There are several motor accidents that happen every day around the globe. The analysis of these motor accidents can help provide critical actionable insights to facilitate informed decisions and drive business outcomes (such as reducing accidents, saving lives etc.) and business applications to reduce costs (Motor Vehicle Repair Claims, Personal Injury Claims etc).

In this project, we aim to answer the following questions :

1. What are the most statistically significant variables , that help us to predict the severity of accidents accurately?
2. Can we evaluate different statastical prediction models and choose the best one depending on the performance?
3. Can we warn the user of high risk accident areas on the route chosen by user in real time?

This project uses the following databases :-

1. Accidents Database : Obtained from govt of UK website https://data.gov.uk/dataset/cb7ae6f0-4be6-4935-9277-47e5ce24a11f/road-safety-data It includes all the details about the accident including the location of the accident, Road/Weather conditions at the time of accident. For the purpose of this project data for 2 years - 2015 and 2016 has been considered.
2. Vehicle Database : This data is also obtained from the above mentioned source and includes details regarding the vehicles involved in the accidents.
3. Traffic Flow Data : This data describes the annual average volume of traffic flow at the location of the accident and has been obtained from https://data.gov.uk/dataset/9562c512-4a0b-45ee-b6ad-afc0f99b841f/highways-england-network-journey-time-and-traffic-flow-data

This analysis part of this Project is done in R, modelling in Python and front end in Rshiny.

After importing the tables thorough data cleaning, data is preprocessed. Some fields are changed so that they become consistent across all the tables. Redundant levels of some categorical columns are removed and datatypes of all the fields are fixed as they should be.

Some feathures that don't seem to be contributing much for the final analysis model are removed after comparing the correlation values and Chi squared vales.

Hour and Month fields are cyclic features. They are encoded using sine and cosine transformations. Road name is another important feature for the predictive model and has more that 5000 different categories, thus making it unsuitable for most of the predictive algorithms. This feature is encoded using mean encoding.

Different plots for all the features in the 3 tables are plotted and detailed analysis is done in order to see which features contribute most in the predictive modeling, removing the ones that are not so useful.

A plot between the different number of samples available for different accident severity values, reveals that the data is highly skewed with maximum number of samples(85%) belonging to the major class- Mild, 14.85% of samples belonging to the serious class and only a small(0.15%) of data belonging to the minority class. In order to deal with this class imbalance, stratified sampling of the training set is done.

Following algorithms are used for predictive modeling in this project:

1. Naive Bayes
2. Multinomial Regression
3. Random Forest
4. Adaboost
5. Training the above models on sampled data
6. Training the above models on cost based learning
7. Ensemble of the above algorithms

All the models above were developed in Python and trained on floyd hub cloud for faster convergence. Weighted Random Forest emerged as a winner in terms of the f2-measure. It provided best f2 measure and the best recall for the class 1.

Front end is developed in Rshiny and it displays the percentage of accident probability on the route chosen by the user and claculates and plots high risk areas along the route. A brief overview of the front end could be find here :- https://www.youtube.com/watch?v=p1gzX0CRXrY&t=57s