Awesome

This is the Python version of the vtreat data preparation system (also available as an R package).

vtreat is a DataFrame processor/conditioner that prepares real-world data for supervised machine learning or predictive modeling in a statistically sound manner.

Installing

Install vtreat with either of:

• pip install vtreat
• pip install https://github.com/WinVector/pyvtreat/raw/master/pkg/dist/vtreat-0.4.6.tar.gz

Video Introduction

Our PyData LA 2019 talk on vtreat is a good video introduction to what problems vtreat can be used to solve. The slides can be found here.

Details

vtreat takes an input DataFrame that has a specified column called "the outcome variable" (or "y") that is the quantity to be predicted (and must not have missing values). Other input columns are possible explanatory variables (typically numeric or categorical/string-valued, these columns may have missing values) that the user later wants to use to predict "y". In practice such an input DataFrame may not be immediately suitable for machine learning procedures that often expect only numeric explanatory variables, and may not tolerate missing values.

To solve this, vtreat builds a transformed DataFrame where all explanatory variable columns have been transformed into a number of numeric explanatory variable columns, without missing values. The vtreat implementation produces derived numeric columns that capture most of the information relating the explanatory columns to the specified "y" or dependent/outcome column through a number of numeric transforms (indicator variables, impact codes, prevalence codes, and more). This transformed DataFrame is suitable for a wide range of supervised learning methods from linear regression, through gradient boosted machines.

The idea is: you can take a DataFrame of messy real world data and easily, faithfully, reliably, and repeatably prepare it for machine learning using documented methods using vtreat. Incorporating vtreat into your machine learning workflow lets you quickly work with very diverse structured data.

To get started with vtreat please check out our documentation:

• Getting started using vtreat for classification.
• Getting started using vtreat for regression.
• Getting started using vtreat for multi-category classification.
• Getting started using vtreat for unsupervised tasks.
• The vtreat Score Frame (a table mapping new derived variables to original columns).
• The original vtreat paper this note describes the methodology and theory. (The article describes the R version, however all of the examples can be found worked in Python here).

Some vtreat common capabilities are documented here:

• Score Frame score_frame_, using the score_frame_ information.
• Cross Validation Customized Cross Plans, controlling the cross validation plan.

vtreat is available as a Python/Pandas package, and also as an R package.

(logo: Julie Mount, source: “The Harvest” by Boris Kustodiev 1914)

vtreat is used by instantiating one of the classes vtreat.NumericOutcomeTreatment, vtreat.BinomialOutcomeTreatment, vtreat.MultinomialOutcomeTreatment, or vtreat.UnsupervisedTreatment. Each of these implements the sklearn.pipeline.Pipeline interfaces expecting a Pandas DataFrame as input. The vtreat steps are intended to be a "one step fix" that works well with sklearn.preprocessing stages.

The vtreat Pipeline.fit_transform() method implements the powerful cross-frame ideas (allowing the same data to be used for vtreat fitting and for later model construction, while mitigating nested model bias issues).

Background

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:

• Treatment of missing values through safe replacement plus an indicator column (a simple but very powerful method when combined with downstream machine learning algorithms).
• Treatment of novel levels (new values of categorical variable seen during test or application, but not seen during training) through sub-models (or impact/effects coding of pooled rare events).
• Explicit coding of categorical variable levels as new indicator variables (with optional suppression of non-significant indicators).
• Treatment of categorical variables with very large numbers of levels through sub-models (again impact/effects coding).
• Correct treatment of nested models or sub-models through data split / cross-frame methods (please see here) or through the generation of “cross validated” data frames (see here); these are issues similar to what is required to build statistically efficient stacked models or super-learners).

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:

• The vtreat package overall.
• Preparing data for analysis using R white-paper
• The types of new variables introduced by vtreat processing (including how to limit down to domain appropriate variable types).
• Statistically sound treatment of the nested modeling issue introduced by any sort of pre-processing (such as vtreat itself): nested over-fit issues and a general cross-frame solution.
• Principled ways to pick significance based pruning levels.

Example

This is an supervised classification example taken from the KDD 2009 cup. A copy of the data and details can be found here: https://github.com/WinVector/PDSwR2/tree/master/KDD2009. The problem was to predict account cancellation ("churn") from very messy data (column names not given, numeric and categorical variables, many missing values, some categorical variables with a large number of possible levels). In this example we show how to quickly use vtreat to prepare the data for modeling. vtreat takes in Pandas DataFrames and returns both a treatment plan and a clean Pandas DataFrame ready for modeling.

to install

!pip install vtreat !pip install wvpy Load our packages/modules.

import pandas
import xgboost
import vtreat
import vtreat.cross_plan
import numpy.random
import wvpy.util
import scipy.sparse

Read in explanitory variables.

# data from https://github.com/WinVector/PDSwR2/tree/master/KDD2009
dir = "../../../PracticalDataScienceWithR2nd/PDSwR2/KDD2009/"
d = pandas.read_csv(dir + 'orange_small_train.data.gz', sep='\t', header=0)
vars = [c for c in d.columns]
d.shape

(50000, 230)

Read in dependent variable we are trying to predict.

churn = pandas.read_csv(dir + 'orange_small_train_churn.labels.txt', header=None)
churn.columns = ["churn"]
churn.shape

(50000, 1)

churn["churn"].value_counts()

-1    46328
 1     3672
Name: churn, dtype: int64

Arrange test/train split.

numpy.random.seed(855885)
n = d.shape[0]
# https://github.com/WinVector/pyvtreat/blob/master/Examples/CustomizedCrossPlan/CustomizedCrossPlan.md
split1 = vtreat.cross_plan.KWayCrossPlanYStratified().split_plan(n_rows=n, k_folds=10, y=churn.iloc[:, 0])
train_idx = set(split1[0]['train'])
is_train = [i in train_idx for i in range(n)]
is_test = numpy.logical_not(is_train)

(The reported performance runs of this example were sensitive to the prevalance of the churn variable in the test set, we are cutting down on this source of evaluation variarance by using the stratified split.)

d_train = d.loc[is_train, :].copy()
churn_train = numpy.asarray(churn.loc[is_train, :]["churn"]==1)
d_test = d.loc[is_test, :].copy()
churn_test = numpy.asarray(churn.loc[is_test, :]["churn"]==1)

Take a look at the dependent variables. They are a mess, many missing values. Categorical variables that can not be directly used without some re-encoding.

d_train.head()

d_train.shape

(45000, 230)

Try building a model directly off this data (this will fail).

fitter = xgboost.XGBClassifier(n_estimators=10, max_depth=3, objective='binary:logistic')
try:
    fitter.fit(d_train, churn_train)
except Exception as ex:
    print(ex)

DataFrame.dtypes for data must be int, float or bool.
                Did not expect the data types in fields Var191, Var192, Var193, Var194, Var195, Var196, Var197, Var198, Var199, Var200, Var201, Var202, Var203, Var204, Var205, Var206, Var207, Var208, Var210, Var211, Var212, Var213, Var214, Var215, Var216, Var217, Var218, Var219, Var220, Var221, Var222, Var223, Var224, Var225, Var226, Var227, Var228, Var229

Let's quickly prepare a data frame with none of these issues.

We start by building our treatment plan, this has the sklearn.pipeline.Pipeline interfaces.

plan = vtreat.BinomialOutcomeTreatment(outcome_target=True)

Use .fit_transform() to get a special copy of the treated training data that has cross-validated mitigations againsst nested model bias. We call this a "cross frame." .fit_transform() is deliberately a different DataFrame than what would be returned by .fit().transform() (the .fit().transform() would damage the modeling effort due nested model bias, the .fit_transform() "cross frame" uses cross-validation techniques similar to "stacking" to mitigate these issues).

cross_frame = plan.fit_transform(d_train, churn_train)

Take a look at the new data. This frame is guaranteed to be all numeric with no missing values, with the rows in the same order as the training data.

cross_frame.head()

cross_frame.shape

(45000, 216)

Pick a recommended subset of the new derived variables.

plan.score_frame_.head()

model_vars = numpy.asarray(plan.score_frame_["variable"][plan.score_frame_["recommended"]])
len(model_vars)

216

Fit the model

cross_frame.dtypes

Var2_is_bad                            float64
Var3_is_bad                            float64
Var4_is_bad                            float64
Var5_is_bad                            float64
Var6_is_bad                            float64
                                  ...         
Var229_logit_code                      float64
Var229_prevalence_code                 float64
Var229_lev__NA_           Sparse[float64, 0.0]
Var229_lev_am7c           Sparse[float64, 0.0]
Var229_lev_mj86           Sparse[float64, 0.0]
Length: 216, dtype: object

# fails due to sparse columns
# can also work around this by setting the vtreat parameter 'sparse_indicators' to False
try:
    cross_sparse = xgboost.DMatrix(data=cross_frame.loc[:, model_vars], label=churn_train)
except Exception as ex:
    print(ex)

DataFrame.dtypes for data must be int, float or bool.
                Did not expect the data types in fields Var193_lev_RO12, Var193_lev_2Knk1KF, Var194_lev__NA_, Var194_lev_SEuy, Var195_lev_taul, Var200_lev__NA_, Var201_lev__NA_, Var201_lev_smXZ, Var205_lev_VpdQ, Var206_lev_IYzP, Var206_lev_zm5i, Var206_lev__NA_, Var207_lev_me75fM6ugJ, Var207_lev_7M47J5GA0pTYIFxg5uy, Var210_lev_uKAI, Var211_lev_L84s, Var211_lev_Mtgm, Var212_lev_NhsEn4L, Var212_lev_XfqtO3UdzaXh_, Var213_lev__NA_, Var214_lev__NA_, Var218_lev_cJvF, Var218_lev_UYBR, Var221_lev_oslk, Var221_lev_zCkv, Var225_lev__NA_, Var225_lev_ELof, Var225_lev_kG3k, Var226_lev_FSa2, Var227_lev_RAYp, Var227_lev_ZI9m, Var228_lev_F2FyR07IdsN7I, Var229_lev__NA_, Var229_lev_am7c, Var229_lev_mj86

# also fails
try:
    cross_sparse = scipy.sparse.csc_matrix(cross_frame[model_vars])
except Exception as ex:
    print(ex)

no supported conversion for types: (dtype('O'),)

# works
cross_sparse = scipy.sparse.hstack([scipy.sparse.csc_matrix(cross_frame[[vi]]) for vi in model_vars])

# https://xgboost.readthedocs.io/en/latest/python/python_intro.html
fd = xgboost.DMatrix(
    data=cross_sparse, 
    label=churn_train)

x_parameters = {"max_depth":3, "objective":'binary:logistic'}
cv = xgboost.cv(x_parameters, fd, num_boost_round=100, verbose_eval=False)

cv.head()

best = cv.loc[cv["test-error-mean"]<= min(cv["test-error-mean"] + 1.0e-9), :]
best

ntree = best.index.values[0]
ntree

21

fitter = xgboost.XGBClassifier(n_estimators=ntree, max_depth=3, objective='binary:logistic')
fitter

XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,
              colsample_bynode=1, colsample_bytree=1, gamma=0,
              learning_rate=0.1, max_delta_step=0, max_depth=3,
              min_child_weight=1, missing=None, n_estimators=21, n_jobs=1,
              nthread=None, objective='binary:logistic', random_state=0,
              reg_alpha=0, reg_lambda=1, scale_pos_weight=1, seed=None,
              silent=None, subsample=1, verbosity=1)

model = fitter.fit(cross_sparse, churn_train)

Apply the data transform to our held-out data.

test_processed = plan.transform(d_test)

Plot the quality of the model on training data (a biased measure of performance).

pf_train = pandas.DataFrame({"churn":churn_train})
pf_train["pred"] = model.predict_proba(cross_sparse)[:, 1]
wvpy.util.plot_roc(pf_train["pred"], pf_train["churn"], title="Model on Train")

0.7424056263753072

Plot the quality of the model score on the held-out data. This AUC is not great, but in the ballpark of the original contest winners.

test_sparse = scipy.sparse.hstack([scipy.sparse.csc_matrix(test_processed[[vi]]) for vi in model_vars])
pf = pandas.DataFrame({"churn":churn_test})
pf["pred"] = model.predict_proba(test_sparse)[:, 1]
wvpy.util.plot_roc(pf["pred"], pf["churn"], title="Model on Test")

0.7328696191869485

Notice we dealt with many problem columns at once, and in a statistically sound manner. More on the vtreat package for Python can be found here: https://github.com/WinVector/pyvtreat. Details on the R version can be found here: https://github.com/WinVector/vtreat.

We can compare this to the R solution (link).

We can compare the above cross-frame solution to a naive "design transform and model on the same data set" solution as we show below. Note we turn off filter_to_recommended as this is computed using cross-frame techniques (and hence is a non-naive estimate).

plan_naive = vtreat.BinomialOutcomeTreatment(
    outcome_target=True,              
    params=vtreat.vtreat_parameters({'filter_to_recommended':False}))
plan_naive.fit(d_train, churn_train)
naive_frame = plan_naive.transform(d_train)

naive_sparse = scipy.sparse.hstack([scipy.sparse.csc_matrix(naive_frame[[vi]]) for vi in model_vars])

fd_naive = xgboost.DMatrix(data=naive_sparse, label=churn_train)
x_parameters = {"max_depth":3, "objective":'binary:logistic'}
cvn = xgboost.cv(x_parameters, fd_naive, num_boost_round=100, verbose_eval=False)

bestn = cvn.loc[cvn["test-error-mean"]<= min(cvn["test-error-mean"] + 1.0e-9), :]
bestn

ntreen = bestn.index.values[0]
ntreen

94

fittern = xgboost.XGBClassifier(n_estimators=ntreen, max_depth=3, objective='binary:logistic')
fittern

XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,
              colsample_bynode=1, colsample_bytree=1, gamma=0,
              learning_rate=0.1, max_delta_step=0, max_depth=3,
              min_child_weight=1, missing=None, n_estimators=94, n_jobs=1,
              nthread=None, objective='binary:logistic', random_state=0,
              reg_alpha=0, reg_lambda=1, scale_pos_weight=1, seed=None,
              silent=None, subsample=1, verbosity=1)

modeln = fittern.fit(naive_sparse, churn_train)

test_processedn = plan_naive.transform(d_test)
test_processedn = scipy.sparse.hstack([scipy.sparse.csc_matrix(test_processedn[[vi]]) for vi in model_vars])

pfn_train = pandas.DataFrame({"churn":churn_train})
pfn_train["pred_naive"] = modeln.predict_proba(naive_sparse)[:, 1]
wvpy.util.plot_roc(pfn_train["pred_naive"], pfn_train["churn"], title="Overfit Model on Train")

0.9492686875296688

pfn = pandas.DataFrame({"churn":churn_test})
pfn["pred_naive"] = modeln.predict_proba(test_processedn)[:, 1]
wvpy.util.plot_roc(pfn["pred_naive"], pfn["churn"], title="Overfit Model on Test")

0.5960012412998182

Note the naive test performance is worse, despite its far better training performance. This is over-fit due to the nested model bias of using the same data to build the treatment plan and model without any cross-frame mitigations.

Solution Details

Some vreat data treatments are “y-aware” (use distribution relations between independent variables and the dependent variable).

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 DataFrame 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”). 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.

• 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).

The above are all awful things that often lurk in real world data. Automating mitigation 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 2nd edition of Practical Data Science with R covers using vtreat in R in chapter 8 "Advanced Data Preparation."

The idea is: DataFrames 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.

References

Some of our related articles (which should make clear some of our motivations, and design decisions):

• The vtreat technical paper.
• Modeling trick: impact coding of categorical variables with many levels
• A bit more on impact coding
• vtreat: designing a package for variable treatment
• A comment on preparing data for classifiers
• Nina Zumel presenting on vtreat

A directory of worked examples can be found here.

We intend to add better Python documentation and a certification suite going forward.

Installation

To install, please run:

# To install:
pip install vtreat

Some notes on controlling vtreat cross-validation can be found here.

Note on data types.

.fit_transform() expects the first argument to be a pandas.DataFrame with trivial row-indexing and scalar column names, (i.e. .reset_index(inplace=True, drop=True)) and the second to be a vector-like object with a len() equal to the number of rows of the first argument. We are working on supporting column types other than string and numeric at this time.