modeler

modeler is an R Package for helping users perform modeling procedures. There are 4 main types of analysis in this package.

1. Exploratory/Summary
2. Variable Transformations
3. Examining Relationships
4. Measuring Model Performance

Most ideas and functions in this package aren't novel but have been written to make actions that I use frequently easier to use. A summary of the functions in each section is included below.

Exploratory Analysis

Everyone has their own EDA steps and ideas but these are some I use a lot.

peruse

peruse examines the variables in a data frame and returns basic summary info about the individual variables.

cars_summary <- peruse(mtcars)
head(cars_summary)
#    Variable   Class    Type Num_Missing Num_Unique             data
#       <chr>   <chr>   <chr>       <chr>      <chr>           <list>
# 1       mpg Numeric Numeric           0         25 <tibble [1 x 7]>
# 2       cyl Numeric Numeric           0          3 <tibble [1 x 7]>
# 3      disp Numeric Numeric           0         27 <tibble [1 x 7]>

Extra information is returned in the data column but will differ between the Variable types. Simply use tidyr's unnest function to view the extra data.

tidyr::unnest(cars_summary, data) %>% head(3)
#   Variable   Class    Type Num_Missing Num_Unique First_Quartile   Max  Mean
#      <chr>   <chr>   <chr>       <chr>      <chr>          <chr> <chr> <chr>
# 1      mpg Numeric Numeric           0         25          15.42  33.9 20.09
# 2      cyl Numeric Numeric           0          3              4     8 6.188
# 3     disp Numeric Numeric           0         27          120.8   472 230.7
#   Median   Min               SD Third_Quartile
#    <chr> <chr>            <chr>          <chr>
# 1   19.2  10.4  6.0269480520891           22.8
# 2      6     4 1.78592164694654              8
# 3  196.3  71.1 123.938693831382            326

The profile function will preform this on a single variable.

sample_groups

sample_groups allows you to sample at the group level. So instead of using sample_n from dplyr to sample n observations you can use sample_groups to return a certain number of groups. This is usually applied when you have nested or hierarchical data. For example if I have a data set with attendance information at a team-game-fan level I may want to return all attendance observations from only a few games (for close data inspection or plotting). I can use sample_groups to accomplish this in one step.

single_iris <- sample_groups(iris, Species, n = 1)
table(single_iris$Species)
# setosa versicolor  virginica 
#      0         50          0 

helpers

I've also included various helper functions:

• deciles - returns the deciles of a numeric vector
• how_many_nas - returns how many nas are in each column of a data frame
• multiplot - plot multiple plots as one
• tableNA - table function that includes NA
• view - like head or tail but returns a random number of observations
• n_combos - counts the number of combinations between different columns in a data frame
• grouped_arrange - includes grouped columns in the arrange call

Variable Transformations

These aren't necessarily typical regression transformations like log or polynomial transformations.

add_pca

add_pca will append a data frame with the pca loadings from a specified set of variables. You can include the new column names (or let them default to .pc1, .pc2, etc...) and specifiy the number of loadings to return (default is to include all).

add_pca(mtcars, mpg:wt, new_column = "car_specs", n = 3) %>% head(3)
#    mpg cyl disp  hp drat    wt  qsec vs am gear carb car_specs.pc1
# 1 21.0   6  160 110 3.90 2.620 16.46  0  1    4    4    -1.0177186
# 2 21.0   6  160 110 3.90 2.875 17.02  0  1    4    4    -0.9093556
# 3 22.8   4  108  93 3.85 2.320 18.61  1  1    4    1    -1.9968043
#   car_specs.pc2 car_specs.pc3
# 1    -0.1514974   -0.02970885
# 2    -0.1032240    0.14450082
# 3     0.2073380    0.07414376

impact_code

impact_code implements some of the methods in this paper to deal with high cardinality categorical variables. The basic thought is that it finds the average of the dependent variable for each level and then shrinks that estimate to the overall average depending on how much evidence is in the data. This is a data processing step that should be done on calibration data and then applied on a training data set to be used in model building.

impact_code(mtcars, am ~ cyl)
#     cyl  estimate
#   <dbl>     <dbl>
# 1     4 0.6844089
# 2     6 0.4294859
# 3     8 0.1771094
dplyr::group_by(mtcars, cyl) %>% dplyr::summarise(mean_am = mean(am))
#     cyl   mean_am
#   <dbl>     <dbl>
# 1     4 0.7272727
# 2     6 0.4285714
# 3     8 0.1428571

So the impact code estimates are all shrinked towards the overall mean of 40.6%

model_matrix

model_matrix is a wrapper around model.matrix but will preserve all levels of character or factor variables.

model.matrix(Petal.Width ~ Petal.Length + Species, data = iris) %>% head(3)
#   (Intercept) Petal.Length Speciesversicolor Speciesvirginica
# 1           1          1.4                 0                0
# 2           1          1.4                 0                0
# 3           1          1.3                 0                0
model_matrix(Petal.Width ~ Petal.Length + Species, .data = iris) %>% head(3)
#   (Intercept) Petal.Length Speciessetosa Speciesversicolor Speciesvirginica
# 1           1          1.4             1                 0                0
# 2           1          1.4             1                 0                0
# 3           1          1.3             1                 0                0

Variable Relationships

find_woe

Weight of Evidence and Information Value are modeling tools that allow you to measure the strength of the relationship between independent variables and a dependent variable. This measure can detect non-linear relationships and handles categorical variables and missing values.

find_woe(mtcars, mpg:qsec, y = am)
#   variable       iv              data
#      <chr>    <dbl>            <list>
# 1      mpg 1.875599 <tibble [10 x 2]>
# 2      cyl 1.067032  <tibble [3 x 2]>
# 3     disp 1.959199 <tibble [10 x 2]>
# 4       hp 1.341193 <tibble [10 x 2]>
# 5     drat 2.513143 <tibble [10 x 2]>
# 6       wt 1.995249 <tibble [10 x 2]>
# 7     qsec 1.201222 <tibble [10 x 2]>

The data column contains the actual splits used.

find_woe(mtcars, mpg:qsec, y = am) %>% 
  dplyr::slice(2) %>% 
  tidyr::unnest(data)
#   variable       iv   bins         woe
#      <chr>    <dbl> <fctr>       <dbl>
# 1      cyl 1.067032      4  1.23357943
# 2      cyl 1.067032      6  0.09496181
# 3      cyl 1.067032      8 -1.26316168

acf_by_group

A lot of data I work with are time series in nature but also nested by groups. In order to examine the nested time series trends acf_by_group will find the auto correlation function by group.

pres_ratings <- data.frame(approval = presidents, 
                           pre_1965 = c(rep(1, 60), rep(0, 60)))
ratings_acf <- acf_by_group(pres_ratings, pre_1965, approval, na.action = na.pass)
#   pre_1965 acf_values   lag
#      <dbl>      <dbl> <int>
# 1        1  1.0000000     0
# 2        1  0.7620363     1
# 3        1  0.6290838     2

acf_by_group uses the stats::acf function so you can pass any extra arguments you want

ratings_pac <- acf_by_group(pres_ratings, pre_1965, approval, na.action = na.pass, type = "partial")
#   pre_1965 acf_values   lag
#      <dbl>      <dbl> <int>
# 1        1  0.7620363     0
# 2        1  0.1153933     1
# 3        1 -0.2246557     2

Right now you actually don't need to provide a grouping variable, it will just return the auto-correlation values for the whole data frame.