TARGET=outcome
DROPS='search_conducted,search_disposition,search_type'
~15 minutes for type inference and univariate metrics ~30 minutes to tune first embedded model (I think LGBM?) ~24 hours to tune second embedded model (Linear?)
Observations:
### Continuous Features (Accuracy: Higher = More important)
| feature | acc |
|:----------------|----------:|
| hour_of_stop | 1.200e-02 |
| year_of_stop | 7.333e-03 |
| license_lat | 1.333e-03 |
| vehicle_year | 0.000e+00 |
| ... | ... |
### Categorical Features (Accuracy: Higher = More important)
| feature | acc |
|:------------------|----------:|
| violation_type | 3.627e-01 |
| chrg_title | 1.140e-01 |
| chrg_sect_mtch | 9.200e-02 |
| chrg_title_mtch | 5.933e-02 |
| stop_chrg_title | 4.867e-02 |
| article | 2.600e-02 |
| race | 2.267e-02 |
| accident | 1.333e-02 |
| acc_blame | 1.333e-02 |
| prop_dmg | 6.000e-03 |
| unmarked_arrest | 4.667e-03 |
| pers_injury | 4.000e-03 |
| licensed_outstate | 4.000e-03 |
| comm_license | 4.000e-03 |
| belts | 3.333e-03 |
| patrol_entity | 2.667e-03 |
| vehicle_make | 2.000e-03 |
| sex | 0.000e+00 |
| ... | ... |
Above suggests charge and violation_type should be removed as features
(are results, not causes). This also means charge-derived features (eg.
"chrg_*_mtch" features) are suspect for leakage. This also includes the
"article" feature.
subagency: categorical
- '3rd District, Silver Spring'
- '6th District, Gaithersburg / Montgomery Village'
- '4th District, Wheaton'
- '2nd District, Bethesda'
- 'Headquarters and Special Operations'
- '5th District, Germantown'
- '1st District, Rockville'
- 'W15'
- 'S15' latitude: continuous
- latitude of stop longitude: continuous
- longitude of stop accident: bool
- is accident or not belts: bool
- seatbelts worn / present pers_injury: bool
- any person injured prop_dmg: bool
- any property damage fatal: bool
- any fatalities comm_license: bool
- commercial license hazmat: bool
- ??? comm_vehicle: bool
- commercial vehicle alcohol: bool
- alcohol involved work_zone: bool
- stop in a work zone search_conducted: bool
- if search conducted on stop search_disposition: categorical
- "nan", "Property Only", None, "Contraband Only", "Contraband and Property" search_type: categorical
- 'nan', 'Both', 'Person', 'Property', None vehicle_type: categorical
- auto, truck, other, cycle, trailer, bus, farm, etc., 10 categories vehicle_year: ordinal
- year of vehicle vehicle_make: categorical
- chevy, honda, toyota, etc. vehicle_color: categorical
- paint color violation_type: categorical
- Citation, ESERO, SERO, Warning article: categorical
- 'Maryland Rules', 'Transportation Article', 'nan', 'other' acc_blame: bool
- accept blame? >97% are true, not a very useful feature race: categorical
- ancestry sex: categorical
- biological sex ("M", "F", "U") patrol_entity: categorical
- e.g. marked patrol, unmarked patrol, marked laser, etc. hour_of_stop: continuous
- 0-24.0 year_of_stop: ordinal
- 2023, ... month_of_stop: ordinal
- 0-11 weeknum_of_stop: ordinal
- week of year weekday_of_stop: ordinal
- day of week reg_lat: continuous
- registration latitude reg_long: continuous
- registration longitude reg_km: continuous
- distance between vehicle registration location and stop/search location home_lat: continuous
- home latitude (converted from state) home_long: continuous
- home longitude (converted from state) license_lat: continuous
- license latitude license_long: continuous
- license longitude chrg_title: ordinal
- charge title number (suspected charge / stop reason) stop_chrg_title: ordinal
- stop actual charge title number chrg_title_mtch: bool
- if suspected charge matches stop charge chrg_sect_mtch: bool
- if suspected charge section number matches stop charge section number unmarked_arrest: bool
- if arrest was unmarked tech_arrest: bool
- if arrest used technology only (e.g. laser, camera) home_outstate: bool
- if driver lives out of state licensed_outstate: bool
- if driver is licensed out of state
outcome: categorical [TARGET]
- 'Citation', 'Arrest', 'nan', 'Warning', 'SERO', None
There are a number of highly-related categorical variables that can each function as a target (i.e. are caused outcomes), and so should never be included in the same model as features. The Cramer V correlations are:
final stop charge charge
charge charge section title search search viol-
title title matches matches condu- dispos search ation
stop stop -cted ion type type article
outcome 0.16 0.12 0.61 0.70 0.82 0.50 0.49 0.52 0.35 article 0.71 0.46 0.21 0.16 0.03 0.02 0.02 0.58 violation_type 0.34 0.14 0.28 0.20 0.16 0.07 0.09 search_type 0.04 0.04 0.03 0.08 1.00 0.61 search_disposition 0.03 0.04 0.03 0.06 1.00 search_conducted 0.06 0.05 0.03 0.08 chrg_title_mtch 0.29 0.15 0.80 chrg_sect_mtch 0.27 0.08 stop_chrg_title 0.60
search_conducted: bool
- if search conducted on stop search_disposition: categorical
- "nan", "Property Only", None, "Contraband Only", "Contraband and Property" search_type: categorical
- 'nan', 'Both', 'Person', 'Property', None
chrg_title: ordinal
- charge title number (suspected charge / stop reason) stop_chrg_title: ordinal
- stop actual charge title number chrg_title_mtch: bool
- if suspected charge matches stop charge chrg_sect_mtch: bool
- if suspected charge section number matches stop charge section number
violation_type: categorical
- Citation, ESERO, SERO, Warning outcome: categorical [TARGET]
- 'Citation', 'Arrest', 'nan', 'Warning', 'SERO', None
article: categorical
- 'Maryland Rules', 'Transportation Article', 'nan', 'other' acc_blame: bool
- accept blame? >97% are true, not a very useful feature
search_conducted: bool
- if search conducted on stop search_disposition: categorical
- "nan", "Property Only", None, "Contraband Only", "Contraband and Property" search_type: categorical
- 'nan', 'Both', 'Person', 'Property', None
chrg_title_mtch: bool
- if suspected charge matches stop charge chrg_sect_mtch: bool
- if suspected charge section number matches stop charge section number
- TARGET=outcome
- TARGET=violation_type
- TARGET=search_conducted
- TARGET=chrg_title_mtch