ARNet

This is a tensorflow implementation of Autoencoder Regularized Network For Driving Style Representation Learning

Requirements

1. Python 2.7
2. Tensorflow 1.3.0
3. Cuda 8.0.61

Data

A sample data file for 5 drivers, with 5 trajectories for each, is shared in 'data' folder. The data file has following columns: Driver, ID, Time, Lat, and Lon.

Experiments

1. Statistical Feature Matrix: In order to create the statistical feature matrix as described in the paper, you need to run 'IBM16_FeatureMatrix.py' which creates two files in data folder.
2. Train ARNet Model: In order to train an ARNet model, you need to run script 'IBM17-ARNet.py'. The input for this script is the statistical feature matrix and output is a trained model.
3. Clustering: In order to build representation for test trajectories and perform the clustering, you can use 'IBM17-Clustering.py'. The input for this script is the trained ARNet model and a set of test trajectories as described in the paper. The output will be the clustering results and also the created emebddings for each trajectory.

Results

Our best results for driver clustering task based on a set of real-world, private, and non-anonymized (based on gps coordinates) datasets are as follows:

#Drivers	#Trajectories/Driver	#Driver_Avg Of Error	#Driver_Std Of Error	AMI Avg	AMI Std
1	200	0.33	0.55	0.71	0.45
2	200	0.28	0.45	0.2	0.22
3	200	0.29	0.45	0.11	0.06
4	200	0.25	0.52	0.13	0.06
5	200	0.44	0.5	0.12	0.05
6	200	0.3	0.46	0.13	0.05
7	200	0.71	0.61	0.12	0.04
8	200	1.08	0.49	0.1	0.02
9	200	1.0	0.74	0.12	0.04
10	200	1.6	0.63	0.12	0.04
	Average	0.63	0.54	0.19	0.1

Note that here we used 25 sets of trajectories with D drivers for each of the rows in above table. We used Affinity Propagation clustering algorithm as suggested by the paper with preference parameter -65 and damping factor 0.5. The results in above table can be compared with the results in Tables 1 and 2 in the ARNet paper.

References

1. Autoencoder Regularized Network For Driving Style Representation Learning, 2017