Introduction to dcmodify

Mark P.J. van der Loo and Edwin de Jonge

Rule-based data processing for data cleaning.

Package version 0.9.0.

Please use citation("dcmodify") to cite the package.

Introduction

The dcmodify package allows users to declare (conditional) data processing steps without hard-coding them in an R script.

Rule-based data processing
Rule-based data processing

The motivating use case is where domain experts need to frequently update (fix) values of a data set depending on data conditions. Such updates require (1) selecting the cases where the conditions apply, and (2) updating the values.

Example

We create a dataset where we asked some imaginary respondents about their average water consumption per day.

> water <- data.frame(
+    name        = c("Ross", "Robert", "Martin", "Brian", "Simon")
+  , consumption = c(110, 105, 0.15, 95, -100) 
+ )
> water
    name consumption
1   Ross      110.00
2 Robert      105.00
3 Martin        0.15
4  Brian       95.00
5  Simon     -100.00

Here, Martin submitted his water consumption in m3 while the others responded in liters. Such a unit of measure error can be detected and fixed. Simon took the engineering approach, interpreted consumption as a sink, and put a negative sign in front of his answer. Again, this is a data error that can be detected and fixed easily.

If such errors occur frequently, it makes sense to store the treatment. In dcmodify this can be solved as follows.

> library(dcmodify)
> # define a rule set (here with one rule)
> rules <- modifier(
+           if ( abs(consumption) <= 1 ) consumption <- 1000*consumption  
+         , if ( consumption < 0 ) consumption <- -1 * consumption )  
> # apply the ruleset to the data
> out <- modify(water, rules)
> out
    name consumption
1   Ross         110
2 Robert         105
3 Martin         150
4  Brian          95
5  Simon         100

In the first step we define a set of conditional data modifying rules of the form:

if (some condition) change somthing

next, using modifier(), these rules are applied record-wise.

What’s the idea

In dcmodify conditional data modifying rules are first class citizens. Modifying rules can be created, deleted, read from or written to file, filtered, selected, and investigated. And of course, applied to data.

In particular, the rules object of the previous example is basically a list of class modifier.

> rules
Object of class modifier with 2 elements:
M1: 
  if (abs(consumption) <= 1) consumption <- 1000 * consumption

M2: 
  if (consumption < 0) consumption <- -1 * consumption

For example, we may select one rule and apply it to our original data set.

> modify(water, rules[2])
    name consumption
1   Ross      110.00
2 Robert      105.00
3 Martin        0.15
4  Brian       95.00
5  Simon      100.00

We can ask which variables are used in the modifying rules (here: only one).

> variables(rules)
[1] "consumption"

Exercises

  1. Load the retailers dataset from the validate package using data("retailers", package="validate").
  2. Create a modifier that (1) sets other.rev to zero if it is missing, (2) replaces negative other.rev with the absolute value.

Import/export of rules and rule metadata

The dcmodify package supports reading/writing rules from free text file, yaml files, or data frames. For example, consider the contents of the file myrules.txt (to try the following code, create such a file yourself).

# myrules.txt

# unit of measure error
if (abs(consumption) <= 1){
  consumption <- 1000*consumption
}

# sign error
if (consumption < 0 ){
  consumption <- -1 * consumption
}

Reading the rules is done with the .file argument.

> rules <- modifier(.file="myrules.txt")
> rules
Object of class modifier with 2 elements:
M1: 
  if (abs(consumption) <= 1) {
      consumption <- 1000 * consumption
  }

M2: 
  if (consumption < 0) {
      consumption <- -1 * consumption
  }

A second way to store rules is in yaml format. This allows one to add metadata to rules, including a name, label or description. To demonstrate this, we will write the rules to yaml file and print the file contents.

> fn <- tempfile()
> # export rules to yaml format
> export_yaml(rules,file=fn)
> # print file contents
> readLines(fn) |> paste(collapse="\n") |> cat()
rules:
- expr:
  - if (abs(consumption) <= 1) {
  - '    consumption <- 1000 * consumption'
  - '}'
  name: M1
  label: ''
  description: ''
  created: 2024-03-27 16:07:30.73769
  origin: myrules.txt
  meta: []
- expr:
  - if (consumption < 0) {
  - '    consumption <- -1 * consumption'
  - '}'
  name: M2
  label: ''
  description: ''
  created: 2024-03-27 16:07:30.73769
  origin: myrules.txt
  meta: []

Finally, it is possible to read rules from (and export to) data frame. A rule data frame must at least contain a character column named rule; all other columns are considered metadata.

> d <- data.frame(
+     name = c("U1","S1")
+   , label = c("Unit error", "sign error")
+ )
> d$rule <- c(
+    "if(abs(consumption)<=1) consuption <- 1000 * consumption"
+   ,"if(consumption < 0) consumption <- -1 * consumption"
+ )
> d
  name      label                                                     rule
1   U1 Unit error if(abs(consumption)<=1) consuption <- 1000 * consumption
2   S1 sign error      if(consumption < 0) consumption <- -1 * consumption

Reading from data frame is done with the .data argument.

> myrules <- modifier(.data=d)
> myrules
Object of class modifier with 2 elements:
U1: Unit error
  if (abs(consumption) <= 1) consuption <- 1000 * consumption

S1: sign error
  if (consumption < 0) consumption <- -1 * consumption

Exercise

Using the retailers dataset of the previous exercises, define a file with rules that

See ?retailers for the meaning of the variables.

Using reference data

It is possible to use data not in the treated dataset in your data modification rules. This so-called reference data needs to be added in the call to modify().

> dat <- data.frame(x=seq_len(nrow(women)))
> m   <- modifier(if (x > 2) x <- ref$height/2)
> out <- modify(dat, m, ref=women)
> head(out)
     x
1  1.0
2  2.0
3 30.0
4 30.5
5 31.0
6 31.5

Note that in this form, it is necessary to use ref$ to refer to the women dataset in the context of modifying rules. This can be customized by passing the reference data as a named list or environment.

> m   <- modifier(if (x > 2) x <- women$height/2)
> out <- modify(dat, m, ref=list(women=women))
> head(out,3)
   x
1  1
2  2
3 30

Or, equivalently

> e <- new.env()
> e$women <- women
> m   <- modifier(if (x > 2) x <- women$height/2)
> out <- modify(dat, m, ref=e)
> head(out,3)
   x
1  1
2  2
3 30

Supported rules

The package supports rules of the form

if (some condition holds){
  change some existing values
} 

where the else clause is optional, and the rules are executed record-by-record. There may be multiple expressions in each {} block, and it is also allowed to have nested if-else statements.

Tracking changes

dcmodify allows rule-by-rule change tracking via integration with the lumberjack package[1]. There are many ways of following what happens to a data file, and we refere to [1] for an overview of the possibilities of the lumberjack package. Here we demonstrate how to use the cellwise logger, which logs all changes cell-by-cell.

> library(lumberjack)
> # create a logger (see ?cellwise)
> lgr <- cellwise$new(key="name")
> # create rules
> 
> rules <- modifier(
+           if ( abs(consumption) <= 1 ) consumption <- 1000*consumption  
+         , if ( consumption < 0 ) consumption <- -1 * consumption )  
> # apply rules, and pass logger object to modify()
> out <- modify(water, rules, logger=lgr)
> # check what happened, by dumping the log and reading in 
> # the csv.
> logfile <- tempfile()
> lgr$dump(file=logfile)
Dumped a log at /tmp/RtmpSEKmUk/file55745013dd2f
 
> read.csv(logfile)
  step                    time         srcref                        expression
1    1 2024-03-27 16:07:30 CET <modifier>#1-1 consumption <- 1000 * consumption
2    2 2024-03-27 16:07:30 CET <modifier>#2-2   consumption <- -1 * consumption
     key    variable     old new
1 Martin consumption    0.15 150
2  Simon consumption -100.00 100

References

[1] van der Loo MPJ (2021). Monitoring Data in R with the lumberjack Package. Journal of Statistical Software, 98(1), 1–13. doi:10.18637/jss.v098.i01.