2017-06-08 59 views
1

這是一些可重複的代碼。我想知道當功能是單熱編碼時每個功能的SE計算是什麼。如果我想給我自己的嘗試:如何解釋一個熱編碼變量的Autoencoder anomoly SE?

它看起來像一些社會企業是1,我想這意味着重建是100%確定它是一回事,但它實際上是另一回事。對於分數誤差,它們是否代表了從softmax分類器分配給該類別的概率的不同程度的錯誤?

library(h2o) 
art <- data.frame(a = c("a","b","a","c","d","e","g","f","a"), 
       b = c("b","c","d","e","b","c","d","e","b"), 
       c = c(4,3,2,5,6,1,2,3,5)) 

dl = h2o.deeplearning(x = c("a","b","c"), training_frame = as.h2o(art), 
         autoencoder = TRUE, 
         reproducible = T, 
         seed = 1234, 
         hidden = c(1), epochs = 1) 
sus.anon = h2o.anomaly(dl, as.h2o(art), per_feature=TRUE) 

回答

0

我不知道h2o autoencoder,但在我看來,autoencoder不能正確使用單熱編碼變量。我嘗試了一切。我沒有嘗試的是'使用Gumbel-Softmax估計器的分類變分自動編碼器'(https://github.com/ericjang/gumbel-softmax)。