Exploratory Data Analysis & Data Preparation with 'funModeling'

by Pablo Casas

funModeling quick-start

This package contains a set of functions related to exploratory data analysis, data preparation, and model performance. It is used by people coming from business, research, and teaching (professors and students).

funModeling is intimately related to the Data Science Live Book -Open Source- (2017) in the sense that most of its functionality is used to explain different topics addressed by the book.

📗 The paperback version is being prepared, get notified by the newsletter or twitter.

Opening the black-box

Some functions have in-line comments so the user can open the black-box and learn how it was developed, or to tune or improve any of them.

All the functions are well documented, explaining all the parameters with the help of many short examples. R documentation can be accessed by: help("name_of_the_function").

Important changes from latest version 1.6.7, (relevant only if you were using previous versions):

From the latest version, 1.6.7 (Jan 21-2018), the parameters str_input, str_target and str_score will be renamed to input, target and score respectively. The functionality remains the same.

If you were using these parameters names on production, they will be still working until next release. this means that for now, you can use for example str_input or input.

The other important change was in discretize_get_bins, which is detailed later in this document.


About this quick-start

This quick-start is focused only on the functions. All explanations around them, and the how and when to use them, can be accessed by following the "***Read more here.***" links below each section, which redirect you to the book.

Below there are most of the funModeling functions divided by category.

Exploratory data analysis

df_status: Dataset health status

Use case: analyze the zeros, missing values (NA), infinity, data type, and number of unique values for a given dataset.

library(funModeling)

df_status(heart_disease)

##                  variable q_zeros p_zeros q_na p_na q_inf p_inf    type
## 1                     age       0    0.00    0 0.00     0     0 integer
## 2                  gender       0    0.00    0 0.00     0     0  factor
## 3              chest_pain       0    0.00    0 0.00     0     0  factor
## 4  resting_blood_pressure       0    0.00    0 0.00     0     0 integer
## 5       serum_cholestoral       0    0.00    0 0.00     0     0 integer
## 6     fasting_blood_sugar     258   85.15    0 0.00     0     0  factor
## 7         resting_electro     151   49.83    0 0.00     0     0  factor
## 8          max_heart_rate       0    0.00    0 0.00     0     0 integer
## 9             exer_angina     204   67.33    0 0.00     0     0 integer
## 10                oldpeak      99   32.67    0 0.00     0     0 numeric
## 11                  slope       0    0.00    0 0.00     0     0 integer
## 12      num_vessels_flour     176   58.09    4 1.32     0     0 integer
## 13                   thal       0    0.00    2 0.66     0     0  factor
## 14 heart_disease_severity     164   54.13    0 0.00     0     0 integer
## 15           exter_angina     204   67.33    0 0.00     0     0  factor
## 16      has_heart_disease       0    0.00    0 0.00     0     0  factor
##    unique
## 1      41
## 2       2
## 3       4
## 4      50
## 5     152
## 6       2
## 7       3
## 8      91
## 9       2
## 10     40
## 11      3
## 12      4
## 13      3
## 14      5
## 15      2
## 16      2

[🔎 Read more here.]


plot_num: Plotting distributions for numerical variables

Plots only numeric variables.

plot_num(heart_disease)

Notes:

  • bins: Sets the number of bins (10 by default).
  • path_out indicates the path directory; if it has a value, then the
    plot is exported in jpeg. To save in current directory path must be
    dot: "."

[🔎 Read more here.]


profiling_num: Calculating several statistics for numerical variables

Retrieves several statistics for numerical variables.

profiling_num(heart_disease)

##                 variable   mean std_dev variation_coef p_01 p_05 p_25
## 1                    age  54.44    9.04           0.17   35   40   48
## 2 resting_blood_pressure 131.69   17.60           0.13  100  108  120
## 3      serum_cholestoral 246.69   51.78           0.21  149  175  211
## 4         max_heart_rate 149.61   22.88           0.15   95  108  134
## 5            exer_angina   0.33    0.47           1.44    0    0    0
## 6                oldpeak   1.04    1.16           1.12    0    0    0
## 7                  slope   1.60    0.62           0.38    1    1    1
## 8      num_vessels_flour   0.67    0.94           1.39    0    0    0
## 9 heart_disease_severity   0.94    1.23           1.31    0    0    0
##    p_50  p_75  p_95  p_99 skewness kurtosis  iqr        range_98
## 1  56.0  61.0  68.0  71.0    -0.21      2.5 13.0        [35, 71]
## 2 130.0 140.0 160.0 180.0     0.70      3.8 20.0      [100, 180]
## 3 241.0 275.0 326.9 406.7     1.13      7.4 64.0   [149, 406.74]
## 4 153.0 166.0 181.9 192.0    -0.53      2.9 32.5 [95.02, 191.96]
## 5   0.0   1.0   1.0   1.0     0.74      1.5  1.0          [0, 1]
## 6   0.8   1.6   3.4   4.2     1.26      4.5  1.6        [0, 4.2]
## 7   2.0   2.0   3.0   3.0     0.51      2.4  1.0          [1, 3]
## 8   0.0   1.0   3.0   3.0     1.18      3.2  1.0          [0, 3]
## 9   0.0   2.0   3.0   4.0     1.05      2.8  2.0          [0, 4]
##         range_80
## 1       [42, 66]
## 2     [110, 152]
## 3 [188.8, 308.8]
## 4   [116, 176.6]
## 5         [0, 1]
## 6       [0, 2.8]
## 7         [1, 2]
## 8         [0, 2]
## 9         [0, 3]

Note:

  • plot_num and profiling_num automatically exclude non-numeric
    variables

[🔎 Read more here.]


freq: Getting frequency distributions for categoric variables

library(dplyr)

# Select only two variables for this example
heart_disease_2=heart_disease %>% select(chest_pain, thal)

# Frequency distribution
freq(heart_disease_2)


##   chest_pain frequency percentage cumulative_perc
## 1          4       144       47.5              48
## 2          3        86       28.4              76
## 3          2        50       16.5              92
## 4          1        23        7.6             100


##   thal frequency percentage cumulative_perc
## 1    3       166      54.79              55
## 2    7       117      38.61              93
## 3    6        18       5.94              99
## 4 <NA>         2       0.66             100

## [1] "Variables processed: chest_pain, thal"

Notes:

  • freq only processes factor and character, excluding
    non-categorical variables.
  • It returns the distribution table as a data frame.
  • If input is empty, then it runs for all categorical variables.
  • path_out indicates the path directory; if it has a value, then the plot is exported in jpeg. To save in current directory path must dot: "."
  • na.rm indicates if NA values should be excluded (FALSE by default).

[🔎 Read more here.]


Correlation

correlation_table: Calculates R statistic

Retrieves R metric (or Pearson coefficient) for all numeric variables, skipping the categoric ones.

correlation_table(heart_disease, "has_heart_disease")

##                 Variable has_heart_disease
## 1      has_heart_disease              1.00
## 2 heart_disease_severity              0.83
## 3      num_vessels_flour              0.46
## 4                oldpeak              0.42
## 5                  slope              0.34
## 6                    age              0.23
## 7 resting_blood_pressure              0.15
## 8      serum_cholestoral              0.08
## 9         max_heart_rate             -0.42

Notes:

  • Only numeric variables are analyzed. Target variable must numeric.
  • If target is categorical, then it will be converted to numeric.

[🔎 Read more here.]


var_rank_info: Correlation based on information theory

Calculates correlation based on several information theory metrics between all variables in a data frame and a target variable.

var_rank_info(heart_disease, "has_heart_disease")

##                       var  en    mi      ig      gr
## 1  heart_disease_severity 1.8 0.995 0.99508 0.53907
## 2                    thal 2.0 0.209 0.20946 0.16805
## 3             exer_angina 1.8 0.139 0.13914 0.15264
## 4            exter_angina 1.8 0.139 0.13914 0.15264
## 5              chest_pain 2.5 0.205 0.20502 0.11803
## 6       num_vessels_flour 2.4 0.182 0.18152 0.11577
## 7                   slope 2.2 0.112 0.11242 0.08688
## 8       serum_cholestoral 7.5 0.561 0.56056 0.07956
## 9                  gender 1.8 0.057 0.05725 0.06330
## 10                oldpeak 4.9 0.249 0.24917 0.06036
## 11         max_heart_rate 6.8 0.334 0.33362 0.05407
## 12 resting_blood_pressure 5.6 0.143 0.14255 0.03024
## 13                    age 5.9 0.137 0.13718 0.02705
## 14        resting_electro 2.1 0.024 0.02415 0.02219
## 15    fasting_blood_sugar 1.6 0.000 0.00046 0.00076

Note: It analyzes numerical and categorical variables. It is also used with the numeric discretization method as before, just as discretize_df.

[🔎 Read more here.]


cross_plot: Distribution plot between input and target variable

Retrieves the relative and absolute distribution between an input and target variable. Useful to explain and report if a variable is important or not.

cross_plot(data=heart_disease, input=c("age", "oldpeak"), target="has_heart_disease")

## [1] "Plotting transformed variable 'age' with 'equal_freq', (too many values). Disable with 'auto_binning=FALSE'"


## [1] "Plotting transformed variable 'oldpeak' with 'equal_freq', (too many values). Disable with 'auto_binning=FALSE'"

Notes:

  • auto_binning: TRUE by default, shows the numerical variable as
    categorical.
  • path_out indicates the path directory; if it has a value, then the
    plot is exported in jpeg.
  • input can be numeric or categoric, and target must be a binary
    (two-class) variable.
  • If input is empty, then it runs for all variables.

[🔎 Read more here.]


plotar: Boxplot and density histogram between input and target variables

Useful to explain and report if a variable is important or not.

Boxplot:

plotar(data=heart_disease, input = c("age", "oldpeak"), target="has_heart_disease", plot_type="boxplot")


[🔎 Read more
here.]


Density histograms:

plotar(data=mtcars, input = "gear", target="cyl", plot_type="histdens")

[🔎 Read more here.]

Notes:

  • path_out indicates the path directory; if it has a value, then the plot is exported in jpeg.
  • If input is empty, then it runs for all numeric (skipping the categorical ones).
  • input must be numeric and target must be categoric.
  • target can be multi-class (not only binary).

categ_analysis: Quantitative analysis for binary outcome

Profile a binary target based on a categorical input variable, the representativeness (perc_rows) and the accuracy (perc_target) for each value of the input variable; for example, the rate of flu infection per country.

df_ca=categ_analysis(data = data_country, input = "country", target = "has_flu")

head(df_ca)

##          country mean_target sum_target perc_target q_rows perc_rows
## 1       Malaysia        1.00          1       0.012      1     0.001
## 2         Mexico        0.67          2       0.024      3     0.003
## 3       Portugal        0.20          1       0.012      5     0.005
## 4 United Kingdom        0.18          8       0.096     45     0.049
## 5        Uruguay        0.17         11       0.133     63     0.069
## 6         Israel        0.17          1       0.012      6     0.007

Note:

  • input variable must be categorical.
  • target variable must be binary (two-value).

This function is used to analyze data when we need to reduce variable cardinality in predictive modeling.

[🔎 Read more here.]


Data preparation

Data discretization

discretize_get_bins + discretize_df: Convert numeric variables to categoric

We need two functions: discretize_get_bins, which returns the thresholds for each variable, and then discretize_df, which takes the result from the first function and converts the desired variables. The binning criterion is equal frequency.

Example converting only two variables from a dataset.

# Step 1: Getting the thresholds for the desired variables: "max_heart_rate" and "oldpeak"
d_bins=discretize_get_bins(data=heart_disease, input=c("max_heart_rate", "oldpeak"), n_bins=5)

## [1] "Variables processed: max_heart_rate, oldpeak"

# Step 2: Applying the threshold to get the final processed data frame
heart_disease_discretized=discretize_df(data=heart_disease, data_bins=d_bins, stringsAsFactors=T)

## [1] "Variables processed: max_heart_rate, oldpeak"

The following image illustrates the result. Please note that the
variable name remains the same.

Notes:

  • This two-step procedure is thought to be used in production with new data.
  • Min and max values for each bin will be -Inf and Inf, respectively.
  • A fix in the latest funModeling release (1.6.7) may change output in certain scenarios. Please check the results if you using version 1.6.6. More info about this here.

[🔎 Read more here.]


convert_df_to_categoric: Convert every column in a data frame to character variables

Binning, or discretization criterion for any numerical variable is equal frequency. Factor variables are directly converted to character variables.

iris_char=convert_df_to_categoric(data = iris, n_bins = 5)

## [1] "Variables processed: Sepal.Length, Sepal.Width, Petal.Length, Petal.Width"
## [1] "Variables processed: Sepal.Length, Sepal.Width, Petal.Length, Petal.Width"

# checking first rows
head(iris_char)

##   Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1  [ 5.1, 5.7) [ 3.5, Inf]  [-Inf, 1.6) [-Inf, 0.3)  setosa
## 2  [-Inf, 5.1) [ 2.8, 3.1)  [-Inf, 1.6) [-Inf, 0.3)  setosa
## 3  [-Inf, 5.1) [ 3.2, 3.5)  [-Inf, 1.6) [-Inf, 0.3)  setosa
## 4  [-Inf, 5.1) [ 3.1, 3.2)  [-Inf, 1.6) [-Inf, 0.3)  setosa
## 5  [-Inf, 5.1) [ 3.5, Inf]  [-Inf, 1.6) [-Inf, 0.3)  setosa
## 6  [ 5.1, 5.7) [ 3.5, Inf]  [ 1.6, 4.0) [ 0.3, 1.2)  setosa

equal_freq: Convert numeric variable to categoric

Converts numeric vector into a factor using the equal frequency criterion.

new_age=equal_freq(heart_disease$age, n_bins = 5)

# checking results
Hmisc::describe(new_age)

## new_age 
##        n  missing distinct 
##      303        0        5 
##                                                   
## Value      [29,46) [46,54) [54,59) [59,63) [63,77]
## Frequency       63      64      71      45      60
## Proportion    0.21    0.21    0.23    0.15    0.20

[🔎 Read more here.]

Notes:

  • Unlike discretize_get_bins, this function doesn't insert -Inf and Inf as the min and max value respectively.

range01: Scales variable into the 0 to 1 range

Convert a numeric vector into a scale from 0 to 1 with 0 as the minimum and 1 as the maximum.

age_scaled=range01(heart_disease$oldpeak)

# checking results
summary(age_scaled)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    0.00    0.00    0.13    0.17    0.26    1.00

Outliers data preparation

hampel_outlier and tukey_outlier: Gets outliers threshold

Both functions retrieve a two-value vector that indicates the thresholds for which the values are considered as outliers. The functions tukey_outlier and hampel_outlier are used internally in prep_outliers.

Using Tukey's method:

tukey_outlier(heart_disease$resting_blood_pressure)

## bottom_threshold    top_threshold 
##               60              200

[🔎 Read more here.]


Using Hampel's method:

hampel_outlier(heart_disease$resting_blood_pressure)

## bottom_threshold    top_threshold 
##               86              174

[🔎 Read more
here.]


prep_outliers: Prepare outliers in a data frame

Takes a data frame and returns the same data frame plus the transformations specified in the input parameter. It also works with a single vector.

Example considering two variables as input:

# Get threshold according to Hampel's method
hampel_outlier(heart_disease$max_heart_rate)

## bottom_threshold    top_threshold 
##               86              220

# Apply function to stop outliers at the threshold values
data_prep=prep_outliers(data = heart_disease, input = c('max_heart_rate','resting_blood_pressure'), method = "hampel", type='stop')

Checking the before and after for variable max_heart_rate:

## [1] "Before transformation -> Min: 71; Max: 202"

## [1] "After transformation -> Min: 86.283; Max: 202"

The min value changed from 71 to 86.23, while the max value remained the
same at 202.

Notes:

  • method can be: bottom_top, tukey or hampel.
  • type can be: stop or set_na. If stop all values flagged outliers will be set to the threshold. If set_na, then the flagged values will set to NA.

[🔎 Read more here.]


Predictive model performance

gain_lift: Gain and lift performance curve

After computing the scores or probabilities for the class we want to predict, we pass it to the gain_lift function, which returns a data frame with performance metrics.

# Create machine learning model and get its scores for positive case 
fit_glm=glm(has_heart_disease ~ age + oldpeak, data=heart_disease, family = binomial)
heart_disease$score=predict(fit_glm, newdata=heart_disease, type='response')

# Calculate performance metrics
gain_lift(data=heart_disease, score='score', target='has_heart_disease')


##    Population Gain Lift Score.Point
## 1          10   21  2.1        0.82
## 2          20   36  1.8        0.70
## 3          30   49  1.6        0.57
## 4          40   61  1.5        0.49
## 5          50   69  1.4        0.40
## 6          60   78  1.3        0.33
## 7          70   88  1.2        0.29
## 8          80   92  1.1        0.25
## 9          90   96  1.1        0.20
## 10        100  100  1.0        0.12

[🔎 Read more here.]


  • Github repository (report bugs or improvements).
  • Creator and maintainer: Pablo Casas | twitter | pcasas.biz [at] gmail.com