Submission For "AI For SEA: Safety"

This is a machine learning model that can detect dangerous driving using telematics data collected during the trip. More details about this project can be found at https://www.aiforsea.com/safety.

About The Repository

1. Complete documentation of this project: documentation.ipynb
2. Python library to predict dangerous driving: predict_model.py
3. Pre-trained model weights directory: model_weights/
4. Demonstratation of python library: demo.ipynb
5. Reusable scripts (Model searching tools): utils.ipynb

Highlights

1. A total of 52 features are generated from raw telematics data using three approaches:
   

• Approach 1: Statistical summary of telematics data
  
• Approach 2: Count outlying driving behaviours based on telematics readings
  
• Approach 3: Sliding windows aggregated features

2. Six blended (stacked) models are used in the prediction pipeline, which consists of 3 layers:
   

• Layer 1: 2 weak learners
  
• Layer 2: 3 strong learners
  
• Layer 3: 1 meta learner

3. The model performance on self-define hold-out test dataset achieved:

• ROC Score: 0.7513
• Accuracy Score: 0.7853

NOTE: Please refer to documentation.ipynb for more details about this project.

Notes To Evaluators

(Tested on Python Version: 3.7)

First, clone the repository and install the dependencies.

git clone https://github.com/kfengtee/grab-aiforsea-safety.git
cd grab-aiforsea-safety
pip install -r requirements.txt

Then, upload your hold-out test set (raw telematics data and labels) to this repository.

To use the library, you can do the following:

import predict_model

classifier = predict_model.DangerousDrivingClassifier('model_weights') # load the pre-trained weights

# replace "dir_to_raw_telematics_data" with actual hold-out test data directory
output = classifier.predict("dir_to_raw_telematics_data") # output: DataFrame, columns = ['bookingID', 'prob', 'label']

Please look at demo.ipynb for more demonstration details.

License

MIT