Table of contents
Introduction: Business Problem
Reducing traffic accidents is an important public safety challenge. Accident prediction is important for optimizing public transportation, enabling safer routes, and cost-effectively improving the transportation infrastructure, all in order to make the roads safer. This project is focusing on predicting the accident severity based on different attributes like locations, weather, road and visibility conditions, cause of different vehicles, collision due to inattention, rough driving by the influence of drugs or alcohol and so on. In an effort to avoid and reduce the frequency of these types of accidents, we will build a model to predict the severity of an accident given the weather, the road, the light conditions and whether or not a driver involved was under the influence of drugs or alcohol. This way we would be able to bring awareness to the drivers and warn people about the possibility of getting into a car accident and its severity if it happens. This way people would drive more carefully or even change the travel if able to.
We will use our data science powers to be answered how severe would be the accident if it happens by knowing the weather, road and visibility conditions.
Data
We have collected the raw data from SDOT Traffic Management Division, Traffic Records Group and contains data of all types of collisions that happened in Seattle city from 2004 to May/2020.
The data contains 194,673 samples and have 37 features. Since we would like to identify the factors that cause the accident and the level of severity, we will use SEVERITYCODE as our dependent variable Y, and try different combinations of independent variables X to get the result. Since the observations are quite large, we may need to filter out the missing value and delete the unrelated columns first. Then we can select the factor which may have more impact on the accidents, such as weather, road condition, light condition, collision type and either driver was involved by influencing drugs or alcohol.
Missing Values
There are missing values on part of the data; some features have over 40% of missing data for that we will not consider them to our model. Removing the irrelevant data attributes away, the variables we will use to classify the severity of the accidents are:
- COLLISIONTYPE: Collision type
- WEATHER: Weather conditions during the time of the collision
- ROADCOND: The condition of the road during the collision
- LIGHTCOND: The light conditions during the collision
- UNDERINFL: Whether or not a driver involved was under the influence of drugs or alcohol
These features contain missing values but it’s below 3% of the total amount of samples.
Target Variable
Our target variable SEVERITYCODE that corresponds to the severity of the collision:
- 1. Property Damage Only Collision which is the same as not injured collision
- 2. Injury Collision which is a binary classification problem
Annual amount of traffic incidents in Seattle
We have noticed that there is a considerably high amount of incidents only discrepancy is from 2020 as it was recorded incidents that occurred till May 2020. We can also infer from the plots that no injury collisions are always more likely to happen.
Collision Type
There is a considerable difference in the collision occurrences according to collision type. The most three accidents were with parked cars, angles and rear-ended. Other types of collisions are also a matter of concern.
Weather Condition
Considering Seattle weather conditions, we notice most incidents happened in a Clear weather. That could be because drivers are less careful when there is no harsh weather condition. It would be interesting to check the correlation between WEATHER and INATTENTIONIND (whether or not collision was due to inattention), but there are too many missing values, 85% of the data is missing.
Road Condition
Considering the road condition we have found that there were more occurrences happened in dry road condition.
Light Condition
Most of the accidents happened in Daylight whereas Dark-Streer Lights On is also considerable for the accident.
Diver Under Influence of Drugs or Alcohol
It has been seen that drivers were not under any influence in most incidents.
Each feature have a different weight of influence on the severity of the collision. Overall, all of them are consistently infering that no-injury accidents in normal driving conditions are more recurrent.
We will use COLLISIONTYPE, WEATHER, ROADCOND, LIGHTCOND and UNDERINFL as attributes to classify SEVERITYCODE. For that, we will need to prepare these features so it is suitable for a binary classification model. We will use some popular machine learning algorithms like SVM, Logistic Regression, Naive Bayes and KNN for build-up models to analyze their performance and predict the collision severity.
Methodology
In this project we will use COLLISIONTYPE, WEATHER, ROADCOND, LIGHTCOND and UNDERINFL as attributes to classify SEVERITYCODE. We will limit our analysis to this four independent variables. For that we will need to prepare this features so it is suitable for a binary classification model.
In the first step we will prepare and clean the dataset to make it readable and suitable for the machine learning algorithms. There are 37 attributes whereas we will consider the features of COLLISIONTYPE, WEATHER, ROADCOND, LIGHTCOND and UNDERINFL to classify SEVERITYCODE. Those attributes has 3% missing data which we will drop them. We will split this dataset as train and test split whereas 70% to train the model and 30% to test the model.
Second step in our analysis will be calculation and exploration of different models to find out the main problem for severity. We will use 3 classification models which are Logistic Regression, Decision Tree an KNN. After obtaining each model's predictions we will evaluate their accuracy, precison, f1-score, log-loss and compare and discuss the results.
Analysis
1. Data Preparation and Cleaning
Data preparation and cleaning is required to make the dataset readable and suitable to the machine learning algorithms.
As we have choose 5 attributes to classify SEVERITYCODE so that we have dropped other attributes out of 37 attributes and also dropped the 3% missing data from those attributes. We have converted the categorical values to numerical values by using label encoding technique.
Then we separated the independent variables to a dataset and dependent variable 'SEVERITYCODE' to another dataset. After that use this data to randomly pick samples and split in below ratio:
- 70% to train model
- 30% to test model
Classification: Modeling and Evaluation
The prepared was used to model three classification models.
KNN
Classifies unseen data through the majority of its ‘neighbours’. In this case we already know K=2 (2 classes of SEVERITY CODES). After obtaining each model’s predictions we will evaluate their accuracy, precison, f1-score, log-loss and compare and discuss the results.
Decision Tree
Classifies by breaking down a dataset into smaller and smaller subsets while at the same time an associated decision tree is incrementally developed.
Logistic Regression
Classifies data by estimating the probability of classes.
Model Evaluation Using Test Set
Among all three models, Jaccard score's measures accuracy is above 70%. The highest accuracy model is the Decision Tree Classifier. The same model also presents the best F1_score and Recall(True positive rate).
Results and Discussion
In this analysis we evaluated the performance of 3 machine learning algorithms on the Seattle Collision dataset to predict the severity of an accident knowing the weather and road conditions.
The three models performed very similary, but Decision Tree stood out with a difference from KNN and Logistic Regression during the evaluation with the model's accuracy.
Conclusion
Purpose of this project was to analyze the relationship between severity of an accident and some characteristics which describe the situation that involved the accident. We picked 5 features out of 37 where it showed to be a reasonable choice to find the answer that were searcing for. It was able to achieve 77% accuracy however there were still significant variances that could not be predicted by the models in this study. Those 5 features have somewhat impact on whether or not travel could result in property damage (class 1) or injury (class 2).
Top comments (0)