Skip to content

Latest commit

 

History

History
1282 lines (1015 loc) · 33.8 KB

Classification.md

File metadata and controls

1282 lines (1015 loc) · 33.8 KB
Raw

Using vtreat with Classification Problems

Nina Zumel and John Mount November 2019

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

Preliminaries

Load modules/packages.

import pkg_resources
import pandas
import numpy
import numpy.random
import seaborn
import matplotlib.pyplot as plt
import vtreat
import vtreat.util
import wvpy.util

numpy.random.seed(2019)

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 NaNs
  • Input x2 is a pure noise variable with no relationship to the output
def make_data(nrows):
    d = pandas.DataFrame({'x': 5*numpy.random.normal(size=nrows)})
    d['y'] = numpy.sin(d['x']) + 0.1*numpy.random.normal(size=nrows)
    d.loc[numpy.arange(3, 10), 'x'] = numpy.nan                           # introduce a nan level
    d['xc'] = ['level_' + str(5*numpy.round(yi/5, 1)) for yi in d['y']]
    d['x2'] = numpy.random.normal(size=nrows)
    d.loc[d['xc']=='level_-1.0', 'xc'] = numpy.nan  # introduce a nan level
    d['yc'] = d['y']>0.5
    return d

d = make_data(500)

d.head()
x y xc x2 yc
0 -1.088395 -0.956311 NaN -1.424184 False
1 4.107277 -0.671564 level_-0.5 0.427360 False
2 7.406389 0.906303 level_1.0 0.668849 True
3 NaN 0.222792 level_0.0 -0.015787 False
4 NaN -0.975431 NaN -0.491017 False 
outcome_name = 'yc'    # outcome variable / column
outcome_target = True  # value we consider positive

Some quick data exploration

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

d['xc'].unique()
array([nan, 'level_-0.5', 'level_1.0', 'level_0.0', 'level_-0.0',
       'level_0.5'], dtype=object)

d['xc'].value_counts(dropna=False)
level_1.0     140
NaN           109
level_-0.5    103
level_0.5      75
level_0.0      37
level_-0.0     36
Name: xc, dtype: int64

Find the prevalence of yc == True (our chosen notion of "positive").

numpy.mean(d[outcome_name] == outcome_target)
0.35

Plot of yc versus x.

seaborn.lineplot(x='x', y='yc', data=d)
<matplotlib.axes._subplots.AxesSubplot at 0x1a19866310>

png

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 NaNs.

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

transform = vtreat.BinomialOutcomeTreatment(
    outcome_name=outcome_name,      # outcome variable
    outcome_target=outcome_target,  # outcome of interest
    cols_to_copy=['y'],  # columns to "carry along" but not treat as input variables
)

Use the training data d to fit the transform and the return a treated training set: completely numeric, with no missing values. Note that for the training data d: transform.fit_transform() is not the same as transform.fit().transform(); 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.transform(o) is an appropriate step.

d_prepared = transform.fit_transform(d, d['yc'])

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.

transform.score_frame_
variable orig_variable treatment y_aware has_range PearsonR R2 significance vcount default_threshold recommended
0 x_is_bad x missing_indicator False True -0.051749 0.002388 0.213707 2.0 0.10 False
1 xc_is_bad xc missing_indicator False True -0.387438 0.169451 0.000000 2.0 0.10 True
2 x x clean_copy False True 0.052826 0.002158 0.237141 2.0 0.10 False
3 x2 x2 clean_copy False True 0.069126 0.003709 0.121205 2.0 0.10 False
4 xc_logit_code xc logit_code True True 0.867915 0.834531 0.000000 1.0 0.20 True
5 xc_prevalence_code xc prevalence_code False True 0.567968 0.341474 0.000000 1.0 0.20 True
6 xc_lev_level_1_0 xc indicator_code False True 0.849837 0.645319 0.000000 4.0 0.05 True
7 xc_lev__NA_ xc indicator_code False True -0.387438 0.169451 0.000000 4.0 0.05 True
8 xc_lev_level_-0_5 xc indicator_code False True -0.373767 0.158566 0.000000 4.0 0.05 True
9 xc_lev_level_0_5 xc indicator_code False True 0.102752 0.007894 0.023774 4.0 0.05 True

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

  • an indicator variable for each possible level (xc_lev_level_*)
  • the value of a (cross-validated) one-variable model for yc as a function of xc (xc_logit_code)
  • a variable that returns how prevalent this particular value of xc is in the training data (xc_prevalence_code)
  • a variable indicating when xc was NaN in the original data (xc_is_bad, x_is_bad)

Any or all of these new variables are available for downstream modeling. x doesn't show as exciting a significance as xc, as we are only checking linear relations, and x is related to y in a very non-linear way.

The recommended column indicates which variables are non constant (has_range == True) and have a significance value smaller than default_threshold. See the section Deriving the Default Thresholds below for the reasoning behind the default thresholds. 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
transform.score_frame_.loc[transform.score_frame_['recommended'], ['variable']]
variable
1 xc_is_bad
4 xc_logit_code
5 xc_prevalence_code
6 xc_lev_level_1_0
7 xc_lev__NA_
8 xc_lev_level_-0_5
9 xc_lev_level_0_5 
# not recommended variables
transform.score_frame_.loc[~transform.score_frame_['recommended'], ['variable']]
variable
0 x_is_bad
2 x
3 x2

Notice that d_prepared only includes recommended variables (along with y and yc):

d_prepared.head()
y yc xc_is_bad xc_logit_code xc_prevalence_code xc_lev_level_1_0 xc_lev__NA_ xc_lev_level_-0_5 xc_lev_level_0_5
0 -0.956311 False 1.0 -5.838582 0.218 0.0 1.0 0.0 0.0
1 -0.671564 False 0.0 -5.837275 0.206 0.0 0.0 1.0 0.0
2 0.906303 True 0.0 1.047934 0.280 1.0 0.0 0.0 0.0
3 0.222792 False 0.0 -5.790939 0.074 0.0 0.0 0.0 0.0
4 -0.975431 False 1.0 -5.837541 0.218 0.0 1.0 0.0 0.0

This is vtreats default behavior; to include all variables in the prepared data, set the parameter filter_to_recommended to False, as we show later, in the Parameters for BinomialOutcomeTreatment section below.

A Closer Look at logit_code variables

Variables of type logit_code are the outputs of a one-variable hierarchical 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_logit_code makes a good one-variable model for yc. It has a large AUC:

wvpy.util.plot_roc(prediction=d_prepared['xc_logit_code'], 
                   istrue=d_prepared['yc'],
                   title = 'performance of xc_logit_code variable')

png

0.9862945054945056

This indicates that xc_logit_code is strongly predictive of the outcome. Negative values of xc_logit_code correspond strongly to negative outcomes, and positive values correspond strongly to positive outcomes.

wvpy.util.dual_density_plot(probs=d_prepared['xc_logit_code'], 
                            istrue=d_prepared['yc'])

png

The values of xc_logit_code are in "link space". We can often visualize the relationship a little better by converting the logistic score to a probability.

from scipy.special import expit  # sigmoid
from scipy.special import logit

offset = logit(numpy.mean(d_prepared.yc))
wvpy.util.dual_density_plot(probs=expit(d_prepared['xc_logit_code'] + offset),
                            istrue=d_prepared['yc'])

png

Variables of type logit_code 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 logit_code 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.

import sklearn.linear_model
import seaborn

not_variables = ['y', 'yc', 'prediction']
model_vars = [v for v in d_prepared.columns if v not in set(not_variables)]

fitter = sklearn.linear_model.LogisticRegression()
fitter.fit(d_prepared[model_vars], d_prepared['yc'])

# now predict
d_prepared['prediction'] = fitter.predict_proba(d_prepared[model_vars])[:, 1]

# look at the ROC curve (on the training data)
wvpy.util.plot_roc(prediction=d_prepared['prediction'], 
                   istrue=d_prepared['yc'],
                   title = 'Performance of logistic regression model on training data')
/Users/johnmount/opt/anaconda3/envs/ai_academy_3_7/lib/python3.7/site-packages/sklearn/linear_model/logistic.py:432: FutureWarning: Default solver will be changed to 'lbfgs' in 0.22. Specify a solver to silence this warning.
  FutureWarning)

png

0.9862945054945055

Now apply the model to new data.

# create the new data
dtest = make_data(450)

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

# apply the model to the prepared data
dtest_prepared['prediction'] = fitter.predict_proba(dtest_prepared[model_vars])[:, 1]

wvpy.util.plot_roc(prediction=dtest_prepared['prediction'], 
                   istrue=dtest_prepared['yc'],
                   title = 'Performance of logistic regression model on test data')

png

0.9726541064569234

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.

vtreat.vtreat_parameters()
{'use_hierarchical_estimate': True,
 'coders': {'clean_copy',
  'deviation_code',
  'impact_code',
  'indicator_code',
  'logit_code',
  'missing_indicator',
  'prevalence_code'},
 'filter_to_recommended': True,
 'indicator_min_fraction': 0.1,
 'cross_validation_plan': vtreat.cross_plan.KWayCrossPlanYStratified(),
 'cross_validation_k': 5,
 'user_transforms': [],
 'sparse_indicators': True,
 'missingness_imputation': <function numpy.mean(a, axis=None, dtype=None, out=None, keepdims=<no value>)>,
 'check_for_duplicate_frames': True,
 'retain_cross_plan': False}

use_hierarchical_estimate:: When True, uses hierarchical smoothing when estimating logit_code variables; when False, uses unsmoothed logistic regression.

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

filter_to_recommended: When True, prepared data only includes variables marked as "recommended" in score frame. When False, prepared data includes all variables. See the Example below.

indicator_min_fraction: For categorical variables, indicator variables (type indicator_code) are only produced for levels that are present at least indicator_min_fraction of the time. A consequence of this is that 1/indicator_min_fraction 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 indicator_min_fraction = 0

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

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

user_transforms: 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.

sparse_indicators: When True, use a (Pandas) sparse representation for indicator variables. This representation is compatible with sklearn; however, it may not be compatible with other modeling packages. When False, use a dense representation.

missingness_imputation The function or value that vtreat uses to impute or "fill in" missing numerical values. The default is numpy.mean(). To change the imputation function or use different functions/values for different columns, see the Imputation example.

Example: Use all variables to model, not just recommended

transform_all = vtreat.BinomialOutcomeTreatment(
    outcome_name='yc',    # outcome variable
    outcome_target=True,  # outcome of interest
    cols_to_copy=['y'],   # columns to "carry along" but not treat as input variables
    params = vtreat.vtreat_parameters({
        'filter_to_recommended': False
    })
)  

transform_all.fit_transform(d, d['yc']).columns
Index(['y', 'yc', 'x_is_bad', 'xc_is_bad', 'x', 'x2', 'xc_logit_code',
       'xc_prevalence_code', 'xc_lev_level_1_0', 'xc_lev__NA_',
       'xc_lev_level_-0_5', 'xc_lev_level_0_5'],
      dtype='object')

transform_all.score_frame_
variable orig_variable treatment y_aware has_range PearsonR R2 significance vcount default_threshold recommended
0 x_is_bad x missing_indicator False True -0.051749 0.002388 0.213707 2.0 0.10 False
1 xc_is_bad xc missing_indicator False True -0.387438 0.169451 0.000000 2.0 0.10 True
2 x x clean_copy False True 0.052826 0.002158 0.237141 2.0 0.10 False
3 x2 x2 clean_copy False True 0.069126 0.003709 0.121205 2.0 0.10 False
4 xc_logit_code xc logit_code True True 0.867909 0.831229 0.000000 1.0 0.20 True
5 xc_prevalence_code xc prevalence_code False True 0.567968 0.341474 0.000000 1.0 0.20 True
6 xc_lev_level_1_0 xc indicator_code False True 0.849837 0.645319 0.000000 4.0 0.05 True
7 xc_lev__NA_ xc indicator_code False True -0.387438 0.169451 0.000000 4.0 0.05 True
8 xc_lev_level_-0_5 xc indicator_code False True -0.373767 0.158566 0.000000 4.0 0.05 True
9 xc_lev_level_0_5 xc indicator_code False True 0.102752 0.007894 0.023774 4.0 0.05 True

Note that the prepared data produced by fit_transform() includes all the variables, including those that were not marked as "recommended".

Types of prepared variables

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

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

prevalence_code: Produced from categorical variables: indicates how often each level of the variable was "on"

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

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

deviation_code: not used by BinomialOutcomeTreatment

impact_code: not used by BinomialOutcomeTreatment

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 logit_code or prevalence_code variables.

transform_thin = vtreat.BinomialOutcomeTreatment(
    outcome_name='yc',    # outcome variable
    outcome_target=True,  # outcome of interest
    cols_to_copy=['y'],   # columns to "carry along" but not treat as input variables
    params = vtreat.vtreat_parameters({
        'filter_to_recommended': False,
        'coders': {'clean_copy',
                   'missing_indicator',
                   'indicator_code',
                  }
    })
)

transform_thin.fit_transform(d, d['yc']).head()
y yc x_is_bad xc_is_bad x x2 xc_lev_level_1_0 xc_lev__NA_ xc_lev_level_-0_5 xc_lev_level_0_5
0 -0.956311 False 0.0 1.0 -1.088395 -1.424184 0.0 1.0 0.0 0.0
1 -0.671564 False 0.0 0.0 4.107277 0.427360 0.0 0.0 1.0 0.0
2 0.906303 True 0.0 0.0 7.406389 0.668849 1.0 0.0 0.0 0.0
3 0.222792 False 1.0 0.0 -0.057044 -0.015787 0.0 0.0 0.0 0.0
4 -0.975431 False 1.0 1.0 -0.057044 -0.491017 0.0 1.0 0.0 0.0 
transform_thin.score_frame_
variable orig_variable treatment y_aware has_range PearsonR R2 significance vcount default_threshold recommended
0 x_is_bad x missing_indicator False True -0.051749 0.002388 0.213707 2.0 0.166667 False
1 xc_is_bad xc missing_indicator False True -0.387438 0.169451 0.000000 2.0 0.166667 True
2 x x clean_copy False True 0.052826 0.002158 0.237141 2.0 0.166667 False
3 x2 x2 clean_copy False True 0.069126 0.003709 0.121205 2.0 0.166667 True
4 xc_lev_level_1_0 xc indicator_code False True 0.849837 0.645319 0.000000 4.0 0.083333 True
5 xc_lev__NA_ xc indicator_code False True -0.387438 0.169451 0.000000 4.0 0.083333 True
6 xc_lev_level_-0_5 xc indicator_code False True -0.373767 0.158566 0.000000 4.0 0.083333 True
7 xc_lev_level_0_5 xc indicator_code False True 0.102752 0.007894 0.023774 4.0 0.083333 True

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.

We noticed, however, that this biases the filtering against numerical variables, since there are at most two derived variables (of types clean_copy and missing_indicator for every numerical variable in the original data. Categorical variables, on the other hand, are expanded to many derived variables: several indicators (one for every common level), plus a logit_code and a prevalence_code. So we now reweight the thresholds.

Suppose you have a (treated) data set with ntreat different types of vtreat variables (clean_copy, indicator_code, etc). There are nT variables of type T. Then the default threshold for all the variables of type T is 1/(ntreat nT). This reweighting helps to reduce the bias against any particular type of variable. The heuristic is still that the set of recommended variables will allow at most one noise variable into the set of candidate variables.

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. Setting the filter_to_recommended parameter to False will keep all derived variables in the treated frame, for the data scientist to filter (or not) as they will.

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:

Some vtreat common capabilities are documented here:

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.