In smaakage85/recordr: Toolkit to Validate New Data for a Predictive Model

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-"
)

recorder

recorder is a lightweight toolkit to validate new observations when computing their corresponding predictions with a predictive model.

With recorder the validation process consists of two steps:

1. record relevant statistics and meta data of the variables in the original training data for the predictive model
2. use these data to run a set of basic validation tests on the new set of observations.

Motivation

There can be many data specific reasons, why you might not be confident in the predictions of a predictive model on new data.

Some of them are obvious, e.g.:

• One or more variables in training data are not found in new data
• The class of a given variable differs in training data and new data

Others are more subtle, for instance if it is the case, that observations in new data are not within the "span" of the training data.

If one or more of the recorder validation tests fail on new data, you might not be confident in the corresponding predictions.

Installation

recorder can be installed from CRAN with install.packages('recorder'). If you want the development version then install directly from GitHub:

devtools::install_github("smaakage85/recorder")

Workflow Example

Get ready by loading the package.

library(recorder)

The famous iris dataset will be used as an example. The data set is divided into training data, that can be used for model development, and new data for predictions after modelling, which can be validated with recorder.

set.seed(1)
trn_idx <- sample(seq_len(nrow(iris)), 100)
data_training <- iris[trn_idx, ]
data_new <- iris[-trn_idx, ]

Record statistics and meta data of the training data with record().

tape <- record(data_training)

Run validation tests on new data with play().

playback <- play(tape, data_new)

Print the over-all results of the validation tests.

playback

The test summary tells us, that one observation (row #11) has a value of the variable "Petal.Length" outside the recorded range in the training data; hence we might not be confident in the prediction of this particular observation.

After running the validation tests, you can extract the results of (any) failed tests for the rows/observations of new data with get_failed_tests().

failed_tests <- get_failed_tests(playback)
# print.
library(knitr)
kable(head(failed_tests, 15))

You might also find the functions get_failed_tests_string() and get_clean_rows() to be useful.

That is basically it. If you to know more about all of the exciting features of recorder, take a look at the vignette.

Also, if you have any feedback on the package, please let me hear from you.

smaakage85/recordr documentation built on June 13, 2019, 7:08 p.m.