quirky caves

Overview

Source: /home/coolhand/html/datavis/data_trove/data/quirky/caves.json

Saturn profiled 69,716 rows across 12 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

id numeric identifier

This column is almost certainly a row identifier: all 69716 values are unique with zero nulls, and the numeric range spans roughly 1.28e7 to 1.354e10 with near-zero skew (0.02) and negative kurtosis (-1.18), consistent with broadly distributed assigned IDs rather than a measured quantity. No outliers or zeros are flagged. Treat the numeric stats as incidental — the magnitudes carry no analytical meaning.

Treatment: Drop from modelling; retain only as a join key.

name text label

Short free-text names of caves/caverns in mixed languages (English 'Cave', French 'Grotte', Italian 'Grotta', Spanish 'Cueva', Cyrillic 'Грот', German 'Bärenhöhle'), averaging 2.5 words and 15.5 characters. Severe duplication (35.1%, 24,487 rows) is dominated by the placeholder 'Unnamed Cave' appearing 19,527 times — over a quarter of all rows are effectively unlabelled. Of 69,716 rows only 45,229 are unique, and the vocabulary of 15,219 tokens reflects the multilingual mix.

Treatment: Treat 'Unnamed Cave' as missing and language-normalise before any text matching or grouping.

lat numeric feature

Latitude coordinates spanning -77.98 to 78.18, so the column covers nearly the full globe from Antarctic to Arctic ranges. The distribution is heavily left-skewed (skew -3.16, kurtosis 11.5) with a tight IQR of 7.26 around a median of 44.14, indicating most points cluster in northern mid-latitudes while a long tail of southern-hemisphere values produces 8,996 outliers (12.9%). Near-unique values (69,544 of 69,716) confirm these are precise geocoordinates rather than bucketed regions.

Treatment: Pair with longitude for geospatial features; consider binning or projecting rather than treating as a raw scalar.

lon numeric feature

This is a longitude coordinate, with values spanning the full -178.00 to 178.80 range and 69,585 unique values across 69,716 rows. The distribution is centered near 12.03 (median 11.38) with IQR 1.24 to 18.18, suggesting a heavy concentration in European/African longitudes, but the std of 40.50 and 16.2% flagged outliers reveal a long global tail. No nulls or zeros are present.

Treatment: Pair with latitude for geospatial features; do not treat outliers as errors since the global range is legitimate.

description text free_text

A free-text 'description' field, but 65189 of 69716 rows (94.7% duplicate rate) are empty strings, leaving only ~4500 populated entries with a mean length of 3.46 characters and median word count of 1. The non-empty values are short multilingual fragments — German ('nicht katasterwürdig', 'unterer Eingang'), French ('Entrée d'une carrière souterraine'), and other tokens like 'cave' and 'Halbhöhle' — suggesting cave/feature annotations rather than prose. With 92.8% one-word rate and a vocab of 4651, this is closer to sparse tagging than narrative text.

Treatment: Treat as sparse optional tag: flag presence as a boolean feature and ignore the text body unless modelling the populated subset.

access categorical feature

Categorical access-permission tag, almost certainly the OSM-style `access` key indicating who may use a feature. 89.7% of the 69,716 rows are empty strings, leaving only ~10% with substantive values like `yes` (2,717), `no` (2,266), `private` (815), `permit` (575), and `permissive` (440). Entropy ratio is just 0.16 and 20 distinct values appear, so the signal is sparse but the long tail (`customers`, `destination`, `restricted`, `unknown`) is meaningful when present.

Treatment: Treat empty string as missing, then collapse rare levels and one-hot encode the survivors.

tourism categorical feature

This is an OSM-style `tourism` tag categorising features like attractions, viewpoints, and museums across 18 distinct values. The column is severely imbalanced: 97.2% of the 69,716 rows are empty strings, with `attraction` (1,670) and `viewpoint` (119) the only non-trivial categories. Entropy ratio of 0.05 confirms almost no information content as-is.

Treatment: Treat empty string as missing and either drop or collapse rare categories into a binary `is_tourism` flag.

wikipedia text metadata

This column holds Wikipedia article references in the OSM-style `lang:Article Title` format (e.g. `de:Einödhöhle`, `fr:Grotte...`), mostly pointing to cave-related pages across many languages. It is overwhelmingly empty: 67,531 of 69,716 rows are blank and the duplicate rate is 0.97, leaving only 2,077 unique values. Among the populated entries, language prefixes span de, fr, pl, es, it, ca, en, ru and more, so any downstream use must handle multilingual strings.

Treatment: Split into language prefix and title, and treat as a sparse optional reference rather than a modelling feature.

website text metadata

This is a website/URL field for each record, but it is overwhelmingly empty: 67,082 of 69,716 rows (n_empty) are blank, leaving only 2,492 unique values across the column. Where populated, entries are single-token URLs (one_word_rate 0.99998, word_mean ~1.0) pointing to varied external domains (angloasianmining.com, unesco.org, termeszetvedelem.hu, etc.). The duplicate_rate of 0.96 is driven almost entirely by the empty string, and url_rate is only 0.038 because the metric is computed over all rows including blanks.

Treatment: Treat as optional reference URL; impute missing as null and avoid using as a feature given >96% blanks.

cave:length categorical metadata

This appears to be a cave length attribute (likely OpenStreetMap-style tag) stored as a string, with 99.08% of the 69,716 rows being empty and only 238 distinct values overall. When populated, values look like small integers (5, 6, 10, 3, 4...) suggesting metres, but the signal is so sparse (entropy ratio 0.0176) that it carries almost no information. The non-null counts in the top values are tiny (≤32 each), indicating this tag is rarely filled in upstream.

Treatment: Drop or convert to a numeric 'has_length' indicator; the column is too sparse to model directly.

cave:depth categorical metadata

This appears to be a cave depth attribute, likely sourced from an OpenStreetMap-style tag, stored as strings. It is effectively empty: 69,419 of 69,716 rows (top_rate 0.9957) carry the blank value "", and the remaining 124 distinct values are tiny integer-like strings ranging from "0" to "30" with single- or low-double-digit counts. Entropy ratio of 0.0092 confirms there is virtually no signal here.

Treatment: Drop; near-constant blank with negligible entropy.

country categorical feature

Two-letter ISO country code, but effectively empty: 69,684 of 69,716 rows (top_rate 0.9995) carry a blank string, leaving only 32 rows spread across 15 actual codes (KY, AU, DE, RO, US, etc.). Entropy ratio of 0.0018 confirms there is essentially no signal here. No nulls are reported because the missingness is encoded as empty string rather than NULL.

Treatment: Drop; near-constant empty string with only 32 populated rows.