我想绘制一个混淆矩阵,但是,我不想只使用热图,因为我认为它们的数值分辨率很差.相反,我还想在正方形的中间绘制频率.例如,我喜欢这个的输出:
library(mlearning); data("Glass", package = "mlbench") Glass$Type <- as.factor(paste("Glass", Glass$Type)) summary(glassLvq <- mlLvq(Type ~ ., data = Glass)); (glassConf <- confusion(predict(glassLvq, Glass, type = "class"), Glass$Type)) plot(glassConf) # Image by default
但是,1.)我不明白"01,02等"是指每个轴.我们怎样才能摆脱这种局面?2.)我希望'Predicted'作为'y'维度的标签,'Actual'作为'x'维度的标签3.)我想用频率/概率替换绝对计数.
或者,是否有其他包可以做到这一点?
从本质上讲,我希望在R中:
http://www.mathworks.com/help/releases/R2013b/nnet/gs/gettingstarted_nprtool_07.gif
要么:
http://c431376.r76.cf2.rackcdn.com/8805/fnhum-05-00189-HTML/image_m/fnhum-05-00189-g009.jpg
该mlearning
封装似乎非常灵活,可以绘制混淆矩阵.
从您的glassConf
对象开始,您可能希望执行以下操作:
prior(glassConf) <- 100 # The above rescales the confusion matrix such that columns sum to 100. opar <- par(mar=c(5.1, 6.1, 2, 2)) x <- x.orig <- unclass(glassConf) x <- log(x + 0.5) * 2.33 x[x < 0] <- NA x[x > 10] <- 10 diag(x) <- -diag(x) image(1:ncol(x), 1:ncol(x), -(x[, nrow(x):1]), xlab='Actual', ylab='', col=colorRampPalette(c(hsv(h = 0, s = 0.9, v = 0.9, alpha = 1), hsv(h = 0, s = 0, v = 0.9, alpha = 1), hsv(h = 2/6, s = 0.9, v = 0.9, alpha = 1)))(41), xaxt='n', yaxt='n', zlim=c(-10, 10)) axis(1, at=1:ncol(x), labels=colnames(x), cex.axis=0.8) axis(2, at=ncol(x):1, labels=colnames(x), las=1, cex.axis=0.8) title(ylab='Predicted', line=4.5) abline(h = 0:ncol(x) + 0.5, col = 'gray') abline(v = 0:ncol(x) + 0.5, col = 'gray') text(1:6, rep(6:1, each=6), labels = sub('^0$', '', round(c(x.orig), 0))) box(lwd=2) par(opar) # reset par
上面的代码使用了confusionImage
被调用的函数的位plot.confusion
.