A 'data.frame' processor/conditioner that prepares real-world data for predictive modeling in a statistically sound manner. 'vtreat' prepares variables so that data has fewer exceptional cases, making it easier to safely use models in production. Common problems 'vtreat' defends against: 'Inf', 'NA', too many categorical levels, rare categorical levels, and new categorical levels (levels seen during application, but not during training). 'vtreat::prepare' should be used as you would use 'model.matrix'.
vtreat is an R data.frame processor/conditioner that prepares real-world data for predictive modeling in a statistically sound manner.
Even with modern machine learning techniques (random forests, support vector machines, neural nets, gradient boosted trees, and so on) or standard statistical methods (regression, generalized regression, generalized additive models) there are common data issues that can cause modeling to fail. vtreat deals with a number of these in a principled and automated fashion.
In particular vtreat emphasizes a concept called "y-aware pre-processing" and implements:
scale
argument to vtreat::prepare()
, using some of the ideas discussed here). This allows correct/sensible application of principal component analysis pre-processing in a machine learning context.The idea is: even with a sophisticated machine learning algorithm there are many ways messy real world data can defeat the modeling process, and vtreat helps with at least ten of them. We emphasize: these problems are already in your data, you simply build better and more reliable models if you attempt to mitigate them. Automated processing is no substitute for actually looking at the data, but vtreat supplies efficient, reliable, documented, and tested implementations of many of the commonly needed transforms.
To help explain the methods we have prepared some documentation:
Install either from CRAN:
install.packages('vtreat')
Or from GitHub:
devtools::install_github('WinVector/vtreat', build_vignettes=TRUE)
And then:
library('vtreat')help('vtreat')
Data treatments are "y-aware" (use distribution relations between independent variables and the dependent variable). For binary classification use 'designTreatmentsC()' and for numeric regression use 'designTreatmentsN()'.
After the design step, 'prepare()' should be used as you would use model.matrix. 'prepare()' treated variables are all numeric and never take the value NA or +-Inf (so are very safe to use in modeling).
In application we suggest splitting your data into three sets: one for building vtreat encodings, one for training models using these encodings, and one for test and model evaluation.
'vtreat' is supplied by Win-Vector LLC under a GPL-3 license, without warranty.
The purpose of 'vtreat' library is to reliably prepare data for supervised machine learning. We try to leave as much as possible to the machine learning algorithms themselves, but cover most of the truly necessary typically ignored precautions. The library is designed to produce a 'data.frame' that is entirely numeric and takes common precautions to guard against the following real world data issues:
Categorical variables with very many levels.
We re-encode such variables as a family of indicator or dummy variables for common levels plus an additional impact code (also called "effects coded" in Jacob Cohen, Patricia Cohen, Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences, 2nd edition, 1983). This allows principled use (including smoothing) of huge categorical variables (like zip-codes) when building models. This is critical for some libraries (such as 'randomForest', which has hard limits on the number of allowed levels).
Rare categorical levels.
Levels that do not occur often during training tend not to have reliable effect estimates and contribute to over-fit. vtreat helps with 2 precautions in this case. First the 'rareLevel' argument suppresses levels with this count our below from modeling, except possibly through a grouped contribution. Also with enough data vtreat attempts to estimate out of sample performance of derived variables. Finally we suggest users reserve a portion of data for vtreat design, separate from any data used in additional training, calibration, or testing.
Novel categorical levels.
A common problem in deploying a classifier to production is: new levels (levels not seen during training) encountered during model application. We deal with this by encoding categorical variables in a possibly redundant manner: reserving a dummy variable for all levels (not the more common all but a reference level scheme). This is in fact the correct representation for regularized modeling techniques and lets us code novel levels as all dummies simultaneously zero (which is a reasonable thing to try). This encoding while limited is cheaper than the fully Bayesian solution of computing a weighted sum over previously seen levels during model application.
Missing/invalid values NA, NaN, +-Inf.
Variables with these issues are re-coded as two columns. The first column is clean copy of the variable (with missing/invalid values replaced with either zero or the grand mean, depending on the user chose of the 'scale' parameter). The second column is a dummy or indicator that marks if the replacement has been performed. This is simpler than imputation of missing values, and allows the downstream model to attempt to use missingness as a useful signal (which it often is in industrial data).
Extreme values.
Variables can be restricted to stay in ranges seen during training. This can defend against some run-away classifier issues during model application.
Constant and near-constant variables.
Variables that "don't vary" or "nearly don't vary" are suppressed.
Need for estimated single-variable model effect sizes and significances.
It is a dirty secret that even popular machine learning techniques need some variable pruning (when exposed to very wide data frames, see here and here). We make the necessary effect size estimates and significances easily available and supply initial variable pruning.
The above are all awful things that often lurk in real world data. Automating these steps ensures they are easy enough that you actually perform them and leaves the analyst time to look for additional data issues. For example this allowed us to essentially automate a number of the steps taught in chapters 4 and 6 of Practical Data Science with R (Zumel, Mount; Manning 2014) into a very short worksheet (though we think for understanding it is essential to work all the steps by hand as we did in the book). The idea is: 'data.frame's prepared with the 'vtreat' library are somewhat safe to train on as some precaution has been taken against all of the above issues. Also of interest are the 'vtreat' variable significances (help in initial variable pruning, a necessity when there are a large number of columns) and 'vtreat::prepare(scale=TRUE)' which re-encodes all variables into effect units making them suitable for y-aware dimension reduction (variable clustering, or principal component analysis) and for geometry sensitive machine learning techniques (k-means, knn, linear SVM, and more). You may want to do more than the 'vtreat' library does (such as Bayesian imputation, variable clustering, and more) but you certainly do not want to do less.
There have been a number of recent substantial improvements to the library, including:
Some of our related articles (which should make clear some of our motivations, and design decisions):
Examples of current best practice using 'vtreat' (variable coding, train, test split) can be found here and here.
Trivial example:
library("vtreat") # categorical exampledTrainC <- data.frame(x=c('a','a','a','b','b',NA,NA), z=c(1,2,3,4,NA,6,NA),y=c(FALSE,FALSE,TRUE,FALSE,TRUE,TRUE,TRUE))dTestC <- data.frame(x=c('a','b','c',NA),z=c(10,20,30,NA)) # help("designTreatmentsC") treatmentsC <- designTreatmentsC(dTrainC,colnames(dTrainC),'y',TRUE, verbose=FALSE)print(treatmentsC$scoreFrame)#> varName varMoves sig needsSplit extraModelDegrees origName#> 1 x_lev_NA TRUE 0.09248399 FALSE 0 x#> 2 x_lev_x.a TRUE 0.26490379 FALSE 0 x#> 3 x_lev_x.b TRUE 0.80967242 FALSE 0 x#> 4 x_catP TRUE 0.26490379 TRUE 2 x#> 5 x_catB TRUE 0.66481792 TRUE 2 x#> 6 z_clean TRUE 0.13176020 FALSE 0 z#> 7 z_isBAD TRUE 0.09248399 FALSE 0 z#> code#> 1 lev#> 2 lev#> 3 lev#> 4 catP#> 5 catB#> 6 clean#> 7 isBAD # help("prepare") dTrainCTreated <- prepare(treatmentsC,dTrainC,pruneSig=1.0,scale=TRUE)varsC <- setdiff(colnames(dTrainCTreated),'y')# all input variables should be mean 0sapply(dTrainCTreated[,varsC,drop=FALSE],mean)#> x_lev_NA x_lev_x.a x_lev_x.b x_catP x_catB #> 3.965082e-18 -1.982154e-17 9.917546e-19 1.585994e-16 0.000000e+00 #> z_clean z_isBAD #> 7.927952e-18 -7.926292e-18# all non NA slopes should be 1sapply(varsC,function(c) { lm(paste('y',c,sep='~'), data=dTrainCTreated)$coefficients[[2]]})#> x_lev_NA x_lev_x.a x_lev_x.b x_catP x_catB z_clean z_isBAD #> 1 1 1 1 1 1 1dTestCTreated <- prepare(treatmentsC,dTestC,pruneSig=c(),scale=TRUE)print(dTestCTreated)#> x_lev_NA x_lev_x.a x_lev_x.b x_catP x_catB z_clean#> 1 -0.1714286 -0.2380952 0.02857143 -0.2380952 -0.1897682 1.194595#> 2 -0.1714286 0.1785714 -0.07142857 0.1785714 -0.1489924 2.951351#> 3 -0.1714286 0.1785714 0.02857143 1.0119048 -0.1320682 4.708108#> 4 0.4285714 0.1785714 0.02857143 0.1785714 0.4336447 0.000000#> z_isBAD#> 1 -0.1714286#> 2 -0.1714286#> 3 -0.1714286#> 4 0.4285714
# numeric exampledTrainN <- data.frame(x=c('a','a','a','a','b','b',NA,NA), z=c(1,2,3,4,5,NA,7,NA),y=c(0,0,0,1,0,1,1,1))dTestN <- data.frame(x=c('a','b','c',NA),z=c(10,20,30,NA))# help("designTreatmentsN")treatmentsN = designTreatmentsN(dTrainN,colnames(dTrainN),'y', verbose=FALSE)print(treatmentsN$scoreFrame)#> varName varMoves sig needsSplit extraModelDegrees origName code#> 1 x_lev_NA TRUE 0.1339746 FALSE 0 x lev#> 2 x_lev_x.a TRUE 0.2070312 FALSE 0 x lev#> 3 x_lev_x.b TRUE 1.0000000 FALSE 0 x lev#> 4 x_catP TRUE 0.2070312 TRUE 2 x catP#> 5 x_catN TRUE 0.6831941 TRUE 2 x catN#> 6 x_catD TRUE 0.6528290 TRUE 2 x catD#> 7 z_clean TRUE 0.1701892 FALSE 0 z clean#> 8 z_isBAD TRUE 0.1339746 FALSE 0 z isBADdTrainNTreated <- prepare(treatmentsN,dTrainN,pruneSig=1.0,scale=TRUE)varsN <- setdiff(colnames(dTrainNTreated),'y')# all input variables should be mean 0sapply(dTrainNTreated[,varsN,drop=FALSE],mean) #> x_lev_NA x_lev_x.a x_lev_x.b x_catP x_catN #> 6.938894e-18 0.000000e+00 7.703720e-34 2.775558e-17 0.000000e+00 #> x_catD z_clean z_isBAD #> -2.775558e-17 4.857226e-17 6.938894e-18# all non NA slopes should be 1sapply(varsN,function(c) { lm(paste('y',c,sep='~'), data=dTrainNTreated)$coefficients[[2]]}) #> x_lev_NA x_lev_x.a x_lev_x.b x_catP x_catN x_catD z_clean #> 1 1 1 1 1 1 1 #> z_isBAD #> 1dTestNTreated <- prepare(treatmentsN,dTestN,pruneSig=c(),scale=TRUE)print(dTestNTreated)#> x_lev_NA x_lev_x.a x_lev_x.b x_catP x_catN x_catD z_clean#> 1 -0.1666667 -0.25 -2.266233e-17 -0.25 -0.25 -0.06743804 0.9952381#> 2 -0.1666667 0.25 6.798700e-17 0.25 0.00 -0.25818161 2.5666667#> 3 -0.1666667 0.25 -2.266233e-17 0.75 0.00 -0.25818161 4.1380952#> 4 0.5000000 0.25 -2.266233e-17 0.25 0.50 0.39305768 0.0000000#> z_isBAD#> 1 -0.1666667#> 2 -0.1666667#> 3 -0.1666667#> 4 0.5000000 # for large data sets you can consider designing the treatments on # a subset like: d[sample(1:dim(d)[[1]],1000),]
Related work: