Fetching contributors… Cannot retrieve contributors at this time
652 lines (536 sloc) 27 KB

Using vtreat with Classification Problems

Nina Zumel and John Mount September 2019

Note this is a description of the R version of vtreat, the same example for the Python version of vtreat can be found here.

Preliminaries

library(rqdatatable)
library(vtreat)
suppressPackageStartupMessages(library(ggplot2))
library(WVPlots)

Generate example data.

• y is a noisy sinusoidal function of the variable x
• yc is the output to be predicted: whether y is > 0.5.
• Input xc is a categorical variable that represents a discretization of y, along some NAs
• Input x2 is a pure noise variable with no relationship to the output
make_data <- function(nrows) {
d <- data.frame(x = 5*rnorm(nrows))
d['y'] = sin(d['x']) + 0.1*rnorm(n = nrows)
d[4:10, 'x'] = NA                  # introduce NAs
d['xc'] = paste0('level_', 5*round(d\$y/5, 1))
d['x2'] = rnorm(n = nrows)
d[d['xc']=='level_-1', 'xc'] = NA  # introduce a NA level
d['yc'] = d[['y']]>0.5
return(d)
}

d = make_data(500)

d %.>%
knitr::kable(.)
x y xc x2 yc
-2.620747 -0.4162074 level_-0.5 -0.6604022 FALSE
10.714563 -0.7431123 level_-0.5 -0.2577693 FALSE
4.138288 -0.7402798 level_-0.5 0.5093482 FALSE
NA 0.9386129 level_1 2.5480866 TRUE
NA 1.0942563 level_1 1.1413150 TRUE
NA 0.1037898 level_0 0.8132199 FALSE

Some quick data exploration

Check how many levels xc has, and their distribution (including NA)

unique(d['xc'])
##           xc
## 1 level_-0.5
## 4    level_1
## 6    level_0
## 7  level_0.5
## 8       <NA>
table(d\$xc, useNA = 'always')
##
## level_-0.5    level_0  level_0.5    level_1       <NA>
##         89         86         96        102        127

Find the mean value of yc

mean(d[['yc']])
##  0.326

Plot of yc versus x.

ggplot(d, aes(x=x, y=as.numeric(yc))) +
geom_line()
## Warning: Removed 7 rows containing missing values (geom_path). Build a transform appropriate for classification problems.

Now that we have the data, we want to treat it prior to modeling: we want training data where all the input variables are numeric and have no missing values or NAs.

First create the data treatment transform design object, in this case a treatment for a binomial classification problem.

We use the training data d to fit the transform and the return a treated training set: completely numeric, with no missing values.

transform_design = vtreat::mkCrossFrameCExperiment(
dframe = d,                                    # data to learn transform from
varlist = setdiff(colnames(d), c('y', 'yc')),  # columns to transform
outcomename = 'yc',                            # outcome variable
outcometarget = TRUE                           # outcome of interest
)
##  "vtreat 1.4.7 start initial treatment design Tue Oct  1 10:35:47 2019"
##  " start cross frame work Tue Oct  1 10:35:47 2019"
##  " vtreat::mkCrossFrameCExperiment done Tue Oct  1 10:35:47 2019"
transform <- transform_design\$treatments
d_prepared <- transform_design\$crossFrame
score_frame <- transform\$scoreFrame
score_frame\$recommended <- score_frame\$varMoves & (score_frame\$sig < 1/nrow(score_frame))

Note that for the training data d: transform_design\$crossFrame is not the same as transform.prepare(d); the second call can lead to nested model bias in some situations, and is not recommended. For other, later data, not seen during transform design transform.preprare(o) is an appropriate step.

Now examine the score frame, which gives information about each new variable, including its type, which original variable it is derived from, its (cross-validated) correlation with the outcome, and its (cross-validated) significance as a one-variable linear model for the outcome.

knitr::kable(score_frame)
varName varMoves rsq sig needsSplit extraModelDegrees origName code recommended
x TRUE 0.0000588 0.8472121 FALSE 0 x clean FALSE
x_isBAD TRUE 0.0070953 0.0343077 FALSE 0 x isBAD TRUE
xc_catP TRUE 0.0075139 0.0294077 TRUE 4 xc catP TRUE
xc_catB TRUE 0.7988184 0.0000000 TRUE 4 xc catB TRUE
x2 TRUE 0.0031878 0.1560070 FALSE 0 x2 clean FALSE
xc_lev_NA TRUE 0.1903343 0.0000000 FALSE 0 xc lev TRUE
xc_lev_x_level_minus_0_5 TRUE 0.1255315 0.0000000 FALSE 0 xc lev TRUE
xc_lev_x_level_0 TRUE 0.1207530 0.0000000 FALSE 0 xc lev TRUE
xc_lev_x_level_0_5 TRUE 0.0773074 0.0000000 FALSE 0 xc lev TRUE
xc_lev_x_level_1 TRUE 0.4599270 0.0000000 FALSE 0 xc lev TRUE

Note that the variable xc has been converted to multiple variables:

• an indicator variable for each possible level (xc_lev_*)
• the value of a (cross-validated) one-variable model for yc as a function of xc (xc_catB)
• a variable that returns how prevalent this particular value of xc is in the training data (xc_catP)
• a variable indicating when xc was NA in the original data (xc_lev_NA for categorical variables, x_isBAD for continuous variables).

Any or all of these new variables are available for downstream modeling.

The recommended column indicates which variables are non constant (varMoves == TRUE) and have a significance value smaller than 1/nrow(score_frame). See the section Deriving the Default Thresholds below for the reasoning behind such a default threshold. Recommended columns are intended as advice about which variables appear to be most likely to be useful in a downstream model. This advice attempts to be conservative, to reduce the possibility of mistakenly eliminating variables that may in fact be useful (although, obviously, it can still mistakenly eliminate variables that have a real but non-linear relationship to the output, as is the case with x, in our example).

Let’s look at the variables that are and are not recommended:

# recommended variables
score_frame[score_frame[['recommended']], 'varName', drop = FALSE]  %.>%
knitr::kable(.)
varName
3 xc_catP
4 xc_catB
6 xc_lev_NA
7 xc_lev_x_level_minus_0_5
8 xc_lev_x_level_0
9 xc_lev_x_level_0_5
10 xc_lev_x_level_1
# not recommended variables
score_frame[!score_frame[['recommended']], 'varName', drop = FALSE] %.>%
knitr::kable(.)
varName
1 x
5 x2

Notice that d_prepared only includes derived variables and the outcome y:

d_prepared %.>%
knitr::kable(.)
x x_isBAD xc_catP xc_catB x2 xc_lev_NA xc_lev_x_level_minus_0_5 xc_lev_x_level_0 xc_lev_x_level_0_5 xc_lev_x_level_1 yc
-2.6207469 0 0.1856287 -12.61272 -0.6604022 0 1 0 0 0 FALSE
10.7145633 0 0.1831832 -12.60092 -0.2577693 0 1 0 0 0 FALSE
4.1382878 0 0.1831832 -12.60092 0.5093482 0 1 0 0 0 FALSE
-0.5287519 1 0.2132132 14.20699 2.5480866 0 0 0 0 1 TRUE
-0.5759212 1 0.2042042 14.15015 1.1413150 0 0 0 0 1 TRUE
-0.3783658 1 0.1766467 -12.56313 0.8132199 0 0 1 0 0 FALSE

A Closer Look at catB variables

Variables of type catB are the outputs of a one-variable regularized logistic regression of a categorical variable (in our example, xc) against the centered output on the (cross-validated) treated training data.

Let’s see whether xc_catB makes a good one-variable model for yc. It has a large AUC:

WVPlots::ROCPlot(
frame = d_prepared,
xvar = 'xc_catB',
truthVar = 'yc',
truthTarget = TRUE,
title = 'performance of xc_catB variable') This indicates that xc_catB is strongly predictive of the outcome. Negative values of xc_catB correspond strongly to negative outcomes, and positive values correspond strongly to positive outcomes.

WVPlots::DoubleDensityPlot(
frame = d_prepared,
xvar = 'xc_catB',
truthVar = 'yc',
title = 'performance of xc_catB variable') The values of xc_catB are in “link space”.

Variables of type catB are useful when dealing with categorical variables with a very large number of possible levels. For example, a categorical variable with 10,000 possible values potentially converts to 10,000 indicator variables, which may be unwieldy for some modeling methods. Using a single numerical variable of type catB may be a preferable alternative.

Using the Prepared Data in a Model

Of course, what we really want to do with the prepared training data is to fit a model jointly with all the (recommended) variables. Let’s try fitting a logistic regression model to d_prepared.

model_vars <- score_frame\$varName[score_frame\$recommended]
f <- wrapr::mk_formula('yc', model_vars, outcome_target = TRUE)

model = glm(f, data = d_prepared)

# now predict
d_prepared['prediction'] = predict(
model,
newdata = d_prepared,
type = 'response')
## Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type
## == : prediction from a rank-deficient fit may be misleading
# look at the ROC curve (on the training data)
WVPlots::ROCPlot(
frame = d_prepared,
xvar = 'prediction',
truthVar = 'yc',
truthTarget = TRUE,
title = 'Performance of logistic regression model on training data') Now apply the model to new data.

# create the new data
dtest <- make_data(450)

# prepare the new data with vtreat
dtest_prepared = prepare(transform, dtest)

# apply the model to the prepared data
dtest_prepared['prediction'] = predict(
model,
newdata = dtest_prepared,
type = 'response')
## Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type
## == : prediction from a rank-deficient fit may be misleading
WVPlots::ROCPlot(
frame = dtest_prepared,
xvar = 'prediction',
truthVar = 'yc',
truthTarget = TRUE,
title = 'Performance of logistic regression model on test data') Parameters for BinomialOutcomeTreatment

We’ve tried to set the defaults for all parameters so that vtreat is usable out of the box for most applications.

suppressPackageStartupMessages(library(printr))
help("mkCrossFrameCExperiment")
## Run categorical cross-frame experiment.
##
## Description:
##
##      Builds a 'designTreatmentsC' treatment plan and a data frame
##      prepared from 'dframe' that is "cross" in the sense each row is
##      treated using a treatment plan built from a subset of dframe
##      disjoint from the given row. The goal is to try to and supply a
##      method of breaking nested model bias other than splitting into
##      calibration, training, test sets.
##
## Usage:
##
##      mkCrossFrameCExperiment(dframe, varlist, outcomename, outcometarget, ...,
##        weights = c(), minFraction = 0.02, smFactor = 0, rareCount = 0,
##        rareSig = 1, collarProb = 0, codeRestriction = NULL,
##        customCoders = NULL, scale = FALSE, doCollar = FALSE,
##        splitFunction = NULL, ncross = 3, forceSplit = FALSE,
##        catScaling = TRUE, verbose = TRUE, parallelCluster = NULL,
##        use_parallel = TRUE)
##
## Arguments:
##
##   dframe: Data frame to learn treatments from (training data), must
##           have at least 1 row.
##
##  varlist: Names of columns to treat (effective variables).
##
## outcomename: Name of column holding outcome variable.
##           dframe[[outcomename]] must be only finite non-missing values.
##
## outcometarget: Value/level of outcome to be considered "success", and
##           there must be a cut such that
##           dframe[[outcomename]]==outcometarget at least twice and
##           dframe[[outcomename]]!=outcometarget at least twice.
##
##      ...: no additional arguments, declared to forced named binding of
##           later arguments
##
##  weights: optional training weights for each row
##
## minFraction: optional minimum frequency a categorical level must have
##           to be converted to an indicator column.
##
## smFactor: optional smoothing factor for impact coding models.
##
## rareCount: optional integer, allow levels with this count or below to
##           be pooled into a shared rare-level.  Defaults to 0 or off.
##
##  rareSig: optional numeric, suppress levels from pooling at this
##           significance value greater.  Defaults to NULL or off.
##
## collarProb: what fraction of the data (pseudo-probability) to collar
##           data at if doCollar is set during 'prepare.treatmentplan'.
##
## codeRestriction: what types of variables to produce (character array of
##           level codes, NULL means no restriction).
##
## customCoders: map from code names to custom categorical variable
##           encoding functions (please see <URL:
##           https://github.com/WinVector/vtreat/blob/master/extras/CustomLevelCoders.md>).
##
##    scale: optional if TRUE replace numeric variables with regression
##           ("move to outcome-scale").
##
## doCollar: optional if TRUE collar numeric variables by cutting off
##           after a tail-probability specified by collarProb during
##           treatment design.
##
## splitFunction: (optional) see vtreat::buildEvalSets .
##
##   ncross: optional scalar>=2 number of cross-validation rounds to
##           design.
##
## forceSplit: logical, if TRUE force cross-validated significance
##           calculations on all variables.
##
## catScaling: optional, if TRUE use glm() linkspace, if FALSE use lm()
##           for scaling.
##
##  verbose: if TRUE print progress.
##
## parallelCluster: (optional) a cluster object created by package
##           parallel or package snow.
##
## use_parallel: logical, if TRUE use parallel methods.
##
## Value:
##
##      list with treatments and crossFrame
##
##
##      'designTreatmentsC', 'designTreatmentsN', 'prepare.treatmentplan'
##
## Examples:
##
##      set.seed(23525)
##      zip <- paste('z',1:100)
##      N <- 200
##      d <- data.frame(zip=sample(zip,N,replace=TRUE),
##                      zip2=sample(zip,20,replace=TRUE),
##                      y=runif(N))
##      del <- runif(length(zip))
##      names(del) <- zip
##      d\$y <- d\$y + del[d\$zip2]
##      d\$yc <- d\$y>=mean(d\$y)
##      cC <- mkCrossFrameCExperiment(d,c('zip','zip2'),'yc',TRUE,
##        rareCount=2,rareSig=0.9)
##      cor(as.numeric(cC\$crossFrame\$yc),cC\$crossFrame\$zip_catB)  # poor
##      cor(as.numeric(cC\$crossFrame\$yc),cC\$crossFrame\$zip2_catB) # better
##      treatments <- cC\$treatments
##      dTrainV <- cC\$crossFrame

Some parameters of note include:

codeRestriction: The types of synthetic variables that vtreat will (potentially) produce. See Types of prepared variables below.

minFraction: For categorical variables, indicator variables (type indicator_code) are only produced for levels that are present at least minFraction of the time. A consequence of this is that 1/minFraction is the maximum number of indicators that will be produced for a given categorical variable. To make sure that all possible indicator variables are produced, set minFraction = 0

splitFunction: The cross validation method used by vtreat. Most people won’t have to change this.

ncross: The number of folds to use for cross-validation

customCoders: For passing in user-defined transforms for custom data preparation. Won’t be needed in most situations, but see here for an example of applying a GAM transform to input variables.

Types of prepared variables

clean: Produced from numerical variables: a clean numerical variable with no NAs or missing values

lev: Produced from categorical variables, one for each (common) level: for each level of the variable, indicates if that level was “on”

catP: Produced from categorical variables: indicates how often each level of the variable was “on”

catB: Produced from categorical variables: score from a one-dimensional model of the centered output as a function of the variable

is_BAD: Produced for both numerical and categorical variables: an indicator variable that marks when the original variable was missing or NaN.

More on the coding types can be found here.

Example: Produce only a subset of variable types

In this example, suppose you only want to use indicators and continuous variables in your model; in other words, you only want to use variables of types (clean_copy, missing_indicator, and indicator_code), and no catB or prevalence_code variables.

transform_design_thin = vtreat::mkCrossFrameCExperiment(
dframe = d,                                    # data to learn transform from
varlist = setdiff(colnames(d), c('y', 'yc')),  # columns to transform
outcomename = 'yc',                            # outcome variable
outcometarget = TRUE,                          # outcome of interest
codeRestriction = c('lev',                     # transforms we want
'clean',
)
##  "vtreat 1.4.7 start initial treatment design Tue Oct  1 10:35:50 2019"
##  " start cross frame work Tue Oct  1 10:35:50 2019"
##  " vtreat::mkCrossFrameCExperiment done Tue Oct  1 10:35:50 2019"
transform_thin <- transform_design_thin\$treatments
d_prepared_thin <- transform_design_thin\$crossFrame
score_frame_thin <- transform_thin\$scoreFrame

d_prepared_thin %.>%
knitr::kable(.)
x x_isBAD x2 xc_lev_NA xc_lev_x_level_minus_0_5 xc_lev_x_level_0 xc_lev_x_level_0_5 xc_lev_x_level_1 yc
-2.6207469 0 -0.6604022 0 1 0 0 0 FALSE
10.7145633 0 -0.2577693 0 1 0 0 0 FALSE
4.1382878 0 0.5093482 0 1 0 0 0 FALSE
-0.4528789 1 2.5480866 0 0 0 0 1 TRUE
-0.4528789 1 1.1413150 0 0 0 0 1 TRUE
-0.5602661 1 0.8132199 0 0 1 0 0 FALSE
knitr::kable(score_frame_thin)
varName varMoves rsq sig needsSplit extraModelDegrees origName code
x TRUE 0.0000588 0.8472121 FALSE 0 x clean
x_isBAD TRUE 0.0070953 0.0343077 FALSE 0 x isBAD
x2 TRUE 0.0031878 0.1560070 FALSE 0 x2 clean
xc_lev_NA TRUE 0.1903343 0.0000000 FALSE 0 xc lev
xc_lev_x_level_minus_0_5 TRUE 0.1255315 0.0000000 FALSE 0 xc lev
xc_lev_x_level_0 TRUE 0.1207530 0.0000000 FALSE 0 xc lev
xc_lev_x_level_0_5 TRUE 0.0773074 0.0000000 FALSE 0 xc lev
xc_lev_x_level_1 TRUE 0.4599270 0.0000000 FALSE 0 xc lev

Deriving the Default Thresholds

While machine learning algorithms are generally tolerant to a reasonable number of irrelevant or noise variables, too many irrelevant variables can lead to serious overfit; see this article for an extreme example, one we call “Bad Bayes”. The default threshold is an attempt to eliminate obviously irrelevant variables early.

Imagine that you have a pure noise dataset, where none of the n inputs are related to the output. If you treat each variable as a one-variable model for the output, and look at the significances of each model, these significance-values will be uniformly distributed in the range [0:1]. You want to pick a weakest possible significance threshold that eliminates as many noise variables as possible. A moment’s thought should convince you that a threshold of 1/n allows only one variable through, in expectation. This leads to the general-case heuristic that a significance threshold of 1/n on your variables should allow only one irrelevant variable through, in expectation (along with all the relevant variables). Hence, 1/n used to be our recommended threshold, when we developed the R version of vtreat.

As noted above, because vtreat estimates variable significances using linear methods by default, some variables with a non-linear relationship to the output may fail to pass the threshold. So it may make sense for the data scientist to filter (or not) as they will.

The variables can be appraised in a non-linear fashion as follows:

d %.>%
value_variables_C(dframe = .,
varlist = setdiff(colnames(.), c('y', 'yc')),
outcomename = 'yc',
outcometarget = TRUE) %.>%
knitr::kable(.)
rsq count sig var
x 0.007095258 2 0.0686153 x
x2 0.007484178 3 0.0891902 x2
xc 0.800235926 2 0.0000000 xc

More on non-linear variable scoring can be found here.

Conclusion

In all cases (classification, regression, unsupervised, and multinomial classification) the intent is that vtreat transforms are essentially one liners.

The preparation commands are organized as follows:

These current revisions of the examples are designed to be small, yet complete. So as a set they have some overlap, but the user can rely mostly on a single example for a single task type.

You can’t perform that action at this time.