animal attacks shark attacks gsaf

Overview

Source: /home/coolhand/html/datavis/data_trove/data/wild/animal_attacks/shark_attacks_gsaf.csv

Saturn profiled 6,462 rows across 257 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).

index numeric identifier

This is a row index running from 0 to 6461 with 6462 unique values and no nulls, matching the row count exactly. Mean and median both equal 3230.5 with zero skew, confirming a perfectly uniform sequential counter rather than a meaningful feature.

Treatment: drop before modelling; it carries no signal.

Case Number text identifier

Almost certainly a case identifier: 6442 unique values across 6462 rows (near_unique), virtually all single tokens (one_word_rate 0.999) of 6-18 chars, and the top repeats look like dates (e.g., '2019.10.08', '2014.08.02'), suggesting the IDs encode an event date. 18 duplicates and 2 nulls exist but are negligible. The 74.8% allcaps rate hints some IDs carry letter suffixes (e.g., '2012.09.02.B') while others are pure numeric dates.

Treatment: Use as a primary key for joins; do not feature-engineer, but consider parsing the leading date for a timestamp column.

Date text timestamp

This is a free-text Date field rather than a parsed timestamp: 87% of values are a single word and the top entries are bare years like '1957' (11) and '1942' (9), alongside non-date tokens such as 'No date' (6) and 'Before 1958' (6). The vocabulary is wide (5,496 unique words across 6,462 rows) and the most frequent word is 'reported' (559), pointing to qualifiers like 'reported before' or 'ca.' that prevent direct date parsing. Duplicate rate is 14% and nulls are negligible (0.0002), but formats are heterogeneous (e.g., '05-Oct-2003').

Treatment: Parse to year with regex, bucketing qualifiers ('ca.', 'before', 'no date') into a separate precision flag.

Year numeric feature

This is a Year column, presumably a release or publication year, with 252 unique values across 6462 rows and almost no nulls (0.05%). The distribution is centered around a median of 1980 with IQR 1943-2006, but extreme outliers stretch the range from 0 to 3019, producing severe negative skew (-6.55) and kurtosis 42.5. About 1.9% of rows are zeros and 4.1% are flagged as outliers, suggesting sentinel values and data-entry errors rather than real years.

Treatment: Clip to a plausible year range (e.g. drop zeros and values >current year) before use.

Type categorical label

Categorical classification of incident type, dominated by 'Unprovoked' at 73% of 6462 rows across 12 distinct values. The long tail reveals redundant labels that look like data-entry drift: 'Watercraft', 'Boat', and 'Boating' are likely the same concept split three ways, and 'Questionable'/'Unconfirmed'/'Unverified' similarly overlap. Null rate is negligible at 0.08%.

Treatment: Consolidate synonymous categories (Boat/Boating/Watercraft; Unconfirmed/Unverified/Questionable) before use as a target or feature.

Country categorical feature

Country code for each record across 205 distinct values, written in uppercase strings. The distribution is heavily Anglosphere-skewed: USA alone accounts for 36.0% of rows (2310), with Australia (1374) and South Africa (585) filling out the top three, and entropy ratio sits at 0.51 indicating concentration well below uniform. Nulls are negligible at 0.79%.

Treatment: Group long tail into 'Other' and one-hot or target-encode before modelling.

Area categorical feature

Geographic area name (state/province), with 810 distinct values across 6462 rows and a 7.16% null rate. Florida dominates at 17.94% of non-null entries, followed by Australian states, Hawaii, California, and South African provinces — suggesting this tracks shark-attack or surf-related locations skewed toward coastal regions. Entropy ratio of 0.64 plus the long_tail alert confirms a heavy head with many rare areas.

Treatment: Group rare areas into 'Other' or target-encode before modelling.

Location text metadata

Free-text geographic descriptors, mostly US coastal place names in 'City, County' form — 'New Smyrna Beach, Volusia County' alone appears 181 times and the words 'county' (1532) and 'beach' (1144) dominate. Values are short (mean 22.75 chars, 3.54 words) but messy: 4148 uniques across 6462 rows with a 29.9% duplicate rate and 8.43% nulls. Language detection flags non-English entries (es 79, pt 65, fr 66, it 58, de 61) suggesting some international shark-attack locations mixed in, though detection on short strings can be noisy.

Treatment: Normalize and geocode to lat/lon or parse into separate city/region/country fields before any spatial analysis.

Activity text feature

Short free-text describing what the victim was doing, dominated by single-word water activities — Surfing (1025), Swimming (932), and Fishing (459) lead the 1516 distinct values. 62.9% of entries are one word and the median length is 8 characters / 1 word, but a long tail stretches to 254 characters, so a minority carry narrative detail. Duplication is heavy (74.3%) and 8.5% are null, indicating this behaves more like a messy categorical than true free text.

Treatment: Normalize case and collapse synonyms, then treat as a categorical with a long-tail 'other' bucket.

Name text free_text

Despite the header, this 'Name' column is a messy mix of actual personal names (john, william, james) and generic descriptors of victims/occupants ('male' appears 579 times, 'female' 106, plus 'boy', 'child', 'sailor', 'Anonymous'). Median length is 13 chars / 2 words, but values run up to 221 chars and include phrases like '2 males' and 'occupants:' constructions, suggesting free-text incident annotations rather than clean names. Duplicate rate is 14.5% and null rate 3.3%, so it cannot be treated as an identifier.

Treatment: Normalize and parse: split out generic descriptors ('male'/'female'/'child') into a separate category feature before any name-level analysis.

Unnamed: 9 categorical metadata

An unnamed trailing column holding a binary M/F flag, almost certainly a stray gender field bled in from a misaligned import. It is 99.57% null with only 28 populated rows (24 M, 4 F), making it unusable as-is.

Treatment: Drop the column; the 99.57% null rate leaves no signal.

Age categorical feature

Age values stored as strings with 154 distinct categories, dominated by teen and young-adult years (17, 18, 20, 19 each appearing 140-160 times). Top value '17' covers only 4.46% and entropy ratio is 0.80, indicating a fairly flat distribution rather than one dominant cohort. The 44.43% null rate is a major concern and the string typing suggests non-numeric entries are mixed in among the 154 levels.

Treatment: Coerce to numeric, investigate the 44% missingness, and impute or bin before modelling.

Injury text free_text

Free-text descriptions of shark-attack injuries, mostly short English phrases (mean 31.5 chars, median 4 words) like 'Foot bitten' or 'Leg bitten'. The dominant value 'FATAL' appears 823 times and drives a 13% all-caps rate, while duplicate_rate of 0.419 (2695 rows) shows many incidents share canned phrasings. A multilingual tail is present but tiny (496 de, 11 fr, 8 nl, 6 es, plus single-row ja/pl/ro/th) against 3812 English entries.

Treatment: Normalize case and map frequent phrases to categorical injury/severity tags (e.g., FATAL, no injury, body-part bitten) before modelling.

Fatal (Y/N) categorical label

Binary fatality flag for shark attack records, with 'N' covering 75.0% of 6462 rows and 'Y' another 1400. Despite being nominally Y/N, the column has 7 distinct values including 71 'UNKNOWN', stray 'F'/'M' (likely sex leaks), a '2017' (year leak), and a lowercase 'y'. Null rate is 8.46%.

Treatment: Normalize case, map UNKNOWN/F/M/2017 to missing, then treat as binary target.

Time categorical metadata

A time-of-day field mixing precise clock values (e.g. '11h00', '15h00') with vague buckets like 'Afternoon' and 'Morning', which together account for the top two categories. Over half the rows are null (null_rate 0.5249) and there are 366 distinct values across only 6462 records, with the modal value covering just 6.3% of non-nulls. The format inconsistency and high missingness make this unreliable as-is.

Treatment: Normalize to a consistent hour bucket (parse 'HHhMM', map vague labels) and add a missingness indicator before use.

Species text free_text

Free-text descriptions of the shark species involved in each incident, ranging from clean labels like "White shark" (166) and "Tiger shark" (81) to size-only guesses like "4' shark" (41) and disclaimer strings like "Shark involvement prior to death was not confirmed" (105). The field is sparsely populated (null_rate 0.4525) and heavily repetitive (duplicate_rate 0.586, only 1466 unique values across 6462 rows), and the language detector flags 14 non-English buckets though the 2582 English count dominates. Notable surprise: several top values are non-identifications or size estimates rather than species, so this column mixes taxonomy, measurements, and investigative status.

Treatment: Normalise and parse into separate fields (species label, size, confirmation status) before using as a feature.

Investigator or Source text metadata

Short attribution strings naming the investigator or source of each shark-incident record, dominated by initialled surnames often suffixed with 'GSAF' (983 occurrences) and a recurring 'V.M. Coppleson' contributor. Despite 4,979 unique values across 6,462 rows, 22.7% are duplicates and the top contributor 'C. Moore, GSAF' appears 106 times, indicating a long tail of repeat investigators. Language detection flags a multilingual mix (4,457 en, 134 es, 100 fr, 62 it, plus 19 other languages), but this is likely a false positive driven by initials and proper nouns rather than real translation.

Treatment: Normalize and split into investigator vs. source/agency fields before using as a categorical feature.

pdf text metadata

This column appears to hold a single PDF filename or reference token per row, with 98.7% of values being one word and a mean length of ~23 characters. It is sparse and high-cardinality: 52.55% null and 3,054 unique values across 6,462 rows, with only 12 duplicates. Top tokens like 'fisherman.pdf', 'bay.pdf', and '1935.06.05.r-solomonislands.pdf' confirm these are document filenames rather than free text.

Treatment: Treat as a document reference; drop for modelling or use only as a join key to source PDFs.

href formula text metadata

This column holds single-token URLs pointing to PDF case files on sharkattackfile.net, with url_rate of 0.9997 and one_word_rate of 0.988. It is sparsely populated (null_rate 0.5262) and nearly unique (3051 distinct values out of 6462, only 11 duplicates). Lengths are tightly bounded between 64 and 95 characters, consistent with a templated href pattern.

Treatment: Drop from modelling; retain as a source-document link for reference.

href text metadata

This column holds URLs to PDF case files (url_rate 0.9997, one_word_rate 0.988), almost certainly links to source documents on sharkattackfile.net. Over half the rows are null (null_rate 0.5262) and of the populated values 3051 out of 6462 are unique with only 11 duplicates, so it functions as a near-unique reference rather than a feature. Lengths are tight (min 64, max 135, median 77), consistent with a single hosted directory pattern.

Treatment: Keep as a source-document link for audit; drop from modelling features.

Case Number.1 text identifier

This is almost certainly a secondary case identifier, likely a date-stamped reference code (e.g. '1966.12.26', '1915.07.06.a.r') with 79% allcaps tokens and 99.8% single-word values averaging 10.6 characters. With 3054 unique values across 6462 rows and a 52.6% null rate, it's near-unique among populated rows and absent on more than half. Only 8 duplicates exist, so it functions as an identifier rather than a categorical feature.

Treatment: Treat as a secondary key for joins; do not use as a model feature.