Correcting Defective Trajectories using Conditional GAN

Abstract

The end-user mobility patterns play a key role in the process of 5G network design. During tracking end-users via GPS, errors can appear. While we are still constrained with a very limited number of commercially available trajectory datasets, a possible solution is to extend an existing dataset using a generative adversarial network (GAN). Our previous work showed that the utilization of GAN in a form of artificial trajectory generator is possible but not flawless as it introduces the issue with unreasonable gaps between trajectory’s consecutive GPS coordinates. To overcome this issue with a generated dataset and with authentic data as well a special type of GAN called conditional GAN can be used. By leveraging this approach we are not only able to generate a potentially unlimited number of new data samples but as well to correct the existing ones. The number of missing datapoints in the trajectory can go as low as 95% of all points. This artificial intelligence approach has the potential to be used in various use cases where trajectory data are defective and need to be corrected.