``An application of frequent pattern mining to traffic accident data" 自動車技術会 2016年春季大会 in パシフィコ横浜 事故分析と安全対策Ⅰ にて発表したスライドです. 時間がオーバーするため泣く泣く非表示にしたスライドを追加してあります. 質疑応答 Q. 事故データは人がラベル付けした? A. 警察の方々がやっています. Q. 3つ目の分析で相関以外の分析を行っては? (線形になるとは限らないので) A. 確かにそうですね(思いつかなかった). supportの増加量・減少量が大きいパターンを抽出してみるのもありだと思います. Q. センサデータを取り入れたビックデータで処理できる? A. 数値データを扱うなら, 別の手法が好ましそうです. センサデータを処理してラベルデータを抽出するならありかもです(事故の衝撃の度合, 車の具体的な損傷等).

1. An application of frequent pattern mining
to traffic accident data
交通事故データへの頻出パターンマイニングの適用
Yuta Takahashi1), Masaru Kiyota1), Yukuo Hayashida1),
Yuichiro Sakamoto1)
1) Saga University

2. Background
• Traffic accidents per hundred thousands in Saga are
most often in Japan while 4 years
2
0
200
400
600
800
1000
1200
2012 2013 2014 2015
NumberofTrafficAccidents
(perhundredthousands)
year
Saga
Average
We haven’t identified the cause
clearly.
Due to complication
Over
2 times!!

3. PDCA cycle for preventing accidents
3
Plan
Do
Check
Action
• Analysis a cause of accidents
• Propose a prevention
method
• Take measures
• Verify the effectiveness of
action
• Consider the lack of
measures
• Schedule next measures
Do
Plan
Check
Action
PCDA cycle

4. PDCA cycle for preventing accidents
4
Plan
Do
Check
Action
PCDA cycle
 Procedure of “Plan”
Select feature
points
Analysis of
accident causes
Planning of
measures
Research entities
Position of
this study

5. Problem of previous analysis
• Ordinary we make many tables and graphs
⇒ It needs much cost to make and check
5
• We attend to introduce a data mining approach

6. Data mining
• A method of gaining knowledge from large data
• It is different from statistical analysis
• Performance takes priority over strictness
6
 SOM
 Clustering
 K-means
 Association rule analysis
 Neural Network
 Decision Tree
 SVM
 Bayesian network
① Learning ② Extract patterns

7. Related works
 Machine Learning Approach
7
Dataset
Fatal injury
Classifier
Incapacitating injury
No injury
…
Model Patterns
Chong, Miao, etc at al. "Traffic accident data mining using machine learning paradigms."
Fourth International Conference on Intelligent Systems Design and Applications (ISDA'04)
Method
 Decision Tree
 Neural Network
 DT & NN

8. Related works
8
 Clustering Approach
Dataset Clustering
C1
C2
C3
C7
…
Pattern1
Pattern2
Pattern3
Pattern7
…
Depaire, Benoît, Geert Wets, and Koen Vanhoof. "Traffic accident segmentation by means
of latent class clustering." Accident Analysis & Prevention 40.4 (2008): 1257-1266.

9. Frequent pattern mining
9
• A method of finding frequent items from the dataset
• It can count item combination efficiently
support 𝑋 =
count 𝑋
𝑀
X : Patterns
M : Amount of transactions
 specify minsup parameter
support(X )
Patterns which are under
minsup will be eliminated
We expect this could easy to extract patterns!!

10. Frequent pattern mining vs statistical analysis
Frequent
pattern mining
Statistical
analysis
strictness △ ○
Analysis cost ○ △
Finding
knowledge ○ ×
10

11. Purpose of this study
Adapt frequent pattern mining to traffic accident data
- Is it efficient to analyze?
- Can it get knowledge?
11
Data & Analysis
- Traffic Accident
Data
45,653 datas
- Algorithm
Apriori
- Programming
Python

12. 3 type analysis with pattern mining
② Vehicle type
12
① Intersection ③ Age
① ー ② ー ③
⇒ Time rules ⇒ Emerging
Patterns, Similarity
⇒ Correlation

13. ① Time rules in each intersection
13① ー ② ー ③
 Intersection analysis
Name of intersection is
labeled by police
• Occur time prediction
• Causes due to time
 Association rule
support 𝑋 ⇒ 𝑌 =
count(𝑋 ∪ 𝑌)
𝑀
rule ∶ 𝑋 ⇒ Y Sunny ⇒ Road is drying

14. Flow of the analysis
14① ー ② ー ③
Dataset
Interse
ction 1
Interse
ction 2
Interse
ction N
Patter
ns 1
Patter
ns N
All rules
Patter
ns 2
Rules
1
Rules
2
Rules
N
…
…
…
 Time span
1. Long span
(Season, Weekday or
Holiday, time per 6
hours)
2. Short span
(Season, Days of the
week, time per 3 hours)

15. Result of rules (Long span)
rule intersection support
{Weekday} ⇒ {12:00-17:00} 下村 0.588
{Autumn} ⇒ {Weekday} 水上交差点 0.583
{Weekday} ⇒ {12:00-17:00} 県庁前交差点 0.571
{Summer} ⇒ {Weekday} 愛敬町 0.545
{Winter} ⇒ {Weekday} 満穴 0.529
{Summer} ⇒ {Weekday} 唐房入口交差点 0.500
{Summer} ⇒ {Weekday} 伊勢町 0.500
{Weekday} ⇒ {12:00-17:00} ハローワーク唐
津入口
0.500
{Summer} ⇒ {Weekday} 栄町北 0.474
{Autumn} ⇒ {Weekday} 村徳永 0.462
15① ー ② ー ③

16. Result of rules (Long span)
rule intersection support
{Winter} ⇒ {Holiday} 幡崎東 0.455
{Holiday} ⇒ {12:00-17:00} 枯木の塔 0.364
{Holiday} ⇒ {12:00-17:00} 長瀬交差点（信号） 0.364
{Autumn} ⇒ {Holiday} 神埼市役所前 0.357
{Holiday} ⇒ {6:00-11:00} 五条（北） 0.333
16① ー ② ー ③

17. Result of rules
rule intersection support
{Spring} ⇒ {Friday} 千布北交差点（信号） 0.364
{Saturday} ⇒ {12:00-14:00} 鏡山入口交差点 0.364
{Winter} ⇒ {Saturday} 幡崎東 0.364
{Winter} ⇒ {Monday} 龍谷短大入口 0.350
{Winter} ⇒ {Sunday} 脇田 0.333
{Summer} ⇒ {12:00-14:00} 浜玉中学校前交差点 0.313
{Spring} ⇒ {18:00-21:00} 中副交差点（Ｒ３８５） 0.308
{Monday} ⇒ {12:00-14:00} 東多久駅前 0.286
{Winter} ⇒ {15:00-17:00} 県庁前交差点 0.286
{Summer} ⇒ {Saturday} 材木町浦島通り 0.273
17① ー ② ー ③

18. ② Vehicle type analysis
18① ー ② ー ③
 Road structure
Extract road structure
patterns by comparing
vehicle type
Determine degree of
similarity with
comparing patterns
Car Walker Bicycle Car Auto bicycle
perpetrator victim

19. Flow of the analysis
19① ー ② ー ③
Dataset
Walker Bicycle
Auto
bicycle
Patter
ns 1
Patter
ns 4
All patterns
Patter
ns 2
Car
Patter
ns 3
Similarity
Emerging
patterns

20. How to determine emerging patterns
20① ー ② ー ③
Walker
Bicycle
Car
Auto bicycle
 Growth rate
GRG 𝑒 =
supportG 𝑒
1
𝑁 𝑖=0
𝑁
supportGi
𝑒
At pattern e
GRG 𝑒 ≥ 1.0
⇩
Emerging pattern

21. Growth rate of each vehicle type
21① ー ② ー ③

22. Elements of the emerging pattern
22
Walker
Bicycle
Car
Auto bicycle
 No traffic light
 Median divider - paint
 Inside of intersection
 Single road・1 lane road
 Near intersection
 Exist pedestrian-vehicle separation
 Median divider - paint
 General traffic location
 No pedestrian-vehicle separation
 No traffic light  Intersection
 No median divider
① ー ② ー ③

23. Degree of similarity between vehicle type
Walker Bicycle Car
Auto
bicycle
Walker - 0.129 0.129 0.112
Bicycle 0.135 - 0.137 0.056
Car 0.129 0.137 - 0.087
Auto
bicycle
0.112 0.056 0.087 -
23① ー ② ー ③
Si G1, G2 =
1
𝑁
𝑖=0
𝑁
𝑠𝑢𝑝𝑝𝑜𝑟𝑡 𝐺1
𝑒𝑖 − 𝑠𝑢𝑝𝑝𝑜𝑟𝑡 𝐺2
𝑒𝑖
Si → 0.0 : Similar
Si → 1.0 : not Similar

24. ③ Correlation between age and support
24① ー ② ー ③
 Correlation patterns
Many patterns
↓
Some patterns have a
correlation with age?
0
200
400
600
800
1000
1200
1400
1600
0 10 20 30 40 50 60 70 80 90
Amountofperpetrators
Age

25. Flow of the analysis
25① ー ② ー ③
Dataset
Age 18 Age 19 Age 80
Patter
ns 1
Patter
ns 62
All patterns
Patter
ns 2
Correlation
patterns
…
…

26. Positive correlation patterns
26① ー ② ー ③ Before noon : 9:00-11:00

27. Negative correlation patterns
27① ー ② ー ③ Beginning of night : 18:00-20:00

28. Conclusions
We tried to adapt frequent pattern mining to traffic
accident data
We did 3 type analysis
• Time rules in intersections
→ Some rules that look related to time
• Vehicle type and road structure
→ Emerging patterns and determine similarity
• Frequent patterns which have correlation with age
→ Extract some patterns with the correlation
28

29. Future works
• Validation of knowledge has not done
• Traffic accident data with location can add more
metadata
• We’d like to introduce this method to real case
29
Dataset
Frequent
pattern
mining
Knowledge ?