Permalink
Fetching contributors…
Cannot retrieve contributors at this time
752 lines (570 sloc) 26.3 KB
---
title: "Assertive R Programming with assertr"
author: "Tony Fischetti"
date: "`r Sys.Date()`"
output: rmarkdown::html_vignette
vignette: >
%\VignetteIndexEntry{Assertive R Programming with assertr}
%\VignetteEngine{knitr::rmarkdown}
\usepackage[utf8]{inputenc}
---
In data analysis workflows that depend on un-sanitized data sets from external
sources, it’s very common that errors in data bring an analysis to a
screeching halt. Oftentimes, these errors occur late in the analysis and
provide no clear indication of which datum caused the error.
On occasion, the error resulting from bad data won’t even appear to be a
data error at all. Still worse, errors in data will pass through analysis
without error, remain undetected, and produce inaccurate results.
The solution to the problem is to provide as much information as you can about
how you expect the data to look up front so that any deviation from this
expectation can be dealt with immediately. This is what the `assertr` package
tries to make dead simple.
Essentially, `assertr` provides a suite of functions designed to verify
assumptions about data early in an analysis pipeline. This package needn't
be used with the `magrittr`/`dplyr` piping mechanism but the examples in this
vignette will use them to enhance clarity.
### concrete data errors
Let’s say, for example, that the R’s built-in car dataset, `mtcars`, was not
built-in but rather procured from an external source that was known for making
errors in data entry or coding.
In particular, the mtcars dataset looks like this:
```{r}
head(mtcars)
```
But let's pretend that the data we got accidentally negated the 5th mpg value:
```{r}
our.data <- mtcars
our.data$mpg[5] <- our.data$mpg[5] * -1
our.data[4:6,]
```
Whoops!
If we wanted to find the average miles per gallon for each number of engine
cylinders, we might do so like this:
```{r message=FALSE}
library(dplyr)
our.data %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
This indicates that the average miles per gallon for a 8 cylinder car is a lowly
12.43. However, in the correct dataset it's really just over 15. Data errors
like that are extremely easy to miss because it doesn't cause an error, and the
results look reasonable.
### enter assertr
To combat this, we might want to use assertr's `verify` function to make sure
that `mpg` is a positive number:
```{r error=TRUE, purl = FALSE, include=TRUE}
library(assertr)
our.data %>%
verify(mpg >= 0) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
If we had done this, we would have caught this data error.
The `verify` function takes a data frame (its first argument is provided by
the `%>%` operator), and a logical (boolean) expression. Then, `verify`
evaluates that expression using the scope of the provided data frame. If any
of the logical values of the expression's result are `FALSE`, `verify` will
raise an error that terminates any further processing of the pipeline.
We could have also written this assertion using `assertr`'s `assert` function...
```{r error=TRUE, purl = FALSE}
our.data %>%
assert(within_bounds(0,Inf), mpg) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
The `assert` function takes a data frame, a predicate function, and an arbitrary
number of columns to apply the predicate function to. The predicate function
(a function that returns a logical/boolean value) is then applied to every
element of the columns selected, and will raise an error when if it finds
violations.
Internally, the `assert` function uses `dplyr`'s `select` function to extract
the columns to test the predicate function on. This allows for complex
assertions. Let's say we wanted to make sure that all values in the dataset
are *greater* than zero (except `mpg`):
```{r error=TRUE, purl = FALSE}
library(assertr)
our.data %>%
assert(within_bounds(0,Inf, include.lower=FALSE), -mpg) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
### verify vs. assert
The first noticable difference between `verify` and `assert` is that `verify`
takes an expression, and `assert` takes a predicate and columns to apply it to.
This might make the `verify` function look more elegant--but there's an
important drawback. `verify` has to evaluate the entire expression first, and
*then* check if there were any violations. Because of this, `verify` can't
tell you the offending datum.
One important drawback to `assert`, and a consequence of its application of
the predicate to *columns*, is that `assert` can't confirm assertions about
the data structure *itself*. For example, let's say we were reading a dataset
from disk that we know has more than 100 observations; we could write a check
of that assumption like this:
```{r eval=FALSE, purl = FALSE}
dat <- read.csv("a-data-file.csv")
dat %>%
verify(nrow(.) > 100) %>%
....
```
This is a powerful advantage over `assert`... but `assert` has one more
advantage of its own that we've heretofore ignored.
### assertr's predicates
`assertr`'s predicates, both built-in and custom, make `assert` very powerful.
The predicates that are built in to `assertr` are
- `not_na` - that checks if an element is not NA
- `within_bounds` - that returns a predicate function that checks if a numeric
value falls within the bounds supplied, and
- `in_set` - that returns a predicate function that checks if an element is
a member of the set supplied.
- `is_uniq` - that checks to see if each element appears only once
We've already seen `within_bounds` in action... let's use the `in_set` function
to make sure that there are only 0s and 1s (automatic and manual, respectively)
values in the `am` column...
```{r, eval=FALSE, purl = FALSE}
our.data %>%
assert(in_set(0,1), am) %>%
...
```
If we were reading a dataset that contained a column representing boroughs of
New York City (named `BORO`), we can verify that there are no mis-spelled
or otherwise unexpected boroughs like so...
```{r, eval=FALSE, purl = FALSE}
boroughs <- c("Bronx", "Manhattan", "Queens", "Brooklyn", "Staten Island")
read.csv("a-dataset.csv") %>%
assert(in_set(boroughs), BORO) %>%
...
```
Rad!
### custom predicates
A convenient feature of `assertr` is that it makes the construction of custom
predicate functions easy.
In order to make a custom predicate, you only have to specify cases where the
predicate should return FALSE. Let's say that a dataset has an ID column
(named `ID`) that we want to check is not an empty string. We can create a
predicate like this:
```{r}
not.empty.p <- function(x) if(x=="") return(FALSE)
```
and apply it like this:
```{r, eval=FALSE, purl = FALSE}
read.csv("another-dataset.csv") %>%
assert(not.empty.p, ID) %>%
...
```
Let's say that the ID column is always a 7-digit number. We can confirm that
all the IDs are 7-digits by defining the following predicate:
```{r}
seven.digit.p <- function(x) nchar(x)==7
```
A powerful consequence of this easy creation of predicates is that the
`assert` function lends itself to use with lambda predicates (unnamed
predicates that are only used once). The check above might be better written as
```{r, eval=FALSE, purl = FALSE}
read.csv("another-dataset.csv") %>%
assert(function(x) nchar(x)==7, ID) %>%
...
```
Neat-o!
### enter `insist` and predicate 'generators'
Very often, there is a need to dynamically determine the predicate function
to be used based on the vector being checked.
For example, to check to see if every element of a vector is within _n_
standard deviations of the mean, you need to create a `within_bounds`
predicate _after_ dynamically determining the bounds by reading and computing
on the vector itself.
To this end, the `assert` function is no good; it just applies a raw predicate
to a vector. We need a function like `assert` that will apply predicate
_generators_ to vectors, return predicates, and _then_ perform `assert`-like
functionality by checking each element of the vectors with its respective custom
predicate. This is precisely what `insist` does.
This is all much simpler than it may sound. Hopefully, the examples will clear
up any confusion.
The primary use case for `insist` is in conjunction with the `within_n_sds` or
`within_n_mads` predicate generator.
Suppose we wanted to check that every `mpg` value in the `mtcars` data set was
within 3 standard deviations of the mean before finding the average miles
per gallon for each number of engine cylinders. We could write something
like this:
```{r purl = FALSE}
mtcars %>%
insist(within_n_sds(3), mpg) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
Notice what happens when we drop that z-score to 2 standard deviations
from the mean
```{r error=TRUE, purl = FALSE}
mtcars %>%
insist(within_n_sds(2), mpg) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
Execution of the pipeline was halted. But now we know exactly which data point
violated the predicate that `within_n_sds(2)(mtcars$mpg)`
returned.
Now that's an efficient car!
After the predicate generator, `insist` takes an arbitrary number of columns
just like `assert` using the syntax of `dplyr`'s `select` function. If you
wanted to check that everything in mtcars is within 10 standard deviations
of the mean (of each column vector), you can do so like this:
```{r purl = FALSE}
mtcars %>%
insist(within_n_sds(10), mpg:carb) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
Aces!
I chose to use `within_n_sds` in this example because people are familiar
z-scores. However, for most practical purposes, the related predicate generator
`within_n_mads` is more useful.
The problem with `within_n_sds` is the mean and standard deviation are so
heavily influenced by outliers, their very presence will compromise attempts
to identify them using these statistics. In contrast with `within_n_sds`,
`within_n_mads` uses the robust statistics, median and median absolute
deviation, to identify potentially erroneous data points.
For example, the vector `<7.4, 7.1, 7.2, 72.1>` almost certainly has an erroneous
data point, but `within_n_sds(2)` will fail to detect it.
```{r purl = FALSE}
example.vector <- c(7.4, 7.1, 7.2, 72.1)
within_n_sds(2)(example.vector)(example.vector)
```
whereas `within_n_mads` will detect it at even lower levels of power....
```{r purl = FALSE}
example.vector <- c(7.4, 7.1, 7.2, 72.1)
within_n_mads(2)(example.vector)(example.vector)
within_n_mads(1)(example.vector)(example.vector)
```
Tubular!
### row-wise assertions and row reduction functions
As cool as it's been so far, this still isn't enough to constitute a complete
grammar of data integrity checking. To see why, check out the following
small example data set:
```{r perl=FALSE}
example.data <- data.frame(x=c(8, 9, 6, 5, 9, 5, 6, 7,
8, 9, 6, 5, 5, 6, 7),
y=c(82, 91, 61, 49, 40, 49, 57,
74, 78, 90, 61, 49, 51, 62, 68))
(example.data)
```
Can you spot the brazen outlier? You're certainly not going to find it by
checking the distribution of each *column*! All elements from both columns are
within 2 standard deviations of their respective means.
Unless you have a *really* good eye, the only way you're going to catch this
mistake is by plotting the data set.
```{r purl=FALSE}
plot(example.data$x, example.data$y, xlab="", ylab="")
```
Ok, so all the `y`s are roughly 10 times the `x`s except the outlying data
point.
The problem having to plot data sets to catch anomalies is that it is *really*
hard to visualize 4-dimensions at once, and it is near impossible with
high-dimensional data.
There's no way of catching this anomaly by looking at each individual
column separately; the only way to catch it is to view each row as a complete
observation and compare it to the rest.
To this end, `assertr` provides three functions that take a data frame, and
reduce each row into a single value. We'll call them *row reduction functions*.
The first one we'll look at is called `maha_dist`. It computes the average
mahalanobis distance (kind of like multivariate z-scoring for outlier
detection) of each row from the whole data set. The big idea is that in the
resultant vector, big/distant values are potential anomalous entries. Let's
look at the distribution of mahalanobis distances for this data set...
```{r purl=FALSE}
maha_dist(example.data)
maha_dist(example.data) %>% hist(main="", xlab="")
```
There's no question here as to whether there's an anomalous entry! But how do
you check for this sort of thing using `assertr` constructs?
Well, `maha_dist` will typically be used with the `insist_rows` function.
`insist_rows` takes a data frame, a row reduction function, a
predicate-generating function, and an arbitrary number of columns to apply
the predicate function to. The row reduction function (`maha_dist` in this case)
is applied to the data frame, and returns a value for each row. The
predicate-generating function is then applied to the vector returned from
the row reduction function and the resultant predicate is applied to each
element of that vector. It will raise an error if it finds any violations.
As always, this undoubtedly sounds far more confusing than it really is. Here's
an example of it in use
```{r purl=FALSE, error=TRUE}
example.data %>%
insist_rows(maha_dist, within_n_mads(3), dplyr::everything())
```
Check that out! To be clear, this function is running the supplied data frame
through the `maha_dist` function which returns a value for each row
corresponding to its mahalanobis distance. (The whole data frame is used because
we used the `everything()` selection function from the `dplyr` package.)
Then, `within_n_mads(3)` computes on that vector and returns a bounds
checking predicate. The bounds checking predicate checks to see that all
mahalanobis distances are within 3 median absolute deviations
of each other. They are not, and the pipeline errors out. Note that the
`data.frame` of errors that is returned by error report contains the
verb used (`insist_rows`), the row reduction function, the predicate,
the column (or columns), the index of the failure and the offendind datum.
This is probably the most powerful construct in `assertr`--it can find a whole
lot of nasty errors that would be very difficult to check for by hand.
Part of what makes it so powerful is how flexible `maha_dist` is. We only used
it, so far, on a data frame of numerics, but it can handle all sorts of data
frames. To really see it shine, let's use it on the iris data set, that contains
a categorical variable in its right-most column...
```{r purl=FALSE}
head(iris)
iris %>% maha_dist %>% hist(main="", xlab="")
```
Looks ok, but what happens when we accidentally enter a row as a different
species...
```{r purl=FALSE}
mistake <- iris
(mistake[149,5])
mistake[149,5] <- "setosa"
mistake %>% maha_dist %>% hist(main="", xlab="")
mistake %>% maha_dist %>% which.max
```
Look at that! This mistake can easily be picked up by any reasonable bounds
checker...
```{r purl=FALSE, error=TRUE}
mistake %>% insist_rows(maha_dist, within_n_mads(7), dplyr::everything())
```
`insist` and `insist_rows` are both similar in that they both take predicate
generators and not actual predicates. What makes `insist_rows` different is
its usage of a row-reduce data frame.
`assert` has a row-oriented counterpart, too; it's called `assert_rows`.
`insist` is to `assert` as `insist_rows` is to `assert_rows`.
`assert_rows` works the same as `insist_rows`, except that instead of using
a predicate generator on the row-reduced data frame, it uses a regular-old
predicate.
For an example of a `assert_rows` use case, let's say that we got a data set
(`another-dataset.csv`) from the web and we don't want to continue processing
the data set if any row contains more than two missing values (NAs). You
can use the row reduction function `num_row_NAs` to reduce all the rows into
the number of NAs they contain. Then, a simple bounds checker will suffice for
ensuring that no element is higher than 2...
```{r, eval=FALSE, purl = FALSE}
read.csv("another-dataset.csv") %>%
assert_rows(num_row_NAs, within_bounds(0,2), dplyr::everything()) %>%
...
```
`assert_rows` can be used for anomaly detection as well. A future version of
`assertr` may contain a cosine distance row reduction function. Since all
cosine distances are constained from -1 to 1, it is easy to use a non-dynamic
predicate to disallow certain values.
### success and error functions
The behavior of functions like `assert`, `assert_rows`,
`insist`, `insist_rows`, `verify` when the assertion
passes or fails is configurable via the `success_fun`
and `error_fun` parameters, respectively.
The `success_fun` parameter takes a function that takes
the data passed to the assertion function as a parameter. You can
write your own success handler function, but there are two
provided by this package:
* `success_continue` - just returns the data that was passed into the
assertion function (this is default)
* `success_logical` - returns TRUE
The `error_fun` parameter takes a function that takes
the data passed to the assertion function as a parameter. You can
write your own error handler function, but there are a few
provided by this package:
* `error_stop` - Prints a summary of the errors and
halts execution (default)
* `error_report` - Prints *all* the information available
about the errors and halts execution.
* `error_append` - Attaches the errors to a special
attribute of `data` and returns the data. This is chiefly
to allow assertr errors to be accumulated in a pipeline so that
all assertions can have a chance to be checked and so that all
the errors can be displayed at the end of the chain.
* `error_logical` - returns FALSE
* `just_warn` - Prints a summary of the errors but does
not halt execution, it just issues a warning.
* `warn_report` - Prints all the information available
about the errors but does not halt execution, it just issues a warning.
### combining chains of assertions
Let's say that as part of an automated pipeline that grabs mtcars from an
untrusted source and finds the average miles per gallon for each number of
engine cylinders, we want to perform the following checks...
- that it has the columns "mpg", "vs", and "am"
- that the dataset contains more than 10 observations
- that the column for 'miles per gallon' (mpg) is a positive number
- that the column for 'miles per gallon' (mpg) does not contain a
datum that is outside 4 standard deviations from its mean, and
- that the am and vs columns (automatic/manual and v/straight engine,
respectively) contain 0s and 1s only
This could be written thusly:
```{r purl = FALSE}
mtcars %>%
verify(has_all_names("mpg", "vs", "am")) %>%
verify(nrow(mtcars) > 10) %>%
verify(mpg > 0) %>%
insist(within_n_sds(4), mpg) %>%
assert(in_set(0,1), am, vs) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
In an assertr chain with default options, `assert`, `assert_rows`,
`insist`, `insist_rows`, and `verify` will stop at the
first assertion that yields an error and not go on to process the
assertions further down in the chain. For some needs, this is sensible
behavior. There are times, however, when we might like to get a report
of all assertion violations. For example, one might want to write an R
program to download some dataset from the internet and get a detailed
report of all deviations from expectation.
The best thing to do for this use case, is to use the `chain_start`,
and `chain_end` functions at the beginning and end of a chain of
assertr assertions. When `chain_start` gets called with data, the
data gets a special tag that tells the assertr assertions that follow
to override their `success_fun` and `error_fun` values and
replace them with `success_continue` (which passes the data along
if the test passes) and `error_append` (which we've just discussed).
After all relevant verifications, `chain_end` will receive the
data (possibly with accumulated error messages attached) and, by default,
print a report of all the errors that have been found since the start of
the chain.
Let's see it in action!
```{r purl = FALSE}
mtcars %>%
chain_start %>%
verify(nrow(mtcars) > 10) %>%
verify(mpg > 0) %>%
insist(within_n_sds(4), mpg) %>%
assert(in_set(0,1), am, vs) %>%
chain_end %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
Now *all* assertions will be checked and reported.
Tip: we can make this whole thing look a lot better by abstracting out
all the assertions:
```{r eval = FALSE, purl = FALSE}
check_me <- . %>%
chain_start %>%
verify(nrow(mtcars) > 10) %>%
verify(mpg > 0) %>%
insist(within_n_sds(4), mpg) %>%
assert(in_set(0,1), am, vs)
chain_end
mtcars %>%
check_me %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
Awesome! Now we can add an arbitrary number of assertions, as the need arises,
without touching the real logic.
### advanced: send email reports using custom error functions
One particularly cool application of `assertr` is to use it as a data integrity
checker for frequently updated data sources. A script can download new data as
it becomes available, and then run `assertr` checks on it. This makes `assertr`
into a sort of "continuous integration" tool (but for data,
not code.)
In an unsupervised "continuous integration" environment, you need a way to
discover that the assertions failed. In CI-as-a-service in the software world,
failed automated checks often send an email of reporting the maintainer of a
botched build; why not bring that functionality to `assertr`?!
As we learned in the last sections, all assertion verbs in `assertr`
support a custom error function. `chain_end` similarly supports custom
error functions. By default, this is `error_stop` (or `error_report` in the
case of `chain_end`) which prints a summary of the errors and halts execution.
You can specify your own, though, to hijack this behavior and redirect
flow-of-control wherever you want.
Your custom error function must take, as its first argument,
a list of `assertr_error` S3 objects. The second argument must be the
`data.frame` that the verb is computing on. Every error function must
take this because there may be some other errors that are attached to the
`data.frame`'s `assertr_errors` attribute leftover from other previous
assertions.
Below we are going to build a function that takes a list of `assertr_errors`,
gets a string representation of the errors and emails it to someone before
halting execution. We will use the `mailR` package to send the mail.
```{r eval=FALSE, purl = FALSE}
library(mailR)
email_me <- function(list_of_errors, data=NULL, ...){
# we are checking to see if there are any errors that
# are still attached to the data.frame
if(!is.null(data) && !is.null(attr(data, "assertr_errors")))
errors <- append(attr(data, "assertr_errors"), errors)
num.of.errors <- length(list_of_errors)
preface <- sprintf("There %s %d error%s:\n",
ifelse(num.of.errors==1,"is", "are"),
num.of.errors,
ifelse(num.of.errors==1,"", "s"))
# all `assertr_error` S3 objects have `print` and `summary` methods
# here, we will call `print` on all of the errors since `print`
# will give us the complete/unabridged error report
error_string <- capture.output(tmp <- lapply(list_of_errors,
function(x){
cat("\n- ");
print(x);
return();}))
error_string <- c(preface, error_string)
error_string <- error_string %>% paste0(collapse="\n")
send.mail(from="assertr@gmail.com", to="YOU@gmail.com",
subject="error from assertr", body=error_string,
smtp = list(host.name="aspmx.l.google.com", port=25),
authenticate = FALSE, send=TRUE)
stop("assertr stopped execution", call.=FALSE)
}
questionable_mtcars %>%
chain_start %>%
verify(nrow(.) > 10) %>%
insist(within_n_sds(4), mpg) %>%
# ...
chain_end(error_fun=email.me)
```
(this particular `send.mail` formulation will only work for gmail
recipients; see the `mailR` documentation for more information)
Now you'll get notified of <s>any</s> all failed assertions via email. Groovy!
### advanced: creating your own predicate generators for `insist`
`assertr` is build with robustness, correctness, and extensibility in mind.
Just like `assertr` makes it easy to create your own custom predicates, so
too does this package make it easy to create your own custom predicate
generators.
Okay... so its, perhaps, not _easy_ because predicate generators by nature
are functions that return functions. But it's possible!
Let's say you wanted to create a predicate generator that checks if all
elements of a vector are within 3 times the vector's interquartile range from
the median. We need to create a function that looks like this
```{r purl = FALSE}
within_3_iqrs <- function(a_vector){
the_median <- median(a_vector)
the_iqr <- IQR(a_vector)
within_bounds((the_median-the_iqr*3), (the_median+the_iqr*3))
}
```
Now, we can use it on `mpg` from `mtcars` like so:
```{r purl = FALSE}
mtcars %>%
insist(within_3_iqrs, mpg) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
There are two problems with this, though...
1. We may want to abstract this so that we can supply an arbitrary number
of IQRs to create the bounds with
2. We lose the ability to choose what optional arguments (if any) that we
give to the returned `within_bounds` predicate.
Now we have to write a function that returns a function that returns
a function...
```{r purl = FALSE}
within_n_iqrs <- function(n, ...){
function(a_vector){
the_median <- median(a_vector)
the_iqr <- IQR(a_vector)
within_bounds((the_median-the_iqr*n), (the_median+the_iqr*n), ...)
}
}
```
Much better! Now, if we want to check that every `mpg` from `mtcars` is
within 5 IQRs of the median and *not allow NA values* we can do so like this:
```{r purl = FALSE}
mtcars %>%
insist(within_n_iqrs(5), mpg) %>%
group_by(cyl) %>%
summarise(avg.mpg=mean(mpg))
```
Super!
### advanced: programming with assertion functions
These assertion functions use the [tidyeval](https://rpubs.com/hadley/dplyr-programming) framework. In the past,
programming in a tidyverse-like setting was accomplished through
standard evaluation versions of verbs, which used functions postfixed
with an underscore: `insist_` instead of `insist`, for example. However,
when tidyeval was made popular with `dplyr` 0.7.0, this usage became deprecated,
and therefore underscore-postfixed functions are no longer part of `assertr`.