Team: Driving It - contact: Janto Oellrich (joellrich@uos.de)

Team members: Scott Hartshorn (Scott H), Janto Oellrich (TheKeymaker), Andrei Varanovich (Dr. Gigabit)

This repository contains our 2nd prize winning solution to AXA Driver Telematics Analysis competition hosted on kaggle.com.

The goal of this challenge was to infer from unlabeled GPS data whether a trip was driven by a given driver or not. The dataset provided by AXA comprised around half a million trips by roughly 2700 drivers.

We tackled the problem by combining the results of a trip matching algorithm and driver signature models. The objective was to identify frequently taken trips as they are likely to be trips from the respective driver and apply supervised models to telematic features extracted from GPS data to produce probabilistic output. We built ensembles of different models in both streams and one factor that was probably responsible for our success was the extensive combination of these solutions.

Trip matching was based on an ensemble of runs of the Douglas-Peucker algorithm applied at different sensitivities. Our final telematics model consisted of an ensemble of 6 different models. We mainly used Random Forests, in combination with Gradient Boosting and Logistic Regression. We trained models on different feature sets and combinations of those feature sets. For each feature set we selected the algorithm (out of our candidates) that yielded the best result on the LB as we expected the LB to be quite informative in this competition. The weights for the ensemble we choose arbitrarily informed by the performance of the individual models.

These are the models we used for the winning solution:

1. Gradient Boosting Machine (feature set 3)
2. Random Forest (feature set 3)
3. An ensemble of Random Forest and Logistic Regression (feature set 1)
4. Random Forest (feature sets 1 and 3 combined)
5. GBM in R (feature set 3)
6. Random Forest (feature set 2)

The final telematics model did pretty well and scored 0.91375 (Private score). Combining this with our trip matching results increased the final AUC score to 0.97398 which ensured us 2nd place.

In order to replicate the winning solution, just follow the few steps below.

If you do not have the raw GPS data, get it here

http://www.kaggle.com/c/axa-driver-telematics-analysis/data

Change the "DATA" path in the file SETTINGS.json to point to your data folder, and the "ROOT" path to match the parent folder of the cloned repository, e.g.

 {
    "DATA": "C:/users/user/FinalSubmission/drivers",
    "ROOT": "C:/users/user/FinalSubmission/"
 }

For the telematics part, we first extracted the 3 feature sets and then trained our various models on them.

To extract the features, go to the Features folder and run

  extract_features

This will create the three feature matrices featurematrix1.csv, featurematrix2.csv and featurematrix3.csv. You can also create the matrices individually by running (e.g. for feature set 3)

  python generate_featureSet3.py

In order to generate feature set 1, the raw GPS files need to be converted to .npy files which makes training a lot faster. To do this, execute python csvToNpy.py which will write .npycopies of the .csv files to the data folders.

All our models can be run by

 train

The 6 individual models are written to the Models folder and are in a submittable format. You can also run models individually by executing (e.g. for a GBM on feature set 2)

  python train_featureSet2.py

The file train_GBM_R.R has to be run independently (change the marked path inside the file before running).

To run our trip matching routine cd to the RDP_Trip_Matching and the RDP_Trip_Matching_Old folder and execute

Run_Trip_Matching

Detailed instructions can be found in the folder's README.md

To ensemble the trip matching results and the telematics models cd to the Create_Ensemble folder and execute

generate_Solution