Data quality diagnosis

knitr::opts_chunk$set(collapse = TRUE, comment = "", out.width = "600px", dpi = 70)
options(tibble.print_min = 4L, tibble.print_max = 4L)

library(dlookr)
library(dplyr)
library(ggplot2)

Preface

After you have acquired the data, you should do the following:

The dlookr package makes these steps fast and easy:

This document introduces Data Quality Diagnosis methods provided by the dlookr package. You will learn how to diagnose the quality of tbl_df data that inherits from data.frame and data.frame with functions provided by dlookr.

dlookr increases synergy with dplyr. Particularly in data exploration and data wrangle, it increases the efficiency of the tidyverse package group.

Supported data structures

Data diagnosis supports the following data structures.

Data: nycflights13

To illustrate basic use of the dlookr package, use the flights data from the nycflights13 package. The flights data frame is data about departure and arrival on all flights departing from NYC in 2013.

library(nycflights13)
dim(flights)
flights

Data diagnosis

dlookr aims to diagnose the data and to select variables that can not be used for data analysis or to find the variables that need to be calibrated.:

General diagnosis of all variables with diagnose()

diagnose() allows to diagnosis a variables in a data frame. Like function of dplyr, the first argument is the tibble (or data frame). The second and subsequent arguments refer to variables within that data frame.

The variables of the tbl_df object returned by diagnose () are as follows.

For example, we can diagnose all variables in flights:

diagnose(flights)

year can be considered not to be used in the analysis model since unique_count is 1. However, you do not have to remove it if you configure date as a combination of year, month, and day.

For example, we can diagnose only a few selected variables:

# Select columns by name
diagnose(flights, year, month, day)
# Select all columns between year and day (include)
diagnose(flights, year:day)
# Select all columns except those from year to day (exclude)
diagnose(flights, -(year:day))

By using with dplyr, variables including missing values can be sorted by the weight of missing values.:

flights %>%
  diagnose() %>%
  select(-unique_count, -unique_rate) %>% 
  filter(missing_count > 0) %>% 
  arrange(desc(missing_count))

Diagnosis of numeric variables with diagnose_numeric()

diagnose_numeric() diagnoses numeric(continuous and discrete) variables in a data frame. Usage is the same as diagnose() but returns more diagnostic information. However, if you specify a non-numeric variable in the second and subsequent argument list, the variable is automatically ignored.

The variables of the tbl_df object returned by diagnose_numeric() are as follows.

The summary() function summarizes the distribution of individual variables in the data frame and outputs it to the console. The summary values of numeric variables are min, Q1, mean, median, Q3 and max, which help to understand the distribution of data.

However, the result displayed on the console has the disadvantage that the analyst has to look at it with the eyes. However, when the summary information is returned in a data frame structure such as tbl_df, the scope of utilization is expanded. diagnose_numeric() supports this.

zero, minus, and outlier are useful measures to diagnose data integrity. For example, numerical data in some cases cannot have zero or negative numbers. A numeric variable called employee salary cannot have negative numbers or zeros. Therefore, this variable should be checked for the inclusion of zero or negative numbers in the data diagnosis process.

diagnose_numeric() can diagnose all numeric variables of flights as follows.:

diagnose_numeric(flights)

If a numeric variable can not logically have a negative or zero value, it can be used with filter() to easily find a variable that does not logically match:

diagnose_numeric(flights) %>% 
  filter(minus > 0 | zero > 0) 

Diagnosis of categorical variables with diagnose_category()

diagnose_category() diagnoses the categorical(factor, ordered, character) variables of a data frame. The usage is similar to diagnose() but returns more diagnostic information. If you specify a non-categorical variable in the second and subsequent argument list, the variable is automatically ignored.

The top argument specifies the number of levels to return for each variable. The default is 10, which returns the top 10 level. Of course, if the number of levels is less than 10, all levels are returned.

The variables of the tbl_df object returned by diagnose_category() are as follows.

`diagnose_category() can diagnose all categorical variables of flights as follows.:

diagnose_category(flights)

In collaboration with filter() in the dplyr package, we can see that the tailnum variable is ranked in top 1 with 2,512 missing values in the case where the missing value is included in the top 10:

diagnose_category(flights) %>% 
  filter(is.na(levels))

The following example returns a list where the level's relative percentage is 0.01% or less. Note that the value of the top argument is set to a large value such as 500. If the default value of 10 was used, values below 0.01% would not be included in the list:

flights %>%
  diagnose_category(top = 500)  %>%
  filter(ratio <= 0.01)

In the analytics model, you can also consider removing levels where the relative frequency is very small in the observations or, if possible, combining them together.

Diagnosing outliers with diagnose_outlier()

diagnose_outlier() diagnoses the outliers of the numeric (continuous and discrete) variables of the data frame. The usage is the same as diagnose().

The variables of the tbl_df object returned by diagnose_outlier() are as follows.

diagnose_outlier() can diagnose outliers of all numerical variables on flights as follows:

diagnose_outlier(flights)

Numeric variables that contained outliers are easily found with filter().:

diagnose_outlier(flights) %>% 
  filter(outliers_cnt > 0) 

The following example finds a numeric variable with an outlier ratio of 5% or more, and then returns the result of dividing mean of outliers by overall mean in descending order:

diagnose_outlier(flights) %>% 
  filter(outliers_ratio > 5) %>% 
  mutate(rate = outliers_mean / with_mean) %>% 
  arrange(desc(rate)) %>% 
  select(-outliers_cnt)

In cases where the mean of the outliers is large relative to the overall average, it may be desirable to impute or remove the outliers.

Visualization of outliers using plot_outlier()

plot_outlier() visualizes outliers of numerical variables(continuous and discrete) of data.frame. Usage is the same diagnose().

The plot derived from the numerical data diagnosis is as follows.

plot_outlier() can visualize an outliers in the arr_delay variable of flights as follows:

flights %>%
  plot_outlier(arr_delay) 

The following example uses diagnose_outlier(), plot_outlier(), and dplyr packages to visualize all numerical variables with an outlier ratio of 5% or higher.

flights %>%
  plot_outlier(diagnose_outlier(flights) %>% 
                 filter(outliers_ratio >= 5) %>% 
                 select(variables) %>% 
                 unlist())

Analysts should look at the results of the visualization to decide whether to remove or replace outliers. In some cases, you should consider removing variables with outliers from the data analysis model.

Looking at the results of the visualization, arr_delay shows that the observed values without outliers are similar to the normal distribution. In the case of a linear model, we might consider removing or imputing outliers.

Visualization for missing values

It is important to look at the missing values of individual variables, but it is also important to look at the relationship between the variables including the missing values.

dlookr provides a visualization tool that looks at the relationship of variables including missing values.

visualize pareto chart using plot_na_pareto()

plot_na_pareto() draws a pareto chart by collecting variables including missing values.

mice::boys %>% 
  plot_na_pareto(col = "blue")

The default value of the only_na argument is FALSE, which includes variables that do not contain missing values, but if this value is set to TRUE, only variables containing missing values are visualized. The variable age was excluded from this plot.

mice::boys %>% 
  plot_na_pareto(only_na = TRUE, main = "Pareto Chart for mice::boys")

The rating of the variable is expressed as a proportion of missing values. It is calculated as the ratio of missing values. If it is [0, 0.05), it is Good, if it is [0.05, 0.4) it is OK, if it is [0.4, 0.8) it is Bad, and if it is [0.8, 1.0] it is Remove. You can override this grade using the grade argument as follows:

mice::boys %>% 
  plot_na_pareto(grade = list(High = 0.1, Middle = 0.6, Low = 1), relative = TRUE)

If the plot argument is set to FALSE, information about missing values is returned instead of plotting.

plot_na_pareto(mice::boys, only_na = TRUE, plot = FALSE)

visualize combination chart using plot_na_hclust()

It is important to look at the relationship between variables, including missing values. plot_na_hclust() visualizes the relationship of variables that contain missing values. This function rearranges the positions of variables using hierarchical clustering. Then, the expression of the missing value is visualized by grouping similar variables.

mice::boys %>% 
  plot_na_hclust(main = "Distribution of missing value")

visualize combination chart using plot_na_intersect()

plot_na_intersect() visualize the combinations of missing value across cases.

The visualization consists of four parts. The bottom left, which is the most basic, visualizes the case of cross(intersection)-combination. The x-axis is the variable including the missing value, and the y-axis represents the case of a combination of variables. And on the marginal of the two axes, the frequency of the case is expressed as a bar graph. Finally, the visualization at the top right expresses the number of variables including missing values in the data set, and the number of observations including missing values and complete cases .

This example visualize the combination variables that is include missing value.

mice::boys %>% 
  plot_na_intersect()

If the n_vars argument is used, only the top n variables containing many missing values are visualized.

mice::boys %>%
  plot_na_intersect(n_vars = 5)

If you use the n_intersacts argument, only the top n numbers of variable combinations(intersection) including missing values are visualized. and If you want to visualize the combination variables that is include missing value and complete case. You just add only_na = FALSE.

mice::boys %>%
  plot_na_intersect(only_na = FALSE, n_intersacts = 7)

Automated report

dlookr provides two automated data diagnostic reports:

Create a diagnostic report using diagnose_web_report()

diagnose_web_report() create dynamic report for object inherited from data.frame(tbl_df, tbl, etc) or data.frame.

Contents of dynamic web report

The contents of the report are as follows.:

Some arguments for dynamic web report

diagnose_web_report() generates various reports with the following arguments.

The following script creates a quality diagnosis report for the tbl_df class object, flights.

flights %>%
  diagnose_web_report(subtitle = "flights", output_dir = "./", 
                      output_file = "Diagn.html", theme = "blue")

Screenshot of dynamic report

knitr::include_graphics('img/diag_web_title.jpg')
knitr::include_graphics('img/diag_web_content.jpg')

Create a diagnostic report using diagnose_paged_report()

diagnose_paged_report() create static report for object inherited from data.frame(tbl_df, tbl, etc) or data.frame.

Contents of static paged report

The contents of the report are as follows.:

Some arguments for static paged report

diagnose_paged_report() generates various reports with the following arguments.

The following script creates a quality diagnosis report for the tbl_df class object, flights.

flights %>%
  diagnose_paged_report(subtitle = "flights", output_dir = "./",
                        output_file = "Diagn.pdf", theme = "blue")

Screenshot of static report

knitr::include_graphics('img/diag_paged_cover.jpg')
knitr::include_graphics('img/diag_paged_content.jpg')

Diagnosing tables in DBMS

The DBMS table diagnostic function supports In-database mode that performs SQL operations on the DBMS side. If the size of the data is large, using In-database mode is faster.

It is difficult to obtain anomaly or to implement the sampling-based algorithm in SQL of DBMS. So some functions do not yet support In-database mode. In this case, it is performed in In-memory mode in which table data is brought to R side and calculated. In this case, if the data size is large, the execution speed may be slow. It supports the collect_size argument, which allows you to import the specified number of samples of data into R.

Preparing table data

Copy the carseats data frame to the SQLite DBMS and create it as a table named TB_CARSEATS. Mysql/MariaDB, PostgreSQL, Oracle DBMS, other DBMS are also available for your environment.

if (!require(DBI)) install.packages('DBI')
if (!require(RSQLite)) install.packages('RSQLite')
if (!require(dplyr)) install.packages('dplyr')
if (!require(dbplyr)) install.packages('dbplyr')

library(dplyr)

carseats <- ISLR::Carseats
carseats[sample(seq(NROW(carseats)), 20), "Income"] <- NA
carseats[sample(seq(NROW(carseats)), 5), "Urban"] <- NA

# connect DBMS
con_sqlite <- DBI::dbConnect(RSQLite::SQLite(), ":memory:")

# copy carseats to the DBMS with a table named TB_CARSEATS
copy_to(con_sqlite, carseats, name = "TB_CARSEATS", overwrite = TRUE)

Diagnose data quality of variables in the DBMS

Use dplyr::tbl() to create a tbl_dbi object, then use it as a data frame object. That is, the data argument of all diagnose function is specified as tbl_dbi object instead of data frame object.

# Diagnosis of all columns
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose()

# Positions values select columns, and In-memory mode
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose(1, 3, 8, in_database = FALSE)

# Positions values select columns, and In-memory mode and collect size is 200
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose(-8, -9, -10, in_database = FALSE, collect_size = 200)

Diagnose data quality of categorical variables in the DBMS

# Positions values select variables, and In-memory mode and collect size is 200
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_category(7, in_database = FALSE, collect_size = 200) 

# Positions values select variables
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_category(-7)

Diagnose data quality of numerical variables in the DBMS

# Diagnosis of all numerical variables
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_numeric()

# Positive values select variables, and In-memory mode and collect size is 200
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_numeric(Sales, Income, collect_size = 200)

Diagnose outlier of numerical variables in the DBMS

con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_outlier()  %>%
  filter(outliers_ratio > 1)

Plot outlier information of numerical data diagnosis in the DBMS

# Visualization of numerical variables with a ratio of
# outliers greater than 1%
# the result is same as a data.frame, but not display here. reference above in document.
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  plot_outlier(con_sqlite %>% 
                 tbl("TB_CARSEATS") %>% 
                 diagnose_outlier() %>%
                 filter(outliers_ratio > 1) %>%
                 select(variables) %>%
                 pull())

Reporting the information of data diagnosis for table of thr DBMS

The following shows several examples of creating an data diagnosis report for a DBMS table.

Using the collect_size argument, you can perform data diagnosis with the corresponding number of sample data. If the number of data is very large, use collect_size.

# create web report file. 
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_web_report()

# create pdf file. file name is Diagn.pdf, and collect size is 350
con_sqlite %>% 
  tbl("TB_CARSEATS") %>% 
  diagnose_paged_report(collect_size = 350, output_file = "Diagn.pdf")


Try the dlookr package in your browser

Any scripts or data that you put into this service are public.

dlookr documentation built on July 30, 2021, 9:07 a.m.