Applying machine learning for generating synthetic geospatial trajectories

The goal of the project is to explore the effectiveness of Recurrent Neural Networks (RNN) and Long Short term Memory (LSTM) to generate synthetic but realistic mobility traffic. RNN’s have been proved successful in various domains such as automatic caption generation for images, constructing new levels in gaming and so on. In this project, you will design and build a synthetic traffic generator that uses machine learning for extracting the behavioural patterns of users from real datasets. The learnt model will be later used to create new and larger datasets, characterised by features that resemble true properties of users from an actual dataset.

Design

We will provide you with real datasets, consisting of mobility traces of more than 100 users. The goal will be to extract and learn existing behaviours, in a way that facilitates generation of new realistic behaviours. You will try out possibilities with different machine learning architectures and models and come up with an ideal design for such a synthetic generator.

Implementation

We will provide you with a base implementation based on Tensorflow and Keras in Python. You will be free to select the tools and libraries of your choice. We will assist you with tutorials with several machine learning libraries such as tensorflow, pybrain etc.

Evaluation

The validation of the newly generated trajectories will be made by comparing their similarity with actual mobility traces. You will select a set of metrics to evaluated the performance of the model.

Takeaway from this project

Completing this project will allow you to gain some practical experience of working with geospatial data and more importantly hands on experience with machine learning tools widely used in many popular applications today. You will be provided with tutorials, documentation, datasets, a friendly working atmosphere and a nice coffee machine 😉

Contact

Vaibhav Kulkarni (firstname.lastname at unil dot ch)