datagrade

Point it at a data frame. Get a score.

datagrade assesses the quality of a tabular data set against the ISO/IEC 25012 data quality model — five inherent characteristics, fifteen quality properties — and returns a report object with tables and figures.

Six of the fifteen properties are computable from a table alone, so it runs on anything with columns. That is the point. The existing R packages for data quality (DQA, DQAstats, dataquieR) all require a metadata repository or data dictionary before they will run at all.

The other nine measure the data against an expectation, and an expectation is something a person declares — so dg_spec() is where you declare it. Supply one and those measures light up. Supply nothing and they report as not applicable, never as zero: a score of zero is a claim about the data, NA is a statement that the standard’s denominator could not be counted.

Install

# install.packages("remotes")
remotes::install_github("ohalukdemirhan/datagrade")

Use

library(datagrade)

report <- dg_assess(ggplot2::diamonds)

── Data quality report ─────────────────────────────────────────────────
Source: in-memory data frame
Size: 53940 rows x 10 columns
Elapsed: 0.15s

── ISO/IEC 25012 characteristics ──

• accuracy: not applicable (0/3 measures)
• completeness: 1.000 (100%) - good (3/4 measures)
• consistency: 0.600 (60%) - serious (2/4 measures)
• credibility: not applicable (0/2 measures)
• currentness: not applicable (0/2 measures)

Overall: 0.800 (80%) - warning

Not ISO: plausibility 0.993 (99.3%) - outliers against a derived interval,
excluded from the overall score.

ℹ 10 of 15 measures need a declared expectation. See `report$measures` and `dg_spec()`.

── Findings ──

• Missing values: 0 across 0 columns
• Duplicate rows: 146 (0.27%)
• Outliers (z_score): 2798 of 377580 numeric values
• No identifier column detected
• Redundant pairs above |r| > 0.5: 10 (strongest carat~x at 0.98)
• Suggested to drop: carat, price, x, and y
• VIF above 5: carat, price, x, y, and z

The printout is a side effect. The object is the product:

report$scores            # five characteristics, ratios in [0, 1]
report$measures          # fifteen properties: code, value, A, B, applicable
report$redundancy$pairs  # data frame of correlated column pairs
summary(report)$columns  # per-column type, missingness, outliers, VIF
as.data.frame(report)    # one row, for stacking across many data sets

Declaring what the data should look like

spec <- dg_spec(
  required = c("id", "dept", "amount"),
  patterns = list(id = "^[A-Z]{2}-[0-9]{4}$"),
  domains  = list(dept = c("SALES", "OPS", "HR")),
  ranges   = list(amount = c(0, 1000)),
  rules    = list(order = ~ start <= end)
)

dg_assess(orders, spec = spec)

Every argument is optional, and each one switches on the measures that depend on it:

report$measures says which, and why the rest are still NA:

          code   characteristic value    a    b                                note
     EXAC_SINT         accuracy 1.000  400  400                                <NA>
    EXAC_SEMAN         accuracy    NA   NA   NA  needs spec$reference and reference_key
      RAN_EXAC         accuracy 0.974  187  192                                <NA>
      COMP_REG     completeness 0.960  192  200                                <NA>
  COMP_VAL_ESP     completeness 0.990  792  800                                <NA>
     COMP_FICH     completeness    NA   NA   NA          needs spec$expected_records
 FAL_COMP_FICH     completeness 1.000  400  400                                <NA>

A is the count of items satisfying the property, B the count of items for which the property is defined at all. That second number is where the metadata lives, and it is the whole reason nine measures need a spec — and the whole reason an unmeasurable property reports NA rather than zero:

Figures come from dg_plots(report):

Also works on files — dg_assess(path = "data.csv") — for .csv, .tsv, .txt, .xls and .xlsx.

It scales

Benchmarked on ggplot2::diamonds (53,940 × 10, MIT licensed), then the same real data replicated with jitter. R 4.5.2, macOS arm64, single core.

Cost per row is 1.37 µs at 54 thousand rows and 0.94 µs at 10 million — it goes down, because fixed startup amortises. The input grows 185-fold; the report object grows 1%, because it holds statistics and never the data. Figures render in constant time at every size. Scores are identical at every size, which is how you know the subsampling isn’t biasing anything.

Reproduce it yourself: Rscript inst/bench/benchmark.R.

Three dependencies: cli, ggplot2, base R.

Documentation

Everything technical lives in the vignettes, not here.

• vignette("datagrade") — the full walkthrough
• vignette("methodology") — every formula, written out
• vignette("scalability") — how the numbers above are achieved
• vignette("deviations") — how this differs from the dissertation it implements

The diagrams above are generated, not drawn. Their sources are in inst/diagrams/*.mmd and regenerate with

mmdc -i inst/diagrams/iso-model.mmd -o man/figures/iso-model.png -b '#fcfcfb' -s 1.5 -w 1600

They use the same palette as the figures, dg_palette in R/palette.R, so a diagram and a plot in the same README are the same colour system.

Background

This implements the framework from:

Demirhan, O. H. (2024). Assessing Data Quality Management Issues in Open-Sourced Data Sets with Statistical Methodology. MSc dissertation, University of Greenwich.

The dissertation scored four dimensions with measures of its own devising. This release keeps its statistics but restructures the scoring to follow the ISO/IEC 25012 model, so numbers here will not always match numbers in the dissertation. vignette("deviations") lists every difference and the reason for it.

The fifteen properties

Two of the six zero-configuration measures deserve a note, because both infer their expectation from the column itself rather than from a declaration:

• FAL_COMP_FICH compares text values against a short, deliberately conservative list of placeholders — "N/A", "unknown", "-999" and a dozen more, matched case-insensitively after trimming. Pass null_equivalents to replace the list, or character(0) to switch the check off.
• CONS_FORM reduces each value to a shape — digit runs become 9, letter runs A — and measures how many share their column’s dominant shape. A column whose values collapse to many shapes has no format to be consistent with, so free text is excluded from the denominator and named in the measure’s note rather than scored near zero.

What is not claimed

plausibility is reported alongside the five characteristics and excluded from the overall score. It is the dissertation’s accuracy measure: outliers against an interval derived from the data (μ ± 3σ, or the IQR fence). Every ISO accuracy property requires an interval, pattern or reference that was declared, so substituting a statistically derived one measures plausibility, not accuracy — an outlier need not be wrong, and a wrong value can sit exactly on the mean. It is kept because it is genuinely useful on undocumented data, and named honestly because it is not the standard’s measure.

Aggregation from properties to characteristics is the unweighted mean over the properties that apply; from characteristics to overall it is the weighted mean over the characteristics that apply. ISO/IEC 25024 does not specify how measures aggregate, so both steps are this package’s choice, not the standard’s.

References

The ISO claims above are sourced. The standards themselves are paywalled. The model’s structure — the five inherent characteristics, the fifteen properties beneath them and their identifiers — is taken from the peer-reviewed literature cited below. The standard’s own measurement functions are not reproduced: each measure here is this package’s implementation of the named property, in the A / B form the standard uses, with its A and B stated in report$measures and derived in vignette("methodology").

• ISO/IEC 25024:2015. Systems and software engineering — Systems and software Quality Requirements and Evaluation (SQuaRE) — Measurement of data quality. International Organization for Standardization. https://www.iso.org/standard/35749.html
• ISO/IEC 25012:2008. Software engineering — SQuaRE — Data quality model. International Organization for Standardization. https://www.iso.org/standard/35736.html
• Gualo, F., Rodriguez, M., Verdugo, J., Caballero, I. and Piattini, M. (2021). ‘Data Quality Certification using ISO/IEC 25012: Industrial Experiences’, Journal of Systems and Software. https://arxiv.org/abs/2102.11527 — source for the X = A/B measurement-function form, for the Accuracy Range (RAN_EXAC) measure whose element B is “the number of data items for which a required interval of values can be defined”, and for the statement that ISO/IEC 25024 does not specify how measures aggregate into a per-characteristic score. Written by an ISO/IEC 25012 accredited evaluation laboratory.
• Calabrese, J., Esponda, S. and Pesado, P. (2020). ‘Framework for Data Quality Evaluation Based on ISO/IEC 25012 and ISO/IEC 25024’, 8th Conference on Cloud Computing, Big Data & Emerging Topics, Universidad Nacional de La Plata. https://sedici.unlp.edu.ar/handle/10915/104778 — source for the five inherent characteristics being Accuracy, Completeness, Consistency, Credibility and Currentness, and for the decomposition of each into the fifteen quality properties and their identifiers (EXAC_SINT, COMP_REG, RIES_INCO, CRED_FUEN, CONV_ACT and the rest) implemented here. What remains unverified against the standard itself is the exact definition of each measure’s quality measure elements — precisely which items ISO counts into A and B for a given property. Those are in Annexes A–E of ISO/IEC 25024, which are not in any public source. Where this package had to choose, the choice is stated in vignette("methodology") and visible in the a and b columns of report$measures, so it can be checked and disagreed with.

Licence

MIT © Omer Haluk Demirhan