Makine \u00f6\u011frenimi, g\u00fcn\u00fcm\u00fcz\u00fcn en heyecan verici ve d\u00f6n\u00fc\u015ft\u00fcr\u00fcc\u00fc teknoloji alanlar\u0131ndan biridir. Verilerden \u00f6\u011frenerek tahminler yapabilen veya kararlar alabilen sistemler olu\u015fturma bilimidir. Bu geni\u015f alan i\u00e7inde, Karar A\u011fa\u00e7lar\u0131 (Decision Trees) hem basitlikleri hem de g\u00fc\u00e7l\u00fc performanslar\u0131 nedeniyle \u00f6ne \u00e7\u0131kan, temel ve anla\u015f\u0131lmas\u0131 kolay algoritmalardan biridir. Bu makalede, Karar A\u011fa\u00e7lar\u0131n\u0131n ne oldu\u011funu, nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131, avantajlar\u0131n\u0131, dezavantajlar\u0131n\u0131 ve modern makine \u00f6\u011frenimindeki yerini derinlemesine ancak anla\u015f\u0131l\u0131r bir dille inceleyece\u011fiz.<\/p>\n
Makine \u00f6\u011frenimi, bilgisayarlara a\u00e7\u0131k\u00e7a programlanmadan \u00f6\u011frenme yetene\u011fi kazand\u0131ran bir yapay zeka dal\u0131d\u0131r. Temelde, algoritmalar b\u00fcy\u00fck veri k\u00fcmelerindeki \u00f6r\u00fcnt\u00fcleri ke\u015ffeder ve bu \u00f6r\u00fcnt\u00fcleri gelecekteki tahminler veya kararlar i\u00e7in kullan\u0131r. Bu algoritmalar genellikle denetimli \u00f6\u011frenme, denetimsiz \u00f6\u011frenme ve peki\u015ftirmeli \u00f6\u011frenme olmak \u00fczere \u00fc\u00e7 ana kategoriye ayr\u0131l\u0131r. Karar A\u011fa\u00e7lar\u0131, genellikle hem s\u0131n\u0131fland\u0131rma (\u00f6rne\u011fin, bir e-postan\u0131n spam olup olmad\u0131\u011f\u0131n\u0131 belirleme) hem de regresyon (\u00f6rne\u011fin, bir evin fiyat\u0131n\u0131 tahmin etme) g\u00f6revleri i\u00e7in kullan\u0131lan denetimli \u00f6\u011frenme algoritmalar\u0131 s\u0131n\u0131f\u0131na girer.<\/p>\n
Karar A\u011fa\u00e7lar\u0131, ad\u0131ndan da anla\u015f\u0131laca\u011f\u0131 gibi, bir dizi karar kural\u0131n\u0131 temsil eden bir a\u011fa\u00e7 yap\u0131s\u0131 olu\u015fturur. Bu a\u011fa\u00e7 yap\u0131s\u0131, bir sorunun \u00e7\u00f6z\u00fcm\u00fcne veya bir hedefin tahminine ula\u015fmak i\u00e7in ard\u0131\u015f\u0131k sorular sormaya benzer. \u00d6rne\u011fin, bir kredi ba\u015fvurusunu de\u011ferlendiren bir banka, ba\u015fvuran\u0131n gelirine, kredi ge\u00e7mi\u015fine ve mevcut bor\u00e7lar\u0131na g\u00f6re ard\u0131\u015f\u0131k kararlar alarak kredi verme veya reddetme sonucuna ula\u015fabilir. Karar a\u011fa\u00e7lar\u0131n\u0131n pop\u00fclaritesi, \u00f6zellikle yorumlanabilirlik ve anla\u015f\u0131labilirlik konusundaki \u00fcst\u00fcnl\u00fcklerinden kaynaklanmaktad\u0131r. Bir\u00e7ok “kara kutu” makine \u00f6\u011frenimi modelinin aksine, Karar A\u011fa\u00e7lar\u0131n\u0131n karar verme s\u00fcreci kolayca g\u00f6rselle\u015ftirilebilir ve insan taraf\u0131ndan anla\u015f\u0131labilir.<\/p>\n
Bir karar a\u011fac\u0131, bir ak\u0131\u015f \u015femas\u0131na benzer bir yap\u0131d\u0131r. Her i\u00e7 d\u00fc\u011f\u00fcm (internal node), bir \u00f6zellik \u00fczerinde bir test veya soru temsil ederken, her dal (branch) testin olas\u0131 sonu\u00e7lar\u0131ndan birini temsil eder. Her yaprak d\u00fc\u011f\u00fcm (leaf node) ise bir s\u0131n\u0131f etiketi (s\u0131n\u0131fland\u0131rma i\u00e7in) veya say\u0131sal bir de\u011fer (regresyon i\u00e7in) ile temsil edilen nihai karar\u0131 veya tahmini verir.<\/p>\n
* K\u00f6k D\u00fc\u011f\u00fcm (Root Node):<\/strong> A\u011fac\u0131n en \u00fcst\u00fcndeki d\u00fc\u011f\u00fcmd\u00fcr ve t\u00fcm veri setini temsil eder. \u0130lk karar veya test burada yap\u0131l\u0131r. Karar a\u011fa\u00e7lar\u0131n\u0131n nas\u0131l in\u015fa edildi\u011finin kalbi, veriyi en iyi \u015fekilde b\u00f6len \u00f6zellikleri ve b\u00f6lme noktalar\u0131n\u0131 se\u00e7me s\u00fcrecidir. Bu s\u00fcrece “b\u00f6lme” (splitting) denir ve d\u00fc\u011f\u00fcmlerin “safl\u0131\u011f\u0131n\u0131” (purity) art\u0131rmay\u0131 hedefler. Bir d\u00fc\u011f\u00fcm ne kadar safsa, o d\u00fc\u011f\u00fcmdeki \u00f6rneklerin \u00e7o\u011fu ayn\u0131 s\u0131n\u0131fa aittir.<\/p>\n S\u0131n\u0131fland\u0131rma problemleri i\u00e7in en yayg\u0131n kullan\u0131lan b\u00f6lme kriterleri Gini Safs\u0131zl\u0131\u011f\u0131 (Gini Impurity) ve Bilgi Kazanc\u0131 (Information Gain) ile Entropi’dir.<\/p>\n * Gini Safs\u0131zl\u0131\u011f\u0131 (Gini Impurity):<\/strong> * Entropi ve Bilgi Kazanc\u0131 (Entropy and Information Gain):<\/strong> Regresyon problemleri i\u00e7in, ama\u00e7 ortalama kare hata (Mean Squared Error – MSE) veya ortalama mutlak hata (Mean Absolute Error – MAE) gibi bir hata \u00f6l\u00e7\u00fcs\u00fcn\u00fc minimize etmektir. En yayg\u0131n kullan\u0131lan\u0131 MSE’dir.<\/p>\n * Ortalama Kare Hata (Mean Squared Error – MSE):<\/strong> Karar a\u011fa\u00e7lar\u0131 genellikle \u00f6zyinelemeli (recursive) ve a\u00e7g\u00f6zl\u00fc (greedy) bir yakla\u015f\u0131mla yukar\u0131dan a\u015fa\u011f\u0131ya do\u011fru in\u015fa edilir. Bu s\u00fcre\u00e7 genellikle CART (Classification and Regression Trees) algoritmas\u0131 gibi algoritmalar taraf\u0131ndan y\u00f6netilir.<\/p>\n 1. En \u0130yi B\u00f6lmeyi Bulma:<\/strong> A\u011fac\u0131n k\u00f6k d\u00fc\u011f\u00fcm\u00fcnden ba\u015flayarak, algoritma mevcut veri k\u00fcmesindeki her bir \u00f6zellik i\u00e7in olas\u0131 t\u00fcm b\u00f6lme noktalar\u0131n\u0131 de\u011ferlendirir. S\u0131n\u0131fland\u0131rma i\u00e7in Gini safs\u0131zl\u0131\u011f\u0131n\u0131 veya bilgi kazanc\u0131n\u0131 en \u00e7ok azaltan, regresyon i\u00e7in ise MSE’yi en \u00e7ok azaltan \u00f6zelli\u011fi ve b\u00f6lme noktas\u0131n\u0131 se\u00e7er. Bu a\u00e7g\u00f6zl\u00fc yakla\u015f\u0131m, her ad\u0131mda yerel olarak en iyi karar\u0131 verir, ancak bu, a\u011fac\u0131n genel olarak k\u00fcresel optimumu buldu\u011fu anlam\u0131na gelmez.<\/p>\n Karar a\u011fa\u00e7lar\u0131, basitlikleri ve esneklikleri nedeniyle bir\u00e7ok avantaj sunar:<\/p>\n * Yorumlanabilirlik ve Anla\u015f\u0131labilirlik:<\/strong> Karar a\u011fa\u00e7lar\u0131n\u0131n en b\u00fcy\u00fck avantaj\u0131, karar verme s\u00fcre\u00e7lerinin \u015feffaf olmas\u0131d\u0131r. Bir karar a\u011fac\u0131 g\u00f6rselle\u015ftirilebilir ve bir insan\u0131n kolayca takip edebilece\u011fi bir dizi “e\u011fer-o zaman” kural\u0131 olarak yorumlanabilir. Bu, “kara kutu” modellerin aksine, neden belirli bir tahminin yap\u0131ld\u0131\u011f\u0131n\u0131 anlamay\u0131 sa\u011flar. Avantajlar\u0131na ra\u011fmen, karar a\u011fa\u00e7lar\u0131n\u0131n baz\u0131 \u00f6nemli dezavantajlar\u0131 da vard\u0131r:<\/p>\n * A\u015f\u0131r\u0131 Uyum (Overfitting):<\/strong> Karar a\u011fa\u00e7lar\u0131n\u0131n en b\u00fcy\u00fck dezavantaj\u0131 a\u015f\u0131r\u0131 uyuma yatk\u0131n olmalar\u0131d\u0131r. A\u011fa\u00e7 \u00e7ok derin b\u00fcy\u00fcrse ve e\u011fitim verisindeki g\u00fcr\u00fclt\u00fcy\u00fc veya ayk\u0131r\u0131 de\u011ferleri \u00f6\u011frenirse, yeni, g\u00f6r\u00fclmemi\u015f veriler \u00fczerinde k\u00f6t\u00fc performans g\u00f6sterebilir. Bu, modelin e\u011fitim verisine \u00e7ok iyi uyum sa\u011flamas\u0131 ancak genelleme yetene\u011fini kaybetmesi anlam\u0131na gelir. A\u015f\u0131r\u0131 uyum, karar a\u011fa\u00e7lar\u0131 i\u00e7in ciddi bir sorun oldu\u011fundan, bu sorunu hafifletmek i\u00e7in “budama” (pruning) ad\u0131 verilen teknikler kullan\u0131l\u0131r. Budama, a\u011fac\u0131n karma\u015f\u0131kl\u0131\u011f\u0131n\u0131 azaltarak genelleme yetene\u011fini art\u0131rmay\u0131 ama\u00e7lar.<\/p>\n * \u00d6n Budama (Pre-pruning \/ Top-down pruning):<\/strong> A\u011fa\u00e7 olu\u015fturulurken b\u00f6lme i\u015flemini durdurarak a\u015f\u0131r\u0131 b\u00fcy\u00fcmeyi engellemeyi i\u00e7erir. Bunun i\u00e7in kullan\u0131lan yayg\u0131n kriterler \u015funlard\u0131r: * Son Budama (Post-pruning \/ Bottom-up pruning):<\/strong> Bu yakla\u015f\u0131mda, a\u011fa\u00e7 \u00f6nce m\u00fcmk\u00fcn oldu\u011funca tam olarak b\u00fcy\u00fct\u00fcl\u00fcr (genellikle durma kriterleri olmadan veya \u00e7ok gev\u015fek kriterlerle). Daha sonra, en alttaki dallardan ba\u015flayarak, her bir dal\u0131n kesilmesinin do\u011frulama seti \u00fczerindeki performans\u0131 nas\u0131l etkiledi\u011fi de\u011ferlendirilir. E\u011fer bir dal\u0131n kesilmesi modelin genelleme performans\u0131n\u0131 art\u0131r\u0131yorsa (veya \u00f6nemli \u00f6l\u00e7\u00fcde k\u00f6t\u00fcle\u015ftirmiyorsa), o dal kesilir ve ilgili d\u00fc\u011f\u00fcm bir yaprak d\u00fc\u011f\u00fcm haline gelir. Maliyet karma\u015f\u0131kl\u0131\u011f\u0131 budamas\u0131 (Cost-Complexity Pruning) veya en zay\u0131f ba\u011flant\u0131 budamas\u0131 (Weakest Link Pruning) gibi y\u00f6ntemler bu kategoriye girer. Son budama, genellikle \u00f6n budamadan daha iyi sonu\u00e7lar verir \u00e7\u00fcnk\u00fc a\u011fac\u0131n t\u00fcm potansiyel dallanmalar\u0131n\u0131 g\u00f6rme f\u0131rsat\u0131 sunar.<\/p>\n Karar a\u011fa\u00e7lar\u0131n\u0131n tek ba\u015f\u0131na baz\u0131 dezavantajlar\u0131 olsa da, “topluluk \u00f6\u011frenmesi” (ensemble learning) ad\u0131 verilen g\u00fc\u00e7l\u00fc bir teknikle birle\u015ftirildiklerinde inan\u0131lmaz derecede g\u00fc\u00e7l\u00fc ve esnek modeller haline gelirler. Topluluk \u00f6\u011frenmesi, birden fazla \u00f6\u011frenme algoritmas\u0131n\u0131n (veya ayn\u0131 algoritman\u0131n farkl\u0131 varyasyonlar\u0131n\u0131n) birle\u015ftirilerek tek bir modelden daha iyi tahminler yapmas\u0131n\u0131 sa\u011flayan bir yakla\u015f\u0131md\u0131r. Karar a\u011fa\u00e7lar\u0131, topluluk \u00f6\u011frenmesi algoritmalar\u0131 i\u00e7in temel yap\u0131 ta\u015f\u0131 olarak s\u0131kl\u0131kla kullan\u0131l\u0131r.<\/p>\n * Rastgele Ormanlar (Random Forests):<\/strong> * Boosting Algoritmalar\u0131 (Gradient Boosting, AdaBoost, XGBoost, LightGBM, CatBoost):<\/strong> Bu topluluk yakla\u015f\u0131mlar\u0131, Karar A\u011fa\u00e7lar\u0131n\u0131n yorumlanabilirlik avantaj\u0131ndan bir miktar \u00f6d\u00fcn verse de (\u00e7\u00fcnk\u00fc y\u00fczlerce a\u011fac\u0131 yorumlamak zordur), genelleme performans\u0131n\u0131 ve do\u011frulu\u011funu dramatik bir \u015fekilde art\u0131r\u0131r.<\/p>\n Karar a\u011fa\u00e7lar\u0131 ve onlara dayal\u0131 topluluk modelleri, \u00e7ok \u00e7e\u015fitli sekt\u00f6rlerde ve problemlerde yayg\u0131n olarak kullan\u0131lmaktad\u0131r:<\/p>\n * T\u0131bbi Te\u015fhis:<\/strong> Hastalar\u0131n semptomlar\u0131na ve t\u0131bbi test sonu\u00e7lar\u0131na dayanarak hastal\u0131klar\u0131 te\u015fhis etmek. Karar A\u011fa\u00e7lar\u0131, makine \u00f6\u011frenimi alan\u0131nda temel, g\u00fc\u00e7l\u00fc ve anla\u015f\u0131lmas\u0131 kolay algoritmalardan biridir. Bir ak\u0131\u015f \u015femas\u0131na benzeyen sezgisel yap\u0131lar\u0131, onlar\u0131 hem yeni ba\u015flayanlar i\u00e7in ideal bir ba\u015flang\u0131\u00e7 noktas\u0131 hem de deneyimli veri bilimcileri i\u00e7in de\u011ferli bir ara\u00e7 haline getirir. Yorumlanabilirlikleri, bir\u00e7ok i\u015f karar\u0131 i\u00e7in kritik \u00f6neme sahip olan modelin nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 anlamam\u0131z\u0131 sa\u011flar.<\/p>\n Tek ba\u015f\u0131na kullan\u0131ld\u0131klar\u0131nda a\u015f\u0131r\u0131 uyum ve karars\u0131zl\u0131k gibi dezavantajlar\u0131 olsa da, budama teknikleri ile bu sorunlar hafifletilebilir. Dahas\u0131, Rastgele Ormanlar ve Boosting gibi topluluk \u00f6\u011frenmesi y\u00f6ntemleriyle birle\u015ftirildiklerinde, Karar A\u011fa\u00e7lar\u0131 inan\u0131lmaz derecede y\u00fcksek performansl\u0131 ve sa\u011flam modeller olu\u015fturabilirler. G\u00fcn\u00fcm\u00fcz\u00fcn en geli\u015fmi\u015f makine \u00f6\u011frenimi algoritimlerinin bir\u00e7o\u011funun temelinde karar a\u011fa\u00e7lar\u0131 yatmaktad\u0131r.<\/p>\n Veri bilimcileri i\u00e7in Karar A\u011fa\u00e7lar\u0131n\u0131 anlamak, sadece bu \u00f6zel algoritmay\u0131 de\u011fil, ayn\u0131 zamanda makine \u00f6\u011freniminin temel prensiplerini ve daha karma\u015f\u0131k modellerin nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 kavramak i\u00e7in de sa\u011flam bir temel olu\u015fturur. Basitlik ve g\u00fcc\u00fcn m\u00fckemmel bir birle\u015fimi olan Karar A\u011fa\u00e7lar\u0131, \u015f\u00fcphesiz makine \u00f6\u011frenimi ara\u00e7 kutusunun vazge\u00e7ilmez bir par\u00e7as\u0131 olmaya devam edecektir.<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"Karar A\u011fa\u00e7lar\u0131 Basitle\u015ftirildi: Makine \u00d6\u011frenimi A\u00e7\u0131klamas\u0131\nMakine \u00f6\u011frenimi, g\u00fcn\u00fcm\u00fcz\u00fcn en heyecan verici ve d\u00f6n\u00fc\u015ft\u00fcr\u00fcc\u00fc teknoloji","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1],"tags":[],"class_list":{"0":"post-33461","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\n
\n* \u0130\u00e7 D\u00fc\u011f\u00fcm (Internal\/Decision Node):<\/strong> K\u00f6k d\u00fc\u011f\u00fcm d\u0131\u015f\u0131ndaki t\u00fcm dallanma d\u00fc\u011f\u00fcmleridir. Bir \u00f6zelli\u011fi test eder ve veriyi daha k\u00fc\u00e7\u00fck alt k\u00fcmelere b\u00f6ler.
\n* Yaprak D\u00fc\u011f\u00fcm (Leaf Node):<\/strong> A\u011fac\u0131n son d\u00fc\u011f\u00fcmleridir ve daha fazla b\u00f6l\u00fcnemezler. Nihai karar\u0131 veya tahmini temsil ederler.<\/p>\nB\u00f6lme Kriterleri<\/h4>\n
S\u0131n\u0131fland\u0131rma A\u011fa\u00e7lar\u0131 \u0130\u00e7in B\u00f6lme Kriterleri<\/h5>\n
\n Gini safs\u0131zl\u0131\u011f\u0131, bir d\u00fc\u011f\u00fcmdeki rastgele se\u00e7ilen bir \u00f6rne\u011fin yanl\u0131\u015f s\u0131n\u0131fland\u0131r\u0131lma olas\u0131l\u0131\u011f\u0131n\u0131n bir \u00f6l\u00e7\u00fcs\u00fcd\u00fcr. D\u00fc\u015f\u00fck Gini safs\u0131zl\u0131\u011f\u0131, d\u00fc\u011f\u00fcm\u00fcn daha saf oldu\u011fu anlam\u0131na gelir. Bir d\u00fc\u011f\u00fcmdeki t\u00fcm \u00f6rnekler ayn\u0131 s\u0131n\u0131fa aitse, Gini safs\u0131zl\u0131\u011f\u0131 0’d\u0131r (maksimum safl\u0131k).
\n Form\u00fcl\u00fc genellikle \u015f\u00f6yledir:
\n $Gini(p) = 1 – \\sum_{i=1}^{c} p_i^2$
\n Burada $p_i$, bir d\u00fc\u011f\u00fcmdeki \u00f6rneklerin $i$ s\u0131n\u0131f\u0131na ait olma olas\u0131l\u0131\u011f\u0131d\u0131r ve $c$ s\u0131n\u0131f say\u0131s\u0131d\u0131r. Algoritma, her b\u00f6l\u00fcnmede Gini safs\u0131zl\u0131\u011f\u0131n\u0131 en \u00e7ok azaltan \u00f6zelli\u011fi ve b\u00f6lme noktas\u0131n\u0131 se\u00e7meye \u00e7al\u0131\u015f\u0131r.<\/p>\n
\n Entropi, bir d\u00fc\u011f\u00fcmdeki veri k\u00fcmesinin d\u00fczensizli\u011finin veya belirsizli\u011finin bir \u00f6l\u00e7\u00fcs\u00fcd\u00fcr. Gini safs\u0131zl\u0131\u011f\u0131 gibi, Entropi de 0 oldu\u011funda maksimum safl\u0131\u011f\u0131 g\u00f6sterir. Bir d\u00fc\u011f\u00fcmdeki \u00f6rnekler farkl\u0131 s\u0131n\u0131flara e\u015fit olarak da\u011f\u0131lm\u0131\u015fsa, Entropi maksimum de\u011ferine ula\u015f\u0131r.
\n Form\u00fcl\u00fc genellikle \u015f\u00f6yledir:
\n $Entropy(p) = – \\sum_{i=1}^{c} p_i \\log_2(p_i)$
\n Bilgi Kazanc\u0131 ise, bir \u00f6zellik \u00fczerinde b\u00f6lme yap\u0131ld\u0131\u011f\u0131nda Entropi’deki azalmay\u0131 \u00f6l\u00e7er. Algoritma, her ad\u0131mda bilgi kazanc\u0131n\u0131 en \u00fcst d\u00fczeye \u00e7\u0131karan \u00f6zelli\u011fi se\u00e7er.
\n $Information Gain(D, A) = Entropy(D) – \\sum_{v \\in Values(A)} \\frac{|D_v|}{|D|} Entropy(D_v)$
\n Burada $D$ ana veri k\u00fcmesi, $A$ b\u00f6lme yap\u0131lan \u00f6zellik, $Values(A)$ \u00f6zelli\u011fin olas\u0131 de\u011ferleri, $|D_v|$ \u00f6zelli\u011fin $v$ de\u011ferine sahip \u00f6rnek say\u0131s\u0131 ve $|D|$ toplam \u00f6rnek say\u0131s\u0131d\u0131r.<\/p>\nRegresyon A\u011fa\u00e7lar\u0131 \u0130\u00e7in B\u00f6lme Kriterleri<\/h5>\n
\n MSE, bir d\u00fc\u011f\u00fcmdeki ger\u00e7ek de\u011ferler ile o d\u00fc\u011f\u00fcmdeki \u00f6rneklerin ortalama de\u011feri aras\u0131ndaki farklar\u0131n karelerinin ortalamas\u0131d\u0131r. Ama\u00e7, her b\u00f6l\u00fcnmede MSE’yi en \u00e7ok azaltan \u00f6zelli\u011fi ve b\u00f6lme noktas\u0131n\u0131 bulmakt\u0131r. D\u00fc\u011f\u00fcmdeki t\u00fcm \u00f6rneklerin hedef de\u011ferleri ayn\u0131ysa, MSE 0’d\u0131r.
\n Form\u00fcl\u00fc genellikle \u015f\u00f6yledir:
\n $MSE = \\frac{1}{N} \\sum_{i=1}^{N} (y_i – \\hat{y})^2$
\n Burada $N$ d\u00fc\u011f\u00fcmdeki \u00f6rnek say\u0131s\u0131, $y_i$ ger\u00e7ek de\u011fer ve $\\hat{y}$ d\u00fc\u011f\u00fcmdeki \u00f6rneklerin ortalama hedef de\u011feridir.<\/p>\nKarar A\u011fa\u00e7lar\u0131 Nas\u0131l Olu\u015fturulur?<\/h3>\n
\n2. D\u00fc\u011f\u00fcm\u00fc B\u00f6lme:<\/strong> Se\u00e7ilen en iyi b\u00f6lme noktas\u0131na g\u00f6re veri k\u00fcmesi iki veya daha fazla alt k\u00fcmeye ayr\u0131l\u0131r ve bu alt k\u00fcmeler i\u00e7in yeni alt d\u00fc\u011f\u00fcmler olu\u015fturulur.
\n3. \u00d6zyinelemeli S\u00fcre\u00e7:<\/strong> Bu s\u00fcre\u00e7, her bir alt d\u00fc\u011f\u00fcm i\u00e7in \u00f6zyinelemeli olarak tekrarlan\u0131r. Her alt d\u00fc\u011f\u00fcm kendi i\u00e7inde bir veri k\u00fcmesi olarak ele al\u0131n\u0131r ve en iyi b\u00f6lme noktas\u0131 aran\u0131r.
\n4. Durma Kriterleri:<\/strong> A\u011fa\u00e7 sonsuza kadar b\u00fcy\u00fcyemez. Belirli durma kriterlerine ula\u015f\u0131ld\u0131\u011f\u0131nda b\u00f6lme i\u015flemi durur ve d\u00fc\u011f\u00fcm bir yaprak d\u00fc\u011f\u00fcm haline gelir. Yayg\u0131n durma kriterleri \u015funlard\u0131r:
\n * Bir d\u00fc\u011f\u00fcmdeki t\u00fcm \u00f6rnekler ayn\u0131 s\u0131n\u0131fa aitse (s\u0131n\u0131fland\u0131rma) veya ayn\u0131 de\u011fere sahipse (regresyon).
\n * A\u011fac\u0131n belirli bir maksimum derinli\u011fe ula\u015fmas\u0131.
\n * Bir d\u00fc\u011f\u00fcmdeki \u00f6rnek say\u0131s\u0131n\u0131n belirli bir minimum e\u015fi\u011fin alt\u0131na d\u00fc\u015fmesi.
\n * B\u00f6lme i\u015fleminden elde edilen bilgi kazanc\u0131n\u0131n veya safs\u0131zl\u0131k azalmas\u0131n\u0131n belirli bir e\u015fi\u011fin alt\u0131nda kalmas\u0131 (yani, daha fazla b\u00f6lmenin \u00f6nemli bir iyile\u015fme sa\u011flamamas\u0131).<\/p>\nKarar A\u011fa\u00e7lar\u0131n\u0131n Avantajlar\u0131<\/h3>\n
\n* Veri \u00d6n \u0130\u015fleme \u0130htiyac\u0131n\u0131n Azalmas\u0131:<\/strong> Karar a\u011fa\u00e7lar\u0131, veri normalizasyonu, standartizasyonu veya \u00f6l\u00e7eklendirmesi gibi yo\u011fun \u00f6n i\u015fleme ad\u0131mlar\u0131na genellikle ihtiya\u00e7 duymazlar. Kategorik ve say\u0131sal \u00f6zellikleri ayn\u0131 anda i\u015fleyebilirler.
\n* Non-lineer \u0130li\u015fkileri Yakalayabilme:<\/strong> Lineer modellere k\u0131yasla, karar a\u011fa\u00e7lar\u0131 karma\u015f\u0131k non-lineer ili\u015fkileri ve etkile\u015fimleri veri i\u00e7inde do\u011fal olarak modelleyebilirler.
\n* H\u0131zl\u0131 E\u011fitim ve Tahmin:<\/strong> Genellikle di\u011fer karma\u015f\u0131k modellere g\u00f6re daha h\u0131zl\u0131 e\u011fitilirler ve tahmin yaparken de olduk\u00e7a h\u0131zl\u0131d\u0131rlar.
\n* \u00c7oklu \u00c7\u0131k\u0131\u015f Problemlerini Destekleme:<\/strong> Tek bir modelle birden fazla hedef de\u011fi\u015fkeni tahmin etmek m\u00fcmk\u00fcnd\u00fcr.
\n* Ayk\u0131r\u0131 De\u011ferlere Kar\u015f\u0131 Dayan\u0131kl\u0131l\u0131k:<\/strong> B\u00f6lme kriterleri genellikle veri s\u0131ralamas\u0131na dayand\u0131\u011f\u0131 i\u00e7in, ayk\u0131r\u0131 de\u011ferlerin a\u011fac\u0131n yap\u0131s\u0131 \u00fczerindeki etkisi di\u011fer baz\u0131 algoritmalara g\u00f6re daha az olabilir.<\/p>\nKarar A\u011fa\u00e7lar\u0131n\u0131n Dezavantajlar\u0131<\/h3>\n
\n* Karars\u0131zl\u0131k (Instability):<\/strong> E\u011fitim verisindeki k\u00fc\u00e7\u00fck de\u011fi\u015fiklikler (\u00f6rne\u011fin, birka\u00e7 \u00f6rne\u011fin eklenmesi veya \u00e7\u0131kar\u0131lmas\u0131), a\u011fac\u0131n yap\u0131s\u0131nda \u00f6nemli de\u011fi\u015fikliklere yol a\u00e7abilir. Bu, modelin olduk\u00e7a de\u011fi\u015fken oldu\u011fu anlam\u0131na gelir.
\n* Lokal Optimum:<\/strong> Karar a\u011fa\u00e7lar\u0131 a\u00e7g\u00f6zl\u00fc bir yakla\u015f\u0131mla in\u015fa edildi\u011fi i\u00e7in, her ad\u0131mda yerel olarak en iyi b\u00f6lmeyi se\u00e7erler. Bu, genel olarak k\u00fcresel optimum \u00e7\u00f6z\u00fcm\u00fc bulamad\u0131klar\u0131 anlam\u0131na gelebilir.
\n* Dengesiz Verilerle Sorunlar:<\/strong> E\u011fer s\u0131n\u0131flar aras\u0131nda b\u00fcy\u00fck bir dengesizlik varsa (bir s\u0131n\u0131f di\u011ferlerinden \u00e7ok daha az \u00f6rne\u011fe sahipse), karar a\u011fac\u0131 az\u0131nl\u0131k s\u0131n\u0131f\u0131n\u0131 g\u00f6z ard\u0131 etme e\u011filiminde olabilir.
\n* Y\u00fcksek Varyans:<\/strong> Tek bir karar a\u011fac\u0131 genellikle y\u00fcksek varyansa sahiptir, yani e\u011fitim verisindeki k\u00fc\u00e7\u00fck de\u011fi\u015fikliklere kar\u015f\u0131 \u00e7ok hassast\u0131r.<\/p>\nA\u015f\u0131r\u0131 Uyumla Ba\u015fa \u00c7\u0131kma: Budama (Pruning)<\/h3>\n
\n * Maksimum a\u011fa\u00e7 derinli\u011fi belirleme.
\n * Bir d\u00fc\u011f\u00fcmdeki minimum \u00f6rnek say\u0131s\u0131n\u0131 belirleme (bu say\u0131n\u0131n alt\u0131na d\u00fc\u015ferse b\u00f6l\u00fcnmez).
\n * B\u00f6lme sonras\u0131 elde edilen bilgi kazanc\u0131n\u0131n veya safs\u0131zl\u0131k azalmas\u0131n\u0131n belirli bir e\u015fi\u011fin alt\u0131nda kalmas\u0131 durumunda b\u00f6lmeyi durdurma.
\n \u00d6n budama, a\u011fac\u0131n daha basit ve daha h\u0131zl\u0131 olu\u015fmas\u0131n\u0131 sa\u011flar, ancak bazen en iyi a\u011fac\u0131 bulmay\u0131 engeller \u00e7\u00fcnk\u00fc baz\u0131 potansiyel olarak faydal\u0131 dallar erken kesilebilir.<\/p>\nTopluluk \u00d6\u011frenmesi ve Karar A\u011fa\u00e7lar\u0131<\/h3>\n
\n Rastgele Ormanlar, en pop\u00fcler ve etkili topluluk \u00f6\u011frenmesi algoritmalar\u0131ndan biridir. Bir\u00e7ok karar a\u011fac\u0131n\u0131n bir araya gelmesiyle olu\u015fur. Her a\u011fa\u00e7, e\u011fitim verisinin rastgele bir alt k\u00fcmesi \u00fczerinde ve \u00f6zelliklerin rastgele bir alt k\u00fcmesi kullan\u0131larak e\u011fitilir. Bu iki rastgelelik boyutu (veri \u00f6rneklemesi ve \u00f6zellik \u00f6rneklemesi), a\u011fa\u00e7lar\u0131n birbirlerinden ba\u011f\u0131ms\u0131z olmas\u0131n\u0131 ve a\u015f\u0131r\u0131 uyumu azaltmas\u0131n\u0131 sa\u011flar. Tahmin yaparken, s\u0131n\u0131fland\u0131rma i\u00e7in a\u011fa\u00e7lar\u0131n \u00e7o\u011funluk oyu al\u0131n\u0131r, regresyon i\u00e7in ise tahminlerin ortalamas\u0131 al\u0131n\u0131r. Rastgele Ormanlar, tek bir karar a\u011fac\u0131n\u0131n y\u00fcksek varyans sorununu \u00e7\u00f6zerken, modelin genelleme yetene\u011fini \u00f6nemli \u00f6l\u00e7\u00fcde art\u0131r\u0131r.<\/p>\n
\n Boosting, zay\u0131f \u00f6\u011frenicileri (genellikle s\u0131\u011f karar a\u011fa\u00e7lar\u0131) art arda e\u011fiterek g\u00fc\u00e7l\u00fc bir \u00f6\u011frenici olu\u015fturan ba\u015fka bir topluluk \u00f6\u011frenmesi tekni\u011fidir. Boosting algoritmalar\u0131, her yeni a\u011fac\u0131 \u00f6nceki a\u011fa\u00e7lar\u0131n yanl\u0131\u015f yapt\u0131\u011f\u0131 \u00f6rnekleri daha iyi tahmin etmeye odaklanarak e\u011fitir. Bu, modelin hatalar\u0131n\u0131 kademeli olarak d\u00fczeltmesini ve performans\u0131 art\u0131rmas\u0131n\u0131 sa\u011flar. Gradient Boosting Machine (GBM), AdaBoost, ve bunlar\u0131n daha geli\u015fmi\u015f versiyonlar\u0131 olan XGBoost, LightGBM ve CatBoost, makine \u00f6\u011frenimi yar\u0131\u015fmalar\u0131nda ve ger\u00e7ek d\u00fcnya uygulamalar\u0131nda s\u0131kl\u0131kla en iyi performans\u0131 g\u00f6steren algoritmalar aras\u0131ndad\u0131r ve hepsi temel olarak karar a\u011fa\u00e7lar\u0131n\u0131 kullan\u0131r.<\/p>\nKarar A\u011fa\u00e7lar\u0131n\u0131n Uygulama Alanlar\u0131<\/h3>\n
\n* M\u00fc\u015fteri Segmentasyonu:<\/strong> M\u00fc\u015fterileri sat\u0131n alma davran\u0131\u015flar\u0131, demografik \u00f6zellikleri veya ilgi alanlar\u0131na g\u00f6re grupland\u0131rmak.
\n* Kredi Riski De\u011ferlendirmesi:<\/strong> Bir ba\u015fvuran\u0131n kredi \u00f6deme olas\u0131l\u0131\u011f\u0131n\u0131 tahmin etmek.
\n* Pazarlama Analizi:<\/strong> Bir m\u00fc\u015fterinin belirli bir \u00fcr\u00fcn\u00fc sat\u0131n alma veya bir kampanyaya yan\u0131t verme olas\u0131l\u0131\u011f\u0131n\u0131 tahmin etmek.
\n* Doland\u0131r\u0131c\u0131l\u0131k Tespiti:<\/strong> Finansal i\u015flemlerde veya sigorta taleplerinde doland\u0131r\u0131c\u0131l\u0131k faaliyetlerini belirlemek.
\n* Do\u011fal Dil \u0130\u015fleme:<\/strong> Metin s\u0131n\u0131fland\u0131rmas\u0131 veya duygu analizi gibi g\u00f6revlerde kullan\u0131labilirler.
\n* G\u00f6r\u00fcnt\u00fc \u0130\u015fleme:<\/strong> Nesne tan\u0131ma veya g\u00f6r\u00fcnt\u00fc s\u0131n\u0131fland\u0131rmas\u0131 gibi temel g\u00f6revlerde \u00f6zellik vekt\u00f6rleriyle birlikte kullan\u0131labilirler.<\/p>\nSonu\u00e7<\/h3>\n