I'm using the mlr package's framework to build a svm model to predict landcover classes in an image. I used the raster package's predict function and also converted the raster to a dataframe and then predicted on that dataframe using the "learner.model" as input. These methods gave me realistic results.
Work well:
> predict(raster, mod$learner.model)
or
> xy <- as.data.frame(raster, xy = T)
> C <- predict(mod$learner.model, xy)
However, if I predict on the dataframe derived from the raster without specifying the learner.model, the results are not the same.
> C2 <- predict(mod, newdata=xy)
C2$data$response is not identical to C. Why?
Here is a reproducible example that demonstrates the problem:
> library(mlr)
> library(kernlab)
> x1 <- rnorm(50)
> x2 <- rnorm(50, 3)
> x3 <- rnorm(50, -20, 3)
> C <- sample(c("a","b","c"), 50, T)
> d <- data.frame(x1, x2, x3, C)
> classif <- makeClassifTask(id = "example", data = d, target = "C")
> lrn <- makeLearner("classif.ksvm", predict.type = "prob", fix.factors.prediction = T)
> t <- train(lrn, classif)
Using automatic sigma estimation (sigest) for RBF or laplace kernel
> res1 <- predict(t, newdata = data.frame(x2,x1,x3))
> res1
Prediction: 50 observations
predict.type: prob
threshold: a=0.33,b=0.33,c=0.33
time: 0.01
prob.a prob.b prob.c response
1 0.2110131 0.3817773 0.4072095 c
2 0.1551583 0.4066868 0.4381549 c
3 0.4305353 0.3092737 0.2601910 a
4 0.2160050 0.4142465 0.3697485 b
5 0.1852491 0.3789849 0.4357659 c
6 0.5879579 0.2269832 0.1850589 a
> res2 <- predict(t$learner.model, data.frame(x2,x1,x3))
> res2
[1] c c a b c a b a c c b c b a c b c a a b c b c c a b b b a a b a c b a c c c
[39] c a a b c b b b b a b b
Levels: a b c
!> res1$data$response == res2
[1] TRUE TRUE TRUE TRUE TRUE TRUE FALSE TRUE TRUE TRUE TRUE FALSE
[13] TRUE TRUE TRUE FALSE TRUE TRUE TRUE FALSE TRUE TRUE TRUE TRUE
[25] TRUE TRUE TRUE TRUE TRUE TRUE FALSE TRUE TRUE TRUE TRUE TRUE
[37] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[49] TRUE FALSE
The predictions are not identical. Following mlr's tutorial page on prediction, I don't see why the results would differ. Thanks for your help.
Update: When I do the same with a random forest model, the two vectors are equal. Is this because SVM is scale dependent and random forest is not?
> library(randomForest)
> classif <- makeClassifTask(id = "example", data = d, target = "C")
> lrn <- makeLearner("classif.randomForest", predict.type = "prob", fix.factors.prediction = T)
> t <- train(lrn, classif)
>
> res1 <- predict(t, newdata = data.frame(x2,x1,x3))
> res1
Prediction: 50 observations
predict.type: prob
threshold: a=0.33,b=0.33,c=0.33
time: 0.00
prob.a prob.b prob.c response
1 0.654 0.228 0.118 a
2 0.742 0.090 0.168 a
3 0.152 0.094 0.754 c
4 0.092 0.832 0.076 b
5 0.748 0.100 0.152 a
6 0.680 0.098 0.222 a
>
> res2 <- predict(t$learner.model, data.frame(x2,x1,x3))
> res2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
a a c b a a a c a b b b b c c a b b a c b a c c b c
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
a a b a c c c b c b c a b c c b c b c a c c b b
Levels: a b c
>
> res1$data$response == res2
[1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[16] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[31] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[46] TRUE TRUE TRUE TRUE TRUE
Another Update: If I change predict.type to "response" from "prob", the two svm prediction vectors agree with each other. I'm going to look into the differences of these types, I had thought that "prob" gave the same results but also gave probabilities. Maybe this isn't the case?
答案 0 :(得分:1)
正如您所知,&#34;错误的来源&#34;是mlr
和kernlab
对预测类型有不同的默认值。
mlr
维持着相当多的内部&#34;状态&#34;并检查每个学习者的学习者参数以及如何处理培训和测试。您可以使用lrn$predict.type
获取学习者将进行的预测类型,在您的情况下,"prob"
会给出mlr
。如果您想了解所有血腥细节,请查看the implementation of classif.ksvm
。
不建议混合mlr
- 包裹的学习者和&#34; raw&#34;像你这样的学习者在这个例子中做,并且不应该这样做。如果你混合了它们,就会发现像你这样的事情 - 所以当使用mlr
时,只使用 mlr
构造来训练模型,做出预测,等
friend_with?
确实有测试以确保&#34; raw&#34;并且包装的学习者产生相同的输出,参见例如the one for classif.ksvm
答案 1 :(得分:0)
答案就在这里:
Why are probabilities and response in ksvm in R not consistent?
简而言之,ksvm type =“probabilities”给出的结果与type =“response”不同。
如果我跑
> res2 <- predict(t$learner.model, data.frame(x2,x1,x3), type = "probabilities")
> res2
然后我得到与上面res1相同的结果(type =“response”是默认值)。
不幸的是,似乎基于概率对图像进行分类并不像使用“响应”那样好。也许这仍然是估算分类确定性的最佳方法?