What Will You Do for the Rest of the Day?: An Approach to Continuous Trajectory Prediction

Understanding and predicting human mobility is vital to a large number of applications, ranging from recommendations to safety and urban service planning. In some travel applications, the ability to accurately predict the user's future trajectory is vital for delivering high quality of service. The accurate prediction of detailed trajectories would empower location-based service providers with the ability to deliver more precise recommendations to users. Existing work on human mobility prediction has mainly focused on the prediction of the next location (or the set of locations) visited by the user, rather than on the prediction of the continuous trajectory (sequences of further locations and the corresponding arrival and departure times). Furthermore, existing approaches often return predicted locations as regions with coarse granularity rather than geographical coordinates, which limits the practicality of the prediction.

In this paper, we introduce a novel trajectory prediction problem: given historical data and a user's initial trajectory in the morning, can we predict the user's full trajectory later in the day (e.g. the afternoon trajectory)? The predicted continuous trajectory includes the sequence of future locations, the stay times, and the departure times. We first conduct a comprehensive analysis about the relationship between morning trajectories and the corresponding afternoon trajectories, and found there is a positive correlation between them. Our proposed method combines similarity metrics over the extracted temporal sequences of locations to estimate similar informative segments across user trajectories.

Our evaluation shows results on both labeled and geographical trajectories with a prediction error reduced by 10-35% in comparison to the baselines. This improvement has the potential to enable precise location services, raising usefulness to users to unprecedented levels. We also present empirical evaluations with Markov model and Long Short Term Memory (LSTM), a state-of-the-art Recurrent Neural Network model. Our proposed method is shown to be more effective when smaller number of samples are used and is exponentially more efficient than LSTM.