Deep Learning Approaches for Mobile Trajectory Prediction

Abstract

Mobility analytics is a very critical research topic related closely with the wide field of human behavior analysis. Mobility trajectory prediction refers to the model-based projection of an individual’s location in the foreseeable future, within the boundaries of a predefined area. This information is proved to be an important input for information systems in multiple ICT-related scientific fields. Various algorithms have been proposed so far for solving this problem, however, traditional predictive approaches are shown to underperform in accuracy and reliability. This leaves room for more sophisticated, deep-learning modeling formulations. Framed within this statement, the focus in this paper is placed on contemporary deep learning approaches for trajectory prediction and more specifically, sequential machine learning models such as the Social GAN, Trajectory Transformer, and the proposed scheme, which is based on a customized GRU neural network extended with the attention mechanism. The three approaches are theoretically described and most importantly, they are implemented, validated, and evaluated on top of a realistic experimental platform and based on both simulated mobility data and open-source mobility datasets. The evaluation process of the models takes into consideration ml regression error metrics, qualitative 2D projection but also system aspects such as training complexity in the form of hyperparameter order. The results indicate that our proposed scheme outweighs SotA deep learning approaches by 29% in terms of Mean Displacement Error and by a factor of thousands in computation complexity, making it a realistic candidate for latency-sensitive applications, placed at edge computing nodes with limited processing capabilities.