我已经运行了具有类别变量和数值变量的逻辑回归模型。该模型试图根据第一周来预测一个月内的网站访问量。显然,第一周的网站访问量是最强的指标。但是,当我使用各种模型和激活功能进行深度学习时,模型的性能非常差。基于var_imp函数,它对非常不重要的变量具有重要性(基于我的逻辑回归模型,这非常好,这是错误的),并且似乎只对分类子集进行了高重要性排序。而且即使在训练数据上,该模型也无法正常运行,这是一个真实的警告信号!所以我只想上传我的代码以检查我是否在做任何不损害模型的事情。后勤回归使它正确无误,但深度学习却使它如此错误,这似乎很奇怪,所以我想我已经做了一些事情!
summary(data)
$ VAR1: Factor w/ 8 levels ,..: 1 5 2 1 7 2 5 1 5 1 ...
$ VAR2: Factor w/ 5 levels ,..: 1 4 1 1 4 4 4 1 1 4 ...
$ VAR3: Factor w/ 2 levels "F","M": 2 2 2 1 2 2 2 2 2 2 ...
$ VAR4: Factor w/ 2 levels : 2 1 2 2 1 1 1 2 2 1 ...
$ VAR5 : num 1000 20 30 20 30 30 30 50 30 400 ...
$ VAR6: Factor w/ 2 levels "N","Y": 1 2 2 1 2 2 2 2 1 2 ...
$ VAR7: Factor w/ 2 levels "N","Y": 1 2 2 1 2 2 2 2 1 2 ...
$ VAR8: num 0 0 0 0 0 0 0 0 0 0 ...
$ VAR9: num 56 52 49 29 28 38 34 79 53 36 ...
$ VAR10: num 3 2 1 3 2 2 3 4 2 2 ...
$ VAR11: num 1 1 1 2 2 1 1 1 1 2 ...
$ VAR12: Factor w/ 2 levels "N","Y": 1 1 1 1 2 1 1 1 1 1 ...
$ VAR13: num 1 0 1 1 1 0 1 0 0 0 ...
$ VAR14: Factor w/ 2 levels "N","Y": 2 1 1 1 1 1 1 1 1 1 ...
$ VAR15: Factor w/ 2 levels "N","Y": 1 1 1 1 1 1 1 1 1 1 ...
$ VAR16: num 1 0 0 1 0 0 0 1 1 0 ...
$ VAR17: num 19 7 1 4 10 2 4 4 7 12 ...
$ VAR18: Factor w/ 2 levels "N","Y": 1 2 2 2 2 2 2 1 2 1 ...
$ VAR19: Factor w/ 2 levels "0","Y": 1 1 2 1 1 1 1 1 1 1 ...
$ VAR20: Factor w/ 2 levels "N","Y": 1 1 2 1 1 1 1 1 1 1 ...
$ VAR21: Factor w/ 2 levels "N","Y": 1 1 1 1 1 1 1 1 1 1 ...
$ VAR22: : num 0.579 0 0 0 0.4 ...
$ VAR23: num 1.89 1 1 1 2.9 ...
$ VAR24: num 0.02962 0.00691 0.05327 0.02727 0.01043 ...
$ VAR25: Factor w/ 3 levels ..: 2 2 2 3 3 2 3 2 1 3 ...
$ VAR26: num 3 2 1 2 3 1 2 1 2 4 ...
$ VAR27: num 3 2 1 1 5 1 1 1 1 2 ...
$ VAR_RESPONSE: num 7 24 4 3 8 12 5 48 2 7 ...
sapply(data,function(x) sum(is.na(x)))
colSums(is.na(data))
data[is.na(data)] = 0
d.hex = as.h2o(data, destination_frame= "d.hex")
Data_g.split = h2o.splitFrame(data = d.hex,ratios = 0.75)
Data_train = Data_g.split[[1]]#75% training data
Data_test = Data_g.split[[2]]
activation_opt <-
c("Rectifier","RectifierWithDropout","Maxout","MaxoutWithDropout",
"Tanh","TanhWithDropout")
hidden_opt <- list(c(10,10),c(20,15),c(50,50,50),c(5,3,2),c(100,100),c(5),c(30,30,30),c(50,50,50,50),c(5,4,3,2))
l1_opt <- c(0,1e-3,1e-5,1e-7,1e-9)
l2_opt <- c(0,1e-3,1e-5,1e-7,1e-9)
hyper_params <- list( activation=activation_opt,
hidden=hidden_opt,
l1=l1_opt,
l2=l2_opt )
search_criteria <- list(strategy = "RandomDiscrete", max_models=30)
dl_grid10 <- h2o.grid("deeplearning"
,grid_id = "deep_learn10"
,hyper_params = hyper_params
,search_criteria = search_criteria
,x = 1:27
,y = "VAR_RESPONSE"
,training_frame = Data_train)
d_grid10 <- h2o.getGrid("deep_learn10",sort_by = "mse")
mn = h2o.deeplearning(x = 1:27,
y = "VAR_RESPONSE",
training_frame = Data_train,
model_id = "mn",
activation = "Maxout",
l1 = 0,
l2 = 1e-9,
hidden = c(100,100),)