\u015f\u0131 kar\u015f\u0131yay\u0131z demektir. Zaten bir d\u00fc\u015f\u00fcnecek olursan\u0131z, dengesiz veri \u00e7o\u011fu s\u0131n\u0131fland\u0131rma probleminin do\u011fas\u0131nda var. Sahte hesaplar ger\u00e7ek hesaplardan daha az olur, biz de bu az g\u00f6zleme sahip s\u0131n\u0131f\u0131 tahmin etmeye \u00e7al\u0131\u015f\u0131r\u0131z.<\/p>\n
Dengesiz veri seti her zaman downside olarak kar\u015f\u0131m\u0131za \u00e7\u0131kmayabilir. S\u0131n\u0131flar aras\u0131 fark k\u00fc\u00e7\u00fckse g\u00f6z ard\u0131 edilebilir.<\/p>\n
Peki ama dengesiz veri seti ne zaman bir downside olarak kar\u015f\u0131m\u0131za \u00e7\u0131kar?<\/p>\n
\n
\n
$\"\"$ <\/picture><\/div>\n<\/div>\n<\/figure>\n
Accuracy<\/strong> <\/em>yani do\u011fruluk, modelin do\u011fru tahmin etti\u011fi \u00f6rneklerin toplam \u00f6rne\u011fe oran\u0131d\u0131r. Ancak do\u011fruluk her zaman g\u00fcvenilir bir metrik de\u011fildir. \u00d6zellikle s\u0131n\u0131flar dengesizse (\u00f6rne\u011fin verilerin %95\u2019i \u201c0\u201d, sadece %5\u2019i \u201c1\u201d ise), mannequin sadece \u201c0\u201d s\u0131n\u0131f\u0131n\u0131 tahmin ederek %95 do\u011fruluk elde edebilir, ama bu y\u00fcksek do\u011fruluk yan\u0131lt\u0131c\u0131d\u0131r \u00e7\u00fcnk\u00fc mannequin az\u0131nl\u0131k s\u0131n\u0131f\u0131n\u0131 tamamen g\u00f6z ard\u0131 etmektedir.<\/p>\n
\u0130\u015fte bu duruma \u201cAccuracy Paradox<\/strong>\u201d yani \u201cDo\u011fruluk Paradoksu<\/strong>\u201d denir.<\/p>\n
\n
\n
$\"\"$ <\/picture><\/div>\n<\/div>\n<\/figure>\n
Accuracy, tahminlerin y\u00fczde ka\u00e7\u0131n\u0131n do\u011fru oldu\u011funu belirtir. Dengesiz veri setlerinde accuracy y\u00fcksek \u00e7\u0131kar, biz de \u00e7ok iyi tahmin yapt\u0131k diye d\u00fc\u015f\u00fcn\u00fcr\u00fcz. \u0130\u015fte bu duruma accuracy paradox <\/em>denir. Tahmin do\u011frulu\u011funu g\u00f6steren bu metrik dengesiz veri setlerinde anlams\u0131z olabilir. Recall ve precision <\/em>de\u011ferleri \u00f6nem kazan\u0131r.<\/p>\n
Yani, accuracy <\/em>y\u00fcksek fakat recall <\/em>veya precision d\u00fc\u015f\u00fck \u00e7\u0131k\u0131yorsa burada bir dengesizlik var san\u0131r\u0131m demelisiniz!<\/p>\n
\u00d6rne\u011fim 100 tane g\u00f6zlem i\u00e7inde 95 regular i\u015flem, 5 doland\u0131r\u0131c\u0131l\u0131k i\u015flemi olsun. Burada zaten t\u00fcm i\u015flemlerin regular oldu\u011funu belirtirsek accuracy<\/strong> de\u011ferimiz y\u00fczde 95 \u00e7\u0131kar.<\/p>\n
\n
\n
$\"\"$ <\/picture><\/div>\n<\/div>\n<\/figure>\n
\n
\n
$\"\"$ <\/picture><\/div>\n<\/div>\n<\/figure>\n
\n
\n
$\"\"$ <\/picture><\/div>\n<\/div>\n<\/figure>\n
\n
Daha fazla veri toplama: <\/strong>Az g\u00f6zlem bulunan s\u0131n\u0131ftaki g\u00f6zlem say\u0131s\u0131 art\u0131r\u0131labilir.<\/li>\n
Undersampling: <\/strong>Az\u0131nl\u0131k s\u0131n\u0131fta yeteri say\u0131da veri varsa, \u00e7o\u011funluk s\u0131n\u0131f\u0131n\u0131n g\u00f6zlem say\u0131s\u0131 rastgele se\u00e7im yap\u0131larak azalt\u0131labilir. Veri boyutu az de\u011filse tercih edilir, bir miktar bilgi kayb\u0131 g\u00f6ze al\u0131n\u0131r.<\/li>\n
Oversampling: <\/strong>Kanaatimce undersampling yap\u0131lam\u0131yorsa tercih edilmelidir. Overfit\u2019e sebebiyet verme ihtimali \u00e7ok y\u00fcksektir. Az\u0131nl\u0131k s\u0131n\u0131ftaki g\u00f6zlem say\u0131s\u0131 yeterli de\u011filse, bu s\u0131n\u0131ftaki g\u00f6zlemlerin kopyas\u0131 olu\u015fturularak \u00e7o\u011funluk s\u0131n\u0131f say\u0131s\u0131na e\u015fitlenebilir.<\/li>\n
Biraz oversampling biraz undersampling: <\/strong>Veri setindeki az\u0131nl\u0131k s\u0131n\u0131ftaki g\u00f6zlem say\u0131s\u0131 az ise, oversampling yap\u0131lmas\u0131 gerektiyse, \u00f6nce \u00e7o\u011funluk s\u0131n\u0131f i\u00e7in undersampling <\/em>sonra az\u0131nl\u0131k s\u0131n\u0131f i\u00e7in oversampling uygulanabilir.<\/li>\n<\/ol>\n
Resampling teknikleri aras\u0131nda en s\u0131k kullan\u0131lan y\u00f6ntem ise SMOTE (Artificial Minority <\/em>Oversampling Approach) olarak bilinir. Klasik veri kopyalama \u015feklinde hareket etmez. Veri kopyalama modele yeni bilgi vermez.<\/p>\n
\n
Okay en yak\u0131n kom\u015fu algoritmas\u0131n\u0131 kullan\u0131r.<\/li>\n
\u00d6ncelikle az\u0131nl\u0131k s\u0131n\u0131ftan random veri se\u00e7er.<\/li>\n
Se\u00e7ilen random veri etraf\u0131ndaki kom\u015fular\u0131 se\u00e7er. Kom\u015fu ile se\u00e7ilen veri aras\u0131ndaki yerde veri \u00fcretir.<\/li>\n
Bu i\u015flem \u00e7o\u011funluk s\u0131n\u0131ftaki g\u00f6zlem say\u0131s\u0131na ula\u015fana dek tekrar eder.<\/li>\n<\/ul>\n
SMOTE sadece continous <\/em>verilere uygulanabilirken SMOTE-NC (Artificial Minority Oversampling Approach \u2014 Nominal Continous) hem nominal <\/em>hem continous <\/em>verilere uygulanabiliyor. Bunlar\u0131n d\u0131\u015f\u0131nda ise Borderline-SMOTE, SVM-SMOTE ve ADASYN gibi resampling teknikleri bulunmakta.<\/p>\n<\/div>\n\n","protected":false},"excerpt":{"rendered":"
Dengesiz veri problemi olarak bilinen s\u0131n\u0131f dengesizli\u011fi, veri bilimi projeleri \u00e7er\u00e7evesinde dikkat edilmesi gereken bir sorundur. S\u0131n\u0131fland\u0131rma algoritmalar\u0131n\u0131n \u00e7o\u011fu, e\u011fitim setlerinin iyi dengelendi\u011fini varsayar. Algoritmalar\u0131n amac\u0131, genellikle, do\u011fru tahmin oran\u0131n\u0131 maksimize etmektir. Ancak varsay\u0131lan bu dengeli da\u011f\u0131l\u0131m \u00e7o\u011fu zaman ger\u00e7ek hayattaki veri k\u00fcmelerinde bulunmaz. S\u0131n\u0131flardan biri \u00e7ok az \u00f6rnekle temsil edilebilirken, di\u011fer s\u0131n\u0131f \u00e7ok say\u0131da […]<\/p>\n","protected":false},"author":2,"featured_media":1414,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[55],"tags":[767,1313,1311,1303,1306,1314,1307,1309,1310,1308,1305,1304,1312],"class_list":["post-1412","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-machine-learning","tag-apr","tag-bugra","tag-calisir","tag-dengesiz","tag-ile","tag-karakollu","tag-makine","tag-modelleri","tag-nasil","tag-ogrenmesi","tag-setleri","tag-veri","tag-yagiz"],"_links":{"self":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/1412","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1412"}],"version-history":[{"count":1,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/1412\/revisions"}],"predecessor-version":[{"id":1413,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/1412\/revisions\/1413"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/media\/1414"}],"wp:attachment":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1412"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1412"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1412"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}