I've got a rather small dataset (162,000 observations with 13 attributes)
that I'm trying to use for modelling, using h2o.GBM. The response variable is categorical with large number of levels (~ 20,000 levels)
The model doesn't run out of memory or give any errors, but it's been going for nearly 24 hours without any progress (says 0% on H2o.GBM reporting)
I finally gave in and stopped it.
I'm wondering if there's anything wrong with my hyperparameters, as data is not particularly large.
here's my code:
library(h2o)
localH2O <- h2o.init(nthreads = -1, max_mem_size = "12g")
train.h20 <- as.h2o(analdata_train)
gbm1 <- h2o.gbm(
y = response_var
, x = independ_vars
, training_frame = train.h20
, ntrees = 3
, max_depth = 5
, min_rows = 10
, stopping_tolerance = 0.001
, learn_rate = 0.1
, distribution = "multinomial"
)
The way H2O GBM multinomial classification works is, when you ask for 1 tree as a parameter, it actually builds a tree for each level in the response column underneath the hood.
So 1 tree really means 20,000 trees in your case.
2 trees would really mean 40,000, and so on...
(Note the binomial classification case takes a shortcut and builds only one tree for both classes.)
So... it will probably finish but it could take quite a long time!
It's probably not a good idea to train a classifier with 20,000 classes -- most GBM implementations won't even let you do that. Can you group/cluster the classes into a smaller number of groups so that you can train a model with a smaller number of classes? If so, then you could perform your training in a two-stage process -- the first model would have K classes (assuming you clustered your classes into K groups). Then you can train secondary models that further classify the observations into your original classes.
This type of two-stage process may make sense if your classes represent groups that naturally clusters into a hierarchy of groups -- such as zip codes or ICD-10 medical diagnostic codes, for example.
If your use-case really demands that you train a 20,000 class GBM (and there's no way around it), then you should get a bigger cluster of machines to use in your H2O cluster (it's unclear how many CPUs you are using currently). H2O GBM should be able to finish training, assuming it has enough memory and CPUs, but it may take a while.
Related
I have a large dataset (3.5+ million observations) of a binary response variable that I am trying to compute a Hierarchical GAM with a global smoother with individual effects that have a Shared penalty (e.g. 'GS' in Pedersen et al. 2019). Specifically I am trying to estimate the following structure: Global > Geographic Zone (N=2) > Bioregion (N=20) > Season (N varies by bioregion). In total, I am trying to estimate 36 different nested parameters.
Here is the the code I am currently using:
modGS <- bam(
outbreak ~
te(days_diff,NDVI_mean,bs=c("tp","tp"),k=c(5,5)) +
t2(days_diff, NDVI_mean, Zone, Bioregion, Season, bs=c("tp", "tp","re","re","re"),k=c(5, 5), m=2, full=TRUE) +
s(Latitude,Longitude,k=50),
family=binomial(),select = TRUE,data=dat)
My main issue is that it is taking a long time (5+ days) to construct the model. This nesting structure cannot be discretized, so I cannot compute it in parallel. Further I have tried gamm4 but I ran into memory limit issues. Here is the gamm4 code:
modGS <- gamm4(
outbreak ~
t2(days_diff,NDVI_mean,bs=c("tp","tp"),k=c(5,5)) +
t2(days_diff, NDVI_mean, Zone, Bioregion, Season, bs=c("tp", "tp","re","re","re"),k=c(5, 5), m=2, full=TRUE) +
s(Latitude,Longitude,k=50),
family=binomial(),select = TRUE,data=dat)
What is the best/most computationally feasible way to run this model?
I cut down the computational time by reducing the amount of bioregion levels and randomly sampling ca. 60% of the data. This actually allow me to calculate OOB error for the model.
There is an article I read recently that has a specific section on decreasing computational time. The main things they highlight are:
Use the bam function with it's useful fREML estimation, which refactorizes the model matrix to make calculation faster. Here it seems you have already done that.
Adding the discrete = TRUE argument, which assumes only a smaller finite number of unique values for estimation.
Manipulating nthreads in this function so it runs more than one core in parallel in your computer.
As the authors caution, the second option can reduce the amount of accuracy in your estimates. I fit some large models recently doing this and found that it was not always the same as the default bam function, so its best to use this as a quick inspection rather than the full result you are looking for.
in traditional gbm, we can use
predict.gbm(model, newsdata=..., n.tree=...)
So that I can compare result with different number of trees for the test data.
In h2o.gbm, although it has n.tree to set, it seems it doesn't have any effect on the result. It's all the same as the default model:
h2o.test.pred <- as.vector(h2o.predict(h2o.gbm.model, newdata=test.frame, n.tree=100))
R2(h2o.test.pred, test.mat$y)
[1] -0.00714109
h2o.test.pred <- as.vector(h2o.predict(h2o.gbm.model, newdata=test.frame, n.tree=10))
> R2(h2o.test.pred, test.mat$y)
[1] -0.00714109
Does anybod have similar problem? How to solve it? h2o.gbm is much faster than gbm, so if it can get detailed result of each tree that would be great.
I don't think H2O supports what you are describing.
BUT, if what you are after is to get the performance against the number of trees used, that can be done at model building time.
library(h2o)
h2o.init()
iris <- as.h2o(iris)
parts <- h2o.splitFrame(iris,c(0.8,0.1))
train <- parts[[1]]
valid <- parts[[2]]
test <- parts[[3]]
m <- h2o.gbm(1:4, 5, train,
validation_frame = valid,
ntrees = 100, #Max desired
score_tree_interval = 1)
h2o.scoreHistory(m)
plot(m)
The score history will show the evaluation after adding each new tree. plot(m) will show a chart of this. Looks like 20 is plenty for iris!
BTW, if your real purpose was to find out the optimum number of trees to use, then switch early stopping on, and it will do that automatically for you. (Just make sure you are using both validation and test data frames.)
As of 3.20.0.6 H2O does support this. The method you are looking for is
staged_predict_proba. For classification models it produces predicted class probabilities after each iteration (tree), for every observation in your testing frame. For regression models (i.e. when response is numerical), although not really documented, it produces the actual prediction for every observation in your testing frame.
From these predictions it is also easy to compute various performance metrics (AUC, r2 etc), assuming that's what you're after.
Python API:
staged_predict_proba = model.staged_predict_proba(test)
R API:
staged_predict_proba <- h2o.staged_predict_proba(model, prostate.test)
I'm using R package randomForest to do a regression on some biological data. My training data size is 38772 X 201.
I just wondered---what would be a good value for the number of trees ntree and the number of variable per level mtry? Is there an approximate formula to find such parameter values?
Each row in my input data is a 200 character representing the amino acid sequence, and I want to build a regression model to use such sequence in order to predict the distances between the proteins.
The default for mtry is quite sensible so there is not really a need to muck with it. There is a function tuneRF for optimizing this parameter. However, be aware that it may cause bias.
There is no optimization for the number of bootstrap replicates. I often start with ntree=501 and then plot the random forest object. This will show you the error convergence based on the OOB error. You want enough trees to stabilize the error but not so many that you over correlate the ensemble, which leads to overfit.
Here is the caveat: variable interactions stabilize at a slower rate than error so, if you have a large number of independent variables you need more replicates. I would keep the ntree an odd number so ties can be broken.
For the dimensions of you problem I would start ntree=1501. I would also recommended looking onto one of the published variable selection approaches to reduce the number of your independent variables.
The short answer is no.
The randomForest function of course has default values for both ntree and mtry. The default for mtry is often (but not always) sensible, while generally people will want to increase ntree from it's default of 500 quite a bit.
The "correct" value for ntree generally isn't much of a concern, as it will be quite apparent with a little tinkering that the predictions from the model won't change much after a certain number of trees.
You can spend (read: waste) a lot of time tinkering with things like mtry (and sampsize and maxnodes and nodesize etc.), probably to some benefit, but in my experience not a lot. However, every data set will be different. Sometimes you may see a big difference, sometimes none at all.
The caret package has a very general function train that allows you to do a simple grid search over parameter values like mtry for a wide variety of models. My only caution would be that doing this with fairly large data sets is likely to get time consuming fairly quickly, so watch out for that.
Also, somehow I forgot that the ranfomForest package itself has a tuneRF function that is specifically for searching for the "optimal" value for mtry.
Could this paper help ?
Limiting the Number of Trees in Random Forests
Abstract. The aim of this paper is to propose a simple procedure that
a priori determines a minimum number of classifiers to combine in order
to obtain a prediction accuracy level similar to the one obtained with the
combination of larger ensembles. The procedure is based on the McNemar
non-parametric test of significance. Knowing a priori the minimum
size of the classifier ensemble giving the best prediction accuracy, constitutes
a gain for time and memory costs especially for huge data bases
and real-time applications. Here we applied this procedure to four multiple
classifier systems with C4.5 decision tree (Breiman’s Bagging, Ho’s
Random subspaces, their combination we labeled ‘Bagfs’, and Breiman’s
Random forests) and five large benchmark data bases. It is worth noticing
that the proposed procedure may easily be extended to other base
learning algorithms than a decision tree as well. The experimental results
showed that it is possible to limit significantly the number of trees. We
also showed that the minimum number of trees required for obtaining
the best prediction accuracy may vary from one classifier combination
method to another
They never use more than 200 trees.
One nice trick that I use is to initially start with first taking square root of the number of predictors and plug that value for "mtry". It is usually around the same value that tunerf funtion in random forest would pick.
I use the code below to check for accuracy as I play around with ntree and mtry (change the parameters):
results_df <- data.frame(matrix(ncol = 8))
colnames(results_df)[1]="No. of trees"
colnames(results_df)[2]="No. of variables"
colnames(results_df)[3]="Dev_AUC"
colnames(results_df)[4]="Dev_Hit_rate"
colnames(results_df)[5]="Dev_Coverage_rate"
colnames(results_df)[6]="Val_AUC"
colnames(results_df)[7]="Val_Hit_rate"
colnames(results_df)[8]="Val_Coverage_rate"
trees = c(50,100,150,250)
variables = c(8,10,15,20)
for(i in 1:length(trees))
{
ntree = trees[i]
for(j in 1:length(variables))
{
mtry = variables[j]
rf<-randomForest(x,y,ntree=ntree,mtry=mtry)
pred<-as.data.frame(predict(rf,type="class"))
class_rf<-cbind(dev$Target,pred)
colnames(class_rf)[1]<-"actual_values"
colnames(class_rf)[2]<-"predicted_values"
dev_hit_rate = nrow(subset(class_rf, actual_values ==1&predicted_values==1))/nrow(subset(class_rf, predicted_values ==1))
dev_coverage_rate = nrow(subset(class_rf, actual_values ==1&predicted_values==1))/nrow(subset(class_rf, actual_values ==1))
pred_prob<-as.data.frame(predict(rf,type="prob"))
prob_rf<-cbind(dev$Target,pred_prob)
colnames(prob_rf)[1]<-"target"
colnames(prob_rf)[2]<-"prob_0"
colnames(prob_rf)[3]<-"prob_1"
pred<-prediction(prob_rf$prob_1,prob_rf$target)
auc <- performance(pred,"auc")
dev_auc<-as.numeric(auc#y.values)
pred<-as.data.frame(predict(rf,val,type="class"))
class_rf<-cbind(val$Target,pred)
colnames(class_rf)[1]<-"actual_values"
colnames(class_rf)[2]<-"predicted_values"
val_hit_rate = nrow(subset(class_rf, actual_values ==1&predicted_values==1))/nrow(subset(class_rf, predicted_values ==1))
val_coverage_rate = nrow(subset(class_rf, actual_values ==1&predicted_values==1))/nrow(subset(class_rf, actual_values ==1))
pred_prob<-as.data.frame(predict(rf,val,type="prob"))
prob_rf<-cbind(val$Target,pred_prob)
colnames(prob_rf)[1]<-"target"
colnames(prob_rf)[2]<-"prob_0"
colnames(prob_rf)[3]<-"prob_1"
pred<-prediction(prob_rf$prob_1,prob_rf$target)
auc <- performance(pred,"auc")
val_auc<-as.numeric(auc#y.values)
results_df = rbind(results_df,c(ntree,mtry,dev_auc,dev_hit_rate,dev_coverage_rate,val_auc,val_hit_rate,val_coverage_rate))
}
}
I've been training randomForest models in R on 7 million rows of data (41 features). Here's an example call:
myModel <- randomForest(RESPONSE~., data=mydata, ntree=50, maxnodes=30)
I thought surely with only 50 trees and 30 terminal nodes that the memory footprint of "myModel" would be small. But it's 65 megs in a dump file. The object seems to be holding all sorts of predicted, actual, and vote data from the training process.
What if I just want the forest and that's it? I want a tiny dump file that I can load later to make predictions off of quickly. I feel like the forest by itself shouldn't be all that large...
Anyone know how to strip this sucker down to just something I can make predictions off of going forward?
Trying to get out of the habit of posting answers as comments...
?randomForest advises against using the formula interface with large numbers of variables... are the results any different if you don't use the formula interface? The Value section of ?randomForest also tells you how to turn off some of the output (importance matrix, the entire forest, proximity matrix, etc.).
For example:
myModel <- randomForest(mydata[,!grepl("RESPONSE",names(mydata))],
mydata$RESPONSE, ntree=50, maxnodes=30, importance=FALSE,
localImp=FALSE, keep.forest=FALSE, proximity=FALSE, keep.inbag=FALSE)
You can make use of tuneRF function in R to know the number of trees and make the size smaller.
tuneRF(data_train, data_train$Response, stepFactor = 1.2, improve = 0.01, plot = T, trace = T)
use ?tuneRF to know more about inside variables.
I am using random forests in a big data problem, which has a very unbalanced response class, so I read the documentation and I found the following parameters:
strata
sampsize
The documentation for these parameters is sparse (or I didn´t have the luck to find it) and I really don´t understand how to implement it. I am using the following code:
randomForest(x=predictors,
y=response,
data=train.data,
mtry=lista.params[1],
ntree=lista.params[2],
na.action=na.omit,
nodesize=lista.params[3],
maxnodes=lista.params[4],
sampsize=c(250000,2000),
do.trace=100,
importance=TRUE)
The response is a class with two possible values, the first one appears more frequently than the second (10000:1 or more)
The list.params is a list with different parameters (duh! I know...)
Well, the question (again) is: How I can use the 'strata' parameter? I am using sampsize correctly?
And finally, sometimes I get the following error:
Error in randomForest.default(x = predictors, y = response, data = train.data, :
Still have fewer than two classes in the in-bag sample after 10 attempts.
Sorry If I am doing so many (and maybe stupid) questions ...
You should try using sampling methods that reduce the degree of imbalance from 1:10,000 down to 1:100 or 1:10. You should also reduce the size of the trees that are generated. (At the moment these are recommendations that I am repeating only from memory, but I will see if I can track down more authority than my spongy cortex.)
One way of reducing the size of trees is to set the "nodesize" larger. With that degree of imbalance you might need to have the node size really large, say 5-10,000. Here's a thread in rhelp:
https://stat.ethz.ch/pipermail/r-help/2011-September/289288.html
In the current state of the question you have sampsize=c(250000,2000), whereas I would have thought that something like sampsize=c(8000,2000), was more in line with my suggestions. I think you are creating samples where you do not have any of the group that was sampled with only 2000.
There are a few options.
If you have a lot of data, set aside a random sample of the data. Build your model on one set, then use the other to determine a proper cutoff for the class probabilities using an ROC curve.
You can also upsample the data in the minority class. The SMOTE algorithm might help (see the reference below and the DMwR package for a function).
You can also use other techniques. rpart() and a few other functions can allow different costs on the errors, so you could favor the minority class more. You can bag this type of rpart() model to approximate what random forest is doing.
ksvm() in the kernlab package can also use unbalanced costs (but the probability estimates are no longer good when you do this). Many other packages have arguments for setting the priors. You can also adjust this to put more emphasis on the minority class.
One last thought: maximizing models based on accuracy isn't going to get you anywhere (you can get 99.99% off the bat). The caret can tune models based on the Kappa statistic, which is a much better choice in your case.
Sorry, I don't know how to post a comment on the earlier answer, so I'll create a separate answer.
I suppose that the problem is caused by high imbalance of dataset (too few cases of one of the classes are present). For each tree in RF the algorithm creates bootstrap sample, which is a training set for this tree. And if you have too few examples of one of the classes in your dataset, then the bootstrap sampling will select examples of only one class (major class). And thus tree cannot be grown on only one class examples. It seems that there is a limit on 10 unsuccessful sampling attempts.
So the proposition of DWin to reduce the degree of imbalance to lower values (1:100 or 1:10) is the most reasonable one.
Pretty sure I disagree with the idea of removing observations from your sample.
Instead you might consider using a stratified sample to set a fixed percentage of each class each time it is resampled. This can be done with the Caret package. This way you will not be omitting observations by reducing the size of your training sample. It will not allow you to over represent your classes but will make sure that each subsample has a representative sample.
Here is an example I found:
len_pos <- nrow(example_dataset[example_dataset$target==1,])
len_neg <- nrow(example_dataset[example_dataset$target==0,])
train_model <- function(training_data, labels, model_type, ...) {
experiment_control <- trainControl(method="repeatedcv",
number = 10,
repeats = 2,
classProbs = T,
summaryFunction = custom_summary_function)
train(x = training_data,
y = labels,
method = model_type,
metric = "custom_score",
trControl = experiment_control,
verbose = F,
...)
}
# strata refers to which feature to do stratified sampling on.
# sampsize refers to the size of the bootstrap samples to be taken from each class. These samples will be taken as input
# for each tree.
fit_results <- train_model(example_dataset
, as.factor(sprintf("c%d", as.numeric(example_dataset$target)))
,"rf"
,tuneGrid = expand.grid(mtry = c( 3,5,10))
,ntree=500
,strata=as.factor(example_dataset$target)
,sampsize = c('1'=as.integer(len_pos*0.25),'0'=as.integer(len_neg*0.8))
)