Context-Aware Destination and Time-To-Destination Prediction Using Machine learning

Abstract

The rapid adoption of Internet-connected devices (i.e., smart phones, smart cars, etc.) in today's society has given rise to a massive amount of data that can be harnessed by intelligent systems to learn and model the behavior of people. One useful set of such data is movement data, which can readily be obtained via GPS or motion-detection sensors, and which can be used to create models of user movement. One relevant application task based on this type of data is destination prediction, where movement data are used to form highly customized models that can forecast intended user destinations based on partially observed trajectories. In this work, we present a two-stage predictive model for destination prediction and Time-To-Destination (TTD) estimation using movement trajectories and contextual information. Our two-stage approach uses a Transformer-based architecture to predict an intended destination and a regression model to estimate how many steps must be traversed before a destination is reached. We showcase experimental results on various trajectory datasets and show that our proposed approach is able to yield significant destination prediction improvements over previous state-of-the-art methods and can also produce accurate TTD estimates.