Variable Selection with Information Criteria

Every data science project hits the point where you investigate which subset of variables you should include in your model.

If you have a handful of variables, you can easily do exhaustive variable selection. But if you have hundreds, or thousands, of possible variables to choose from, you may want to find a way to select the best subset to start with.

This app provides a graphical interface that applies several information criteria metrics to a data set, and assesses how predictive the variables are in relation to a defined target.

The app will provide several pieces of functionality on a toy data set, the mtcars data set.

1. A sidebar to select the relevant target and predictor columns
2. A tab to view information criteria metrics to evaluate predictivity of variables
3. A tab to view exploratory plots of relevant variables to the target variable
4. A tab to view the raw data

Several methods are available for assessing predictivity of variables: Akaike Information Criteria, Bayesian Information Criteria, and Adjusted R-Squared

In this app, these metrics are applied to a simple univariate linear model for each predictor, according to the example code below. In this case, mpg is the target variable.

# AIC 
attach(mtcars)
AIC(lm(mpg ~ wt))
AIC(lm(mpg ~ drat))
# Way to go weight!

## [1] 166.0294
## [1] 190.7999

#Adjusted R-Squared
attach(mtcars)

## The following objects are masked from mtcars (pos = 3):
## 
##     am, carb, cyl, disp, drat, gear, hp, mpg, qsec, vs, wt

summary(lm(mpg ~ wt))$adj.r.sq
summary(lm(mpg ~ drat))$adj.r.sq
# Way to go weight!

## [1] 0.7445939
## [1] 0.4461283

The app is located on shinyapps.io here

General Usage of the App can be found by selecting the "Description" tab.

The code for the shiny app and this presentation can be located at my github page

Feel free to play around with the app, and leave any comments!