Prediction in Traffic Accident Duration Based on Heterogeneous Ensemble Learning

Figures & data

Figure 1. Heterogeneous integration stacking process.

Table 1. Original basic characteristics

Attribute	Characteristic variable
Traffic	TMC, Start time, End time, Start lat(latitude), Start lng(longitude)
Location	Street, Side(right/left), City, County, State, Zipcode, Airport code
Weather	Temperature(F), Wind chill(F), Humidity(%), Pressure(in), Visibility(mi), Wind direction, Wind speed(mph), Precipitation(in), Weather condition
POI	Amenity(true or false), Bump, Crossing, Give Way, Junction, No Exit, Railway, Roundabout, Station, Stop, Traffic Calming, Traffic Signal, Turning Loop
Time	Timezone, Sunrise sunset(day or night), Civil twilight, Nautical twilight, Astronomical twilight

Table 2. Missing data

Feature	Missing rate(%)	Feature	Missing rate(%)
City	0.0021	Wind direction	1.4068
Zipcode	0.0128	Wind speed(mph)	14.0116
Timezone	0.0854	Precipitation(in)	62.9072
Airport code	0.1708	Weather condition	1.9682
Temperature(F)	1.6342	Sunrise sunset	0.0022
Wind chill(F)	58.7179	Civil twilight	0.0022
Humidity(%)	1.7422	Nautical twilight	0.0022
Pressure(in)	1.4040	Astronomical twilight	0.0022
Visibility(mi)	1.9697

Table 3. Effects of three missing filling schemes

Scheme	Feature	Validation MSE	Offline test MSE
(${\mathrm{A}}_1$)	Pressure(in)	0.0202	0.0316
	Temperature(F)	10.94	72.4596
	Humidity(%)	81.4766	263.0995
	Visibility(mi)	2.3016	4.5374
	Wind speed(mph)	9.2113	21.0961
(${\mathrm{A}}_2$)	Pressure(in)	0.0202	0.0316
	Temperature(F)	10.1585	68.4917
	Humidity(%)	55.6164	259.3242
	Visibility(mi)	2.0031	4.2732
	Wind speed(mph)	8.5267	21.8461
(${\mathrm{A}}_3$)	Pressure(in)	0.0202	0.0293
	Temperature(F)	7.2011	57.1106
	Humidity(%)	54.3567	177.6861
	Visibility(mi)	1.9709	4.2178
	Wind speed(mph)	8.5209	21.7633

Table 4. Constructed features

Dimension	Statistical objects	Unit	Statistics
Time	Start time	year, month, day, hour, minute, second	year, month, day, hour, minute, second
		day	day of the week, is weekend, is holiday
		hour	morning peak, evening peak, is period of high accident duration
Space	Country, State, County, City, Street	number of accident records	accumulation of historical accidents
	State, County, City, Street, Zipcode, 5-digit Zipcode	yearly, monthly	maximum, minimum, mean, median, standard deviation, cumulative count
Second-order historical information	State, County, City, Street, Zipcode, 5-digit Zipcode cross with Wind direction and Weather condition	yearly, monthly	maximum, minimum, mean, median, standard deviation, cumulative count

Figure 2. Accident record cumulative SHAP values.

Table 5. The prediction effects of different models

Model	MAPE(%)	MAE	MSE
XGBoost	36.6540	32.1356	4413.2723
LightGBM	35.4587	31.1413	4314.8268
CatBoost	36.2398	31.1533	4302.6698
Decision tree	52.9972	32.1943	6902.1944
Elastic net	60.2050	34.0313	3861.6144
Heterogeneous ensemble	35.6101	30.7432	4252.1728

Figure 3. SHAP value.

Figure 4. Feature importance of combinatorial models.