Introduction

Binary classification models are perhaps the most common use-case in predictive analytics. The reason is that many key client actions across a wide range of industries are binary in nature, such as defaulting on a loan, clicking on an ad, or terminating a subscription.

Prior to building a binary classification model, a common step is to perform variable screening and exploratory data analysis. This is the step where we get to know the data and weed out variables that are either ill-conditioned or simply contain no information that will help us predict the action of interest. Note that the purpose of this step should not to be confused with that of multiple-variable selection techniques, such as stepwise regression and lasso, where the variables that go into the final model are selected. Rather, this is a precursory step designed to ensure that the approaches deployed during the final modeling phases are set up for success.

The weight of evidence (WOE) and information value (IV) provide a great framework for for exploratory analysis and variable screening for binary classifiers. WOE and IV have been used extensively in the credit risk world for several decades, and the underlying theory dates back to the 1950s. However, it is still not widely used outside the credit risk world.

WOE and IV enable one to:

• Consider each variable’s independent contribution to the outcome.
• Detect linear and non-linear relationships.
• Rank variables in terms of “univariate” predictive strength.
• Visualize the correlations between the predictive variables and the binary outcome.
  
• Seamlessly compare the strength of continuous and categorical variables without creating dummy variables.
• Seamlessly handle missing values without imputation.
• Assess the predictive power of missing values.

In this paper we will provide a brief description of WOE and IV as well as the Information R package, which provides an easy way to perform variable screening and exploratory analysis with WOE and IV for traditional binary classifiers as well as uplift models.

Information Theory

Weight of evidence (WOE) and information value are closely related to concepts from information theory where one of the goals is to understand the uncertainty involved in predicting the outcome of random events given varying degrees of knowledge of other variables (see [2], [3], and [4]). Thus this is a perfect framework for variable screening and exploratory analysis for predictive modeling.

Information theory kicked off when Claude Shannon (1948) wrote down an expression for the entropy of a probability distribution. Entropy is often described as the level of “disorder,” but in the context of information theory it is better to think of it as one’s level of uncertainty. High entropy means high uncertainty (you don’t know what tomorrow holds) and low entropy mean low uncertainty (there will be no surprises tomorrow).

Suppose a friend tosses a coin and asks you to predict whether the toss came up heads or tails. There’s a 50 percent probability of each possibility, so your prediction isn’t worth very much. This turns out to be the maximum amount of entropy for an event with two possible outcomes. Suppose instead that you know the coin is weighted so that it comes up heads 90 percent of the time. Now your job of making a prediction is easier – Shannon’s formula tells you exactly how much information you’ve gained compared to the previous situation.

Mutual information is a way of summarizing much knowing the value of one random variable tells you about another random variable (Shannon + Weaver 1949). If \( X \) and \( Y \) are independent, the mutual information is zero. Any correlations (positive, negative, or nonlinear) will result in positive mutual information. It turns out the mutual information of \( X \) and \( Y \) is nicely expressed in terms of entropy (\( H \)):

MI(X, Y) = H(X) + H(Y) − H(X, Y).

Mutual information is the difference in predictability that you get from knowing the joint distribution \( p(X,Y) \) compared to knowing only the two marginal distributions \( p(X) \) and \( p(Y) \).

Information value (IV), which we will define below, is quite closely connected to mutual information (MI), and both are intended to do the same thing: summarize how much knowing the value of \( X \) in a given trial helps you predict the value of \( Y \) for the same trial.

Defining WOE and IV

Let’s say that we have a binary dependent variable \( Y \) and a set of predictive variables \( X_1, , X_p \). As mentioned above, \( Y \) can capture a wide range of outcomes, such as defaulting on a loan, clicking on an ad, or terminating a subscription.

WOE and IV play two distinct roles when analyzing data:

• WOE describes the relationship between a predictive variable and a binary target variable.
• IV measures the strength of that relationship.

Extensions to Exploratory Analysis for Uplift Models

Consider a direct marketing program where a test group received an offer of some sort and the control group did not receive anything. The test and control groups are based on a random split. The lift of the campaign is defined as the difference in success rates between the test and control groups. In other words, the program can only be deemed successful if the offer outperforms the “do nothing” (a.k.a baseline) scenario.

The purpose of uplift models is to estimate the difference between the test and control groups and then using the resulting model to target persuadables – i.e., potential or existing clients that are on the fence and need some sort of offer to buy a product. Thus, when preparing to build an uplift model, we cannot only focus on the log odds of \( Y=1 \), we also need to consider the log odds ratio of \( Y=1 \) for the test group versus the control group. This leads to the net weight of evidence (NWOE) and the net information value (NIV). (Actual uplift modeling techniques are beyond the scope of this paper; we will stick to model exploration and variable screening.)

The net weight of evidence (NWOE) is the difference between the WOEs for the test group and the control group

NWOE = WOE_t − WOE_c,

which is equivalent to

$$ \text{NWOE}_j = \log \frac{f(x_j|Y=1)_{t}f(x_j|Y=0)_{c}}{f(x_j|Y=1)_{c}f(x_j|Y=0)_{t}}. \nonumber $$

The net information value for variable \( X_j \) is then defined as

$$ 10 \times \int \left( f(X_j|Y=1)_{t}f(X_j|Y=0)_{c} - f(X_j|Y=1)_{c}f(X_j|Y=0)_{t} \right) \times \log \frac{f(X_j|Y=1)_{t}f(X_j|Y=0)_{c}}{f(X_j|Y=1)_{c}f(X_j|Y=0)_{t}} dX_j. $$

where multiplication of 10 is simply used to make NIV look more like IV. Note that NWOE and NIV work just like IV and WOE, just from an uplift perspective. NIV measures the strength of a given variable while NWOE describes the pattern of the relationship. Specifically, the higher the value of NIV, the better the given variable is at separating self-selectors – i.e., people who are self-motivated to buy – and persuadables that need to be motivated.

The Information R Package

The Information package is designed to perform exploratory data analysis and variable screening for binary classification models using WOE and IV. To make the package as efficient as possible aggregations are done in data.table and creation of WOE vectors can be distributed across multiple cores.

There are a number of R packages available that support WOE and IV (see a partial list below), but they are either primarily designed to build WOE-based models – e.g., naive Bayes classifiers – or they do not support the uplift use-case. The Information package is specifically designed to perform data exploration by producing easy-to-read tables and graphs.

library(Information)
options(scipen=10)

### Loading the data
data(train, package="Information")
data(valid, package="Information")


### Exclude the control group
train <- subset(train, TREATMENT==1)

### Ranking variables using penalized IV. 
## Using ncore=2 because more than 2 is no allowed by CRAN in examples  
## For real applications, leave ncore as NULL to get the default
IV <- create_infotables(data=train, 
                       valid=valid,
                       y="PURCHASE",
                       parallel=FALSE)

knitr::kable(head(IV$Summary))

### Note that the elements of the list IV$Tables are named according to the variable names
names <- names(IV$Tables)

knitr::kable(IV$Tables$N_OPEN_REV_ACTS)

plot_infotables(IV, "N_OPEN_REV_ACTS", show_values=TRUE)

plot_infotables(IV, IV$Summary$Variable[1:4], same_scales=TRUE)

# Get the names and loop through to create individual plots
names <- names(IV$Tables)
plots <- list()
for (i in 1:length(names)){
   plots[[i]] <- plot_infotables(IV, names[i])
}

# Showing the top 18 variables
plots[1:4]

IV <- create_infotables(data=train, 
                   y="PURCHASE")

# Note that this example is running on parallel mode
IV <- create_infotables(data=train, 
                   valid=valid, 
                   y="PURCHASE", 
                   bins=20)

## Read the datasets, including the control group this time
data(train, package="Information")
data(valid, package="Information")

## This time use cross validation
NIV <- create_infotables(data=train, 
                  valid=valid, 
                  y="PURCHASE", 
                  trt="TREATMENT", 
                  parallel=FALSE)
                  
knitr::kable(head(NIV$Summary))

knitr::kable(NIV$Tables$N_OPEN_REV_ACTS)

plot_infotables(NIV, "N_OPEN_REV_ACTS")

library(ClustOfVar)
library(reshape2)
library(plyr)

data(train, package="Information")
data(valid, package="Information")
train <- subset(train, TREATMENT==1)
valid <- subset(valid, TREATMENT==1)

tree <- hclustvar(train[,!(names(train) %in% c("PURCHASE", "TREATMENT"))])
nvars <- length(tree[tree$height<0.7])
part_init<-cutreevar(tree,nvars)$cluster
kmeans<-kmeansvar(X.quanti=train[,!(names(train) %in% c("PURCHASE", "TREATMENT"))],init=part_init)
clusters <- cbind.data.frame(melt(kmeans$cluster), row.names(melt(kmeans$cluster)))
names(clusters) <- c("Cluster", "Variable")
clusters <- join(clusters, IV$Summary, by="Variable", type="left")
clusters <- clusters[order(clusters$Cluster),]
clusters$Rank <- stats::ave(-clusters$AdjIV, clusters$Cluster, FUN=rank)
selected_members <- subset(clusters, Rank==1)
selected_members$Rank <- NULL

# Using variable clustering in combination with IV cuts the number of variables from 68 to 21.
print(selected_members, row.names=FALSE)

Summary

• The purpose of exploratory analysis and variable screening is to get to know the data and assess “univariate” predictive strength, before we deploy more sophisticated variable selection approaches.

• The weight of evidence (WOE) and information value (IV) provide a great framework for performing exploratory analysis and variable screening prior to building a binary classifier (e.g., logistic regression). It seamlessly handles missing values and character variables, and the output is easy to interpret.

• The information value originates from information theory and is closely related to the concept of *mutual information (see [3], [4]).

• The Information package is specifically written to perform this type of analysis using parallel processing. It also supports exploratory analysis for uplift models, a growing area within marketing analytics. The Information package is not designed to transfer data into WOE vectors for Naive Bayes models, although this feature could be added later.

References

[1] Hastie, Trevor, Tibshirani, Robert and Friedman, Jerome. (1986), Elements of Statistical Learning, Second Edition, Springer, 2009.

[2] Kullback S., Information Theory and Statistics, John Wiley and Sons, 1959.

[3] Shannon, C.E., A Mathematical Theory of Communication, Bell System Technical Journal, 1948.

[4] Shannon, CE. and Weaver, W. The Mathematical Theory of Communication. Univ of Illinois Press, 1949.

[5] GAM: the Predictive Modeling Silver Bullet, (via)

Other Packages that support information theory

woe package, (link)

klaR package, (link)

infotheo package, (link)