<\/p>\n
Genellikle makine \u00f6\u011frenimi s\u0131n\u0131fland\u0131rma algoritmalar\u0131yla<\/a> \u00e7al\u0131\u015f\u0131rken veri k\u00fcmesindeki s\u0131n\u0131flar dengesiz olacakt\u0131r.<\/span><\/span><\/p>\n \u00d6rne\u011fin:<\/span><\/p>\n Bu dengesiz s\u0131n\u0131flar nedeniyle olu\u015fturdu\u011funuz tahmine dayal\u0131 model az\u0131nl\u0131k s\u0131n\u0131f\u0131nda iyi \u00e7al\u0131\u015fmayabilir.<\/span><\/p>\n Daha da k\u00f6t\u00fcs\u00fc, \u00e7o\u011funlukla tahmin etmeyi en \u00e7ok istedi\u011fimiz s\u0131n\u0131f az\u0131nl\u0131k s\u0131n\u0131f\u0131d\u0131r.<\/span><\/p>\n Bu dengesizlik sorununu \u00e7\u00f6zmenin bir yolu, genellikle SMOTE<\/strong> olarak k\u0131salt\u0131lan sentetik az\u0131nl\u0131k a\u015f\u0131r\u0131 \u00f6rnekleme tekni\u011fini<\/strong> kullanmakt\u0131r.<\/span><\/p>\n Bu teknik, daha dengeli s\u0131n\u0131flara sahip bir veri k\u00fcmesi \u00fcreten az\u0131nl\u0131k s\u0131n\u0131f\u0131ndan g\u00f6zlemlerin a\u015f\u0131r\u0131 \u00f6rneklenmesiyle yeni bir veri k\u00fcmesi olu\u015fturulmas\u0131n\u0131 i\u00e7erir.<\/span><\/p>\n R’de SMOTE kullanman\u0131n en kolay yolu DMwR<\/strong> paketindeki SMOTE()<\/strong> i\u015flevini kullanmakt\u0131r.<\/span><\/p>\n Bu i\u015flev a\u015fa\u011f\u0131daki temel s\u00f6zdizimini kullan\u0131r:<\/span><\/span><\/p>\n Alt\u0131n:<\/span><\/p>\n A\u015fa\u011f\u0131daki \u00f6rnekte bu fonksiyonun pratikte nas\u0131l kullan\u0131laca\u011f\u0131 g\u00f6sterilmektedir.<\/span><\/p>\n Yan\u0131t de\u011fi\u015fkeni i\u00e7in 90’\u0131n\u0131n “Evet” s\u0131n\u0131f\u0131na ve 10’unun “Hay\u0131r” s\u0131n\u0131f\u0131na sahip oldu\u011fu, R’de 100 g\u00f6zlem<\/a> i\u00e7eren a\u015fa\u011f\u0131daki veri setine sahip oldu\u011fumuzu varsayal\u0131m:<\/span><\/p>\n Bu, dengesiz bir veri k\u00fcmesinin klasik bir \u00f6rne\u011fidir \u00e7\u00fcnk\u00fc tahmin etti\u011fimiz yan\u0131t de\u011fi\u015fkeni, “Evet” s\u0131n\u0131f\u0131na sahip 90 g\u00f6zleme ve “Hay\u0131r” s\u0131n\u0131f\u0131na sahip yaln\u0131zca 10 g\u00f6zleme sahiptir.<\/span><\/span><\/p>\n Daha dengeli bir veri k\u00fcmesi olu\u015fturmak i\u00e7in DMwR<\/strong> paketindeki SMOTE()<\/strong> i\u015flevini kullanabiliriz:<\/span><\/p>\n Ortaya \u00e7\u0131kan veri seti, s\u0131n\u0131f olarak \u201cHay\u0131r\u201d olan 210 g\u00f6zlem ve s\u0131n\u0131f olarak \u201cEvet\u201d olan 800 g\u00f6zlem i\u00e7ermektedir.<\/span><\/p>\n SMOTE fonksiyonunun bu yeni veri setini tam olarak nas\u0131l \u00fcretti\u011fi a\u015fa\u011f\u0131da a\u00e7\u0131klanm\u0131\u015ft\u0131r:<\/span><\/p>\n Nihai sonu\u00e7, h\u00e2l\u00e2 daha fazla \u00e7o\u011funluk s\u0131n\u0131f\u0131 i\u00e7eren ancak yine de orijinal veri k\u00fcmesinden daha dengeli bir veri k\u00fcmesidir.<\/span><\/p>\n Art\u0131k se\u00e7ti\u011finiz s\u0131n\u0131fland\u0131rma algoritmas\u0131n\u0131 bu yeni veri k\u00fcmesine uyarlayabilirsiniz; bu yeni veri k\u00fcmesinde az\u0131nl\u0131k s\u0131n\u0131f\u0131ndan daha fazla g\u00f6zlem oldu\u011fundan az\u0131nl\u0131k s\u0131n\u0131f\u0131nda daha iyi performans g\u00f6stermesi gerekir.<\/span><\/p>\n Not<\/strong> : \u0130htiya\u00e7lar\u0131n\u0131za uygun bir veri k\u00fcmesi elde etmek i\u00e7in SMOTE i\u015flevindeki perc.over<\/strong> ve perc.under<\/strong> arg\u00fcmanlar\u0131yla oynamaktan \u00e7ekinmeyin.<\/span><\/p>\n A\u015fa\u011f\u0131daki e\u011fitimlerde R’de di\u011fer ortak g\u00f6revlerin nas\u0131l ger\u00e7ekle\u015ftirilece\u011fi a\u00e7\u0131klanmaktad\u0131r:<\/span><\/p>\n R’de \u00f6zet tablolar nas\u0131l olu\u015fturulur<\/a> Genellikle makine \u00f6\u011frenimi s\u0131n\u0131fland\u0131rma algoritmalar\u0131yla \u00e7al\u0131\u015f\u0131rken veri k\u00fcmesindeki s\u0131n\u0131flar dengesiz olacakt\u0131r. \u00d6rne\u011fin: \u00dcniversite oyuncular\u0131n\u0131n NBA’e se\u00e7ilip se\u00e7ilmedi\u011fine ili\u015fkin bilgi i\u00e7eren bir veri k\u00fcmesinde oyuncular\u0131n %98’i se\u00e7ilmemi\u015f ve %2’si draft edilmi\u015f olabilir. Hastalar\u0131n kanser olup olmad\u0131\u011f\u0131na dair bilgi i\u00e7eren bir veri seti, kanserli olmayan hastalar\u0131n %99’unu ve kanserli hastalar\u0131n yaln\u0131zca %1’ini i\u00e7erebilir. Banka doland\u0131r\u0131c\u0131l\u0131\u011f\u0131 bilgilerini i\u00e7eren […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11],"tags":[],"class_list":["post-3061","post","type-post","status-publish","format-standard","hentry","category-rehber"],"yoast_head":"\n\n
SMOTE(form, data, perc. over<\/span> = 200<\/span> , perc. under<\/span> = 200<\/span> , ...)\n<\/strong><\/pre>\n\n
\u00d6rnek: R’de SMOTE nas\u0131l kullan\u0131l\u0131r?<\/strong><\/span><\/h3>\n
#make this example reproducible\n<\/span>set. seed<\/span> ( 0<\/span> )\n\n#create data frame with one response variable and two predictor variables\n<\/span>df <- data. frame<\/span> (y=rep(as. factor<\/span> (c(' Yes<\/span> ', ' No<\/span> ')), times=c( 90<\/span> , 10<\/span> )),\n x1=rnorm( 100<\/span> ),\n x2=rnorm( 100<\/span> ))\n\n#view first six rows of data frame\n<\/span>head(df)\n\n y x1 x2\n1 Yes 1.2629543 0.7818592\n2 Yes -0.3262334 -0.7767766\n3 Yes 1.3297993 -0.6159899\n4 Yes 1.2724293 0.0465803\n5 Yes 0.4146414 -1.1303858\n6 Yes -1.5399500 0.5767188\n \n#view distribution of response variable\n<\/span>table(df$y)\n\n Drowned \n 10 90<\/strong><\/pre>\n library<\/span> ( DMwR)\n\n<\/span>#use SMOTE to create new dataset that is more balanced\n<\/span>new_df <- SMOTE(y ~ ., df, perc. over<\/span> = 2000<\/span> , perc. under<\/span> = 400<\/span> )\n\n#view distribution of response variable in new dataset\n<\/span>table(new_df$y)\n\n Drowned \n210,800\n<\/strong><\/pre>\n\n
Ek kaynaklar<\/strong><\/span><\/h3>\n
R’deki veriler nas\u0131l normalle\u015ftirilir<\/a>
R’de ayk\u0131r\u0131 de\u011ferler nas\u0131l kald\u0131r\u0131l\u0131r<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"