Insighters Hack the Crash! at HackUPC 2019 for McKinsey & Company

We developed a Machine Learning pipeline to predict accident severity and a ChatBot that provides bite-sized advice to drivers!

Winners at at HackUPC 2019 for McKinsey & Company's challenge in Barcelona

The submission for the predictions is here

Motivation

Rate of road-accidents and fatality in the UK is very high

This challenges focuses on predicting accident severity based on the following parameters:

Predicting accident severity will help the authorities closely monitor accident prone road segments and do a root cause analysis for such mishaps

Accident training data	Vehicle data
137,599	455,652

Exploratory Data Analysis: EDA by looking at all the fields and their values
- Accidents
- Vehicles
- Test
Data Wrangling:
- Data Clearning & Conversion
Data Summarization:
- Correlations & Scatter plots
- Data Distributions
Feature Engineering: handcraft a bunch of features from the fields
Feature Selection & Dimensionality Reduction: PCA, tSNE, UMAP
Split the data: split the dataset into Train and Validation sets
Modelling a classification model:
- Model a Random Forest classifier and analyse the power of their features
- Transform the text of the information regarding the accident and the vehicles involved using NLP techniques such as TF-IDF & Embeddings
- Ensembling our individual models
Ablation studies: fitting a Linear Regression model and running ablation studies to measure the variance explained by the features
Mutual Information Exploration: measuring the mutual information (mutual dependence between two variables) of each of the features we figured as important (based on the ablation study) and the precision mark.
Hyper-parameter tuning: tune the model
Train the final model: train the final model and build the submission file