Exploratory framework for analysing road traffic accident data with validation on Gauteng province data

Figures & data

Table 1. Methods applied during experiments

Methods
Imputation methods	Mean
	k-nearest neighbor
Dimensionality reduction	Principal component analysis
	Linear discriminant analysis
Classifiers	Naïve Bayes
	Logistic regression
	k-nearest neighbor
Metrics	Accuracy
	Root means square error

Figure 1. Road traffic accident experimental framework

Table 2. Features used for investigation

#	Features	Description
1	InvolvedVehicles	Refers to the number of vehicles that were involved in the road accident.
2	TotalNoVictims	Refers to the number of people/individuals that lost their lives during a specific road accident.
3	Gender	Refers to the different gender involved during a traffic accident such as F, M and Others.
4	Location	Refers to the place where an incident/road accident occurred.
5	Season	Refers to one of the four seasons: Spring, Summer, Winter and Autumn.
6	Month	Refers to the different number of months in a year with peak and off-peak RTAs.
7	DayOfWeek	Refers to the days of a week, this covers the number of RTAs for weekdays and weekends.
8	CrashDate	Refers to the date on which an RTA occurred.

Table 3. Dataset class value distribution

#	Events	Values	%
1	Pedestrian	71	32.57
2	Overturned	28	12.84
3	Unknown	20	9.17
4	Fixed Object	19	8.72
5	Head-On	16	7.34
6	Head-Rear	15	6.88
7	Collision	11	5.05
8	Sideswipe	10	4.59
9	Rear-End	9	4.13
10	Multiple vehicles	7	3.21
11	Hit and run	6	2.75
12	Motorcycle	6	2.75
Total		218	100

Figure 2. Summary of RTAs yearly

Figure 3. Data distribution (a) Gender, (b) Month, (c) Location and (d) Season

Figure 4. Histogram (a) DayOfWeek and (b) Event

Figure 5. Boxplot (a) TotalNoVictims vs Event, (b) InvolvedVehicles vs Event

Table 4. Number of PCs by feature importance

#	Original Features
1	Gender
2	Season
3	CrashDate
4	DayOfWeek
5	TotalNoVictims
6	InvolvedVehicles
7	Month
8	Gender

Figure 6. MEAN IM Data distribution PC1 and PC2 grouped by classes

Figure 7. K-NN IM Data distribution PC1 and PC2 grouped by classes

Figure 8. MEAN IM Data distribution LD1 and LD2 grouped by classes

Figure 9. K-NN IM data distribution LD1 and LD2 grouped by classes

Table 5. Summary of the model results

Models Default Settings
Missing value method: Mean						Missing value method: k-NN
#	Model/metrics	Accuracy1 %		RMSE		Accuracy2 %	RMSE
1	NB	7		13.233		12	15.953
2	LR	37		15.628		42	14.465
3	k-NN	28		20.86		37	18.744
Principal Component Analysis
Missing value method: Mean
		n = 8		n = 3		n = 2
#	Model/metrics	Accuracy %	RMSE	Accuracy %	RMSE	Accuracy %	RMSE
4	NB	37	13.977	37	15.442	37	13.419
5	LR	37	15.628	33	13.07	33	11.953
6	k-NN	28	20.860	26	18.395	23	21.558
Missing value method: k-NN
#	Model/metrics	Accuracy %	RMSE	Accuracy %	RMSE	Accuracy %	RMSE
7	NB	35	17.326	37	11.767	35	13.326
8	LR	42	14.465	44	9.163	37	11.861
9	k-NN	37	18.744	37	16.721	19	17.279
Linear Discriminant Analysis
Missing value method: Mean
		n = 8		n = 3		n = 2
#	Model/metrics	Accuracy %	RMSE	Accuracy %	RMSE	Accuracy %	RMSE
10	NB	33	17.163	40	16.233	37	16.581
11	LR	40	15.512	37	18.953	37	18.558
12	k-NN	30	21.674	30	24	28	24.884
Missing value method: k-NN
#	Model/metrics	Accuracy %	RMSE	Accuracy%	RMSE	Accuracy %	RMSE
13	NB	47	11.558	49	14.047	51	13.023
14	LR	42	14.465	49	10.674	49	10.977
15	k-NN	40	18.047	40	12.093	42	12.512

Figure 10. RTA Classification Models (a) Default settings (b)PCA and LDA components 2, (c) PCA and LDA components 3 and (d)PCA and LDA components 8