urban parking violations sample

Overview

Source: /home/coolhand/html/datavis/data_trove/data/urban/parking_violations_sample.csv

Saturn profiled 10,000 rows across 9 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).

summons_number numeric identifier

Every one of the 10,000 rows carries a distinct value (n_unique = 10000, null_rate = 0.0), and the magnitudes (min 1.12e9, max 9.26e9) match the size of NYC parking summons numbers. The wide spread (std ≈ 2.71e9) and lack of outliers reflect identifier allocation rather than a measurable quantity. Despite being typed numeric, no arithmetic interpretation applies.

Treatment: drop from modelling; retain only as a row key for joins.

issue_date categorical timestamp

This is an issue_date column stored as ISO datetime strings, treated here as categorical across 687 distinct days with no nulls. The distribution is severely concentrated: 65.42% of all 10000 rows fall on 2025-12-28, with another 1594 on 2025-12-30 and 356 on 2025-12-29, meaning roughly 85% of issuance clusters in late December 2025 before tapering into a long tail through 2026. Entropy ratio of 0.29 confirms the heavy skew, and the year-end spike looks like a backfill or batch-load artifact worth confirming before treating this as a true event date.

Treatment: Parse to datetime and bucket by month or week; investigate the 2025-12-28 spike before using as a feature.

violation_code numeric feature

This is almost certainly a categorical violation code stored as a number, with 62 distinct values across 10,000 rows and no nulls. Despite the numeric type, the distribution is meaningless as a quantity: values span 4 to 99, the IQR runs 21–36, and 10.07% of rows (1,007) flag as outliers under a numeric rule, with skew 1.51 and kurtosis 2.99. Median equals Q3 at 36, suggesting a heavy concentration at one or two dominant codes.

Treatment: Treat as categorical and one-hot or target-encode rather than using the raw integer.

violation_description categorical feature

Categorical column describing the parking/traffic violation issued, with 74 distinct codes across 10000 rows. It is heavily concentrated: 'PHTO SCHOOL ZN SPEED VIOLATION' alone covers 52.03% of records, with 'No Parking Street Cleaning' a distant second at 1428, yielding low entropy ratio 0.452. Note 15.13% of values are null, and the labels mix numeric-prefixed legal codes (e.g. '14-No Standing') with free-form descriptions like 'Fire Hydrant', some of which appear to duplicate the coded versions ('40-Fire Hydrant' vs 'Fire Hydrant').

Treatment: Normalise the coded vs free-text duplicates, then group rare categories before one-hot encoding.

street_name text feature

Street-name strings, mostly truncated NYC traffic-camera or incident locations like 'SB CROSS BAY BLVD @' with directional prefixes (sb/wb/nb/eb) and '@' delimiters dominating the top words. Values are heavily repeated (68.8% duplicate rate, only 3,115 uniques in 10,000 rows) and 77.3% are all-caps; lengths cap sharply at 20 characters, suggesting upstream truncation. Language detection flags ja (614) and zh (59) alongside 3,596 en, but this is almost certainly a false positive on short ALL-CAPS abbreviations rather than genuine multilingual content.

Treatment: Normalise case, split off the directional prefix and the '@' cross-street, and treat as a categorical location feature.

vehicle_make categorical feature

Categorical vehicle manufacturer codes, with HONDA leading at 13.4% of 9,922 non-null rows and TOYOT close behind at 1,302. Values appear truncated to 5 characters (TOYOT, NISSA, ME/BE, CHEVR, HYUND, SUBAR), which will fragment any join against full make names. Cardinality is 126 with entropy ratio 0.668, indicating a long but moderately concentrated tail, and nulls are negligible at 0.78%.

Treatment: Normalize the truncated codes to canonical make names, then target- or frequency-encode for modelling.

vehicle_color categorical feature

Vehicle color codes, but encoded inconsistently: short codes like GY (2079), BK (1784), WH (1579) coexist with verbose forms WHITE (347), BLACK (273), GREY (239), and alternate abbreviations BLK (275), GRY (167) for the same underlying colors. With 99 distinct values across 10000 rows and a 9.43% null rate, the cardinality is inflated by these duplicate encodings rather than true diversity. Entropy ratio of 0.55 reflects a heavy concentration in the gray/black/white tail.

Treatment: Normalize synonymous codes (e.g., BK/BLK/BLACK → black) before one-hot or target encoding.

violation_time text timestamp

This column encodes time-of-day stamps in a compact HHMM+AM/PM format (e.g. '0839A', '1200P'), with every value exactly 5 characters and uppercase. Duplication is high (85.67%) across 1,432 unique stamps, which is expected for clock times sampled across 10,000 records. Null rate is negligible (0.0004) and there are no empty strings.

Treatment: Parse the HHMMa/p format into a proper time-of-day (minutes since midnight) before modelling.

issuing_agency categorical feature

Single-letter codes for the agency issuing each record, drawn from a closed set of 20 values with no nulls. Distribution is heavily concentrated: 'V' alone covers 44.16% (4,416/10,000) and the top four codes (V, T, S, P) account for the bulk of rows, while letters like K, N, A, Y, M, O appear fewer than 50 times each. Entropy ratio of 0.46 confirms the imbalance.

Treatment: One-hot encode the top categories and bucket the long tail into 'other' before modelling.