本文介绍了使用 H2O R 包中的 h2o.anomaly 函数重建 MSE 计算的处理方法,对大家解决问题具有一定的参考价值,需要的朋友们下面随着小编来一起学习吧!

问题描述

我试图执行自动编码器以进行异常检测.我使用 H2O R 包使用 h2o.anomaly 函数为样本数据生成重建 MSE.但是,我也尝试根据以下文档链接中的 MSE 公式自己手动计算它:http://docs.h2o.ai/h2o/latest-stable/h2o-docs/performance-and-prediction.html#mse-mean-squared-error

I was trying to perform Autoencoder for anomaly detection. I used H2O R package to generate reconstruction MSE for a sample data using h2o.anomaly function. However, I have also tried to manually calculate it by myself according the the MSE formula from the documentation link below:http://docs.h2o.ai/h2o/latest-stable/h2o-docs/performance-and-prediction.html#mse-mean-squared-error

我用来构建模型的由三个特征和5行组成的训练数据如下:

The training data consisting of three features and 5 rows that I used to build the model is as below:

head(train_dat)

  Feature1  Feature2 Feature3
1    68.18 0.1806535 3.871201
2    71.51 0.3987761 2.484907
3    67.77 0.4285304 3.332205
4    69.58 0.1823216 2.890372
5    70.98 0.4134333 1.791759

我用于预测的由三个特征和5行组成的测试数据如下:

The test data consisting of three features and 5 rows that I used for prediction is as below:

head(test_dat)

  Feature1  Feature2 Feature3
1 68.33000 0.4350239 2.708050
2 73.98000 0.5550339 3.044522
3 67.11000 0.7323679 2.639057
4 69.90395 0.9999787 4.499810
5 71.28867 0.4882539 3.091042

运行训练和预测后,重建的特征如下:

After running training and prediction, the reconstructed features are as below:

head(mod.out)

  reconstr_Feature1 reconstr_Feature2 reconstr_Feature3
1          69.66297         0.4239244          2.346250
2          69.88329         0.3963843          2.381598
3          69.46544         0.4610502          2.233164
4          68.96117         0.4229165          2.676295
5          69.63208         0.3895452          2.530025

当我使用 h2o.anomaly 函数进行 MSE 计算时,我收到如下 MSE 输出:

When I used the h2o.anomaly function for MSE calculation, I received MSE output as below:

head(mse.list)

  Reconstruction.MSE
1         0.05310159
2         0.57037600
3         0.54427385
4         2.08407248
5         0.14251951

然而,当我尝试通过应用下面的函数来计算 MSE 时,我得到了不同的 MSE 输出:

However, when I tried to calculate the MSE by applying the function below, I obtained different MSE output:

mod.anon.validate <- apply((test_dat - mod.out)^2, 1, mean)
mse.list.validate <- as.data.frame(mod.anon.validate)
head(mse.list.validate)

  mod.anon.validate
1         0.6359438
2         5.7492281
3         1.9288268
4         1.5156829
5         1.0229217

我想知道我在手动 MSE 计算中做错了什么?当它被称为重建MSE"时,它与一般的MSE有什么不同?完整的 R 脚本如下:

I was wondering what I have done wrong in my manual MSE calculation? When it is called "Reconstruction MSE", is it different from the general MSE? The full R script is as below:

### H2O Autoencoder test run ###

#Load test and training data.
test_dat <- read.table("sample.test.dat", header=TRUE)
train_dat <- read.table("sample.train.dat", header=TRUE)

#Start H2O
library(h2o)
localH2O <- h2o.init(port =54321)

#Training and deep learning

feature_names <- names(train_dat[1:3])

unmod.hex <- as.h2o(train_dat, destination_frame="train.hex") ; mod.hex=as.h2o(test_dat, destination_frame="test.hex")

unmod.dl <- h2o.deeplearning(x=feature_names,
        training_frame=unmod.hex,
        autoencoder = TRUE,
        reproducible = T,
        hidden = c(3,2,3), epochs = 50,
        activation = "Tanh")

#Output result

mod.out <- as.data.frame(h2o.predict(unmod.dl,mod.hex,type=response))

mod.anon <- h2o.anomaly(unmod.dl, mod.hex, per_feature=FALSE)
mse.list <- as.data.frame(mod.anon)

mod.anon.validate <- apply((test_dat - mod.out)^2, 1, mean)
mse.list.validate <- as.data.frame(mod.anon.validate)

感谢您的帮助.

推荐答案

计算不匹配,因为 MSE 是在规范化空间中计算的.如果您在 h2o.deeplearning() 中设置 standardize=FALSE 参数,它将匹配:

The calculations don't match because MSE is calculated in the normalised space. If you set standardize=FALSE param in h2o.deeplearning() it will match:

unmod.dl <- h2o.deeplearning(x=feature_names, standardize = FALSE,
                             training_frame=unmod.hex,
                             autoencoder = TRUE,
                             reproducible = T,
                             hidden = c(3,2,3), epochs = 50,
                             activation = "Tanh")

mod.out <- as.data.frame(h2o.predict(unmod.dl, mod.hex, type=response))

mod.anon <- h2o.anomaly(unmod.dl, mod.hex, per_feature=FALSE)
mse.list <- as.data.frame(mod.anon)
mse.list

> mse.list
  Reconstruction.MSE
1           1512.740
2           1777.491
3           1458.438
4           1587.593
5           1648.999

> mod.anon.validate <- apply((test_dat - mod.out)^2, 1, mean)
> mse.list.validate <- as.data.frame(mod.anon.validate)
> mse.list.validate
  mod.anon.validate
1          1512.740
2          1777.491
3          1458.438
4          1587.593
5          1648.999

这篇关于使用 H2O R 包中的 h2o.anomaly 函数重建 MSE 计算的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持!

07-09 19:12