NextRoute: a lossless model for accurate mobility prediction

Abstract

Mobility prediction of vehicles has become a key feature in mobility management in smart cities, mobile computing environments, intelligent transportation systems, vehicular networks, and location-based services. It has several important applications in traffic congestion forecasting, location-based routing protocol designs and targeting advertisements generation, etc. Mobility prediction (aka route prediction) consists of forecasting future routes to be traversed by a vehicle. Several models have been proposed for route prediction. These models are generally probabilistic models (e.g., Markov models) or data mining-based models (e.g., sequential rule-based models), and are trained with historical location data. Although these models were shown to perform well, one important drawback is that they are lossy. In other words, a large amount of information found in the location data is discarded by their training processes. Consequently, these models do not perform well for predicting routes where detailed information is required to perform accurate predictions. Besides, several prediction models assume the Markovian hypothesis that the next location of a user only depends on his/her current location and any previous movement of the user is ignored in prediction. This hypothesis has been employed, such as in Lian et al. (ACM Trans Intell Syst Technol 6:1–27, 2015. https://doi.org/10.1145/2629557), as a simplifying assumption because it allows building models having a small size. However, this assumption is often unrealistic and thus greatly decreases the accuracy of route prediction. Moreover, these models are noise sensitive as they do not tolerate the smallest deviation in location data principally prone to several disturbances and uncertainty issues. To address these limitations, this paper introduces a novel route prediction model named NextRoute. The proposed model is lossless as it compresses location data in a prediction tree without information loss, and it is designed to use all the relevant information contained in the training data to perform prediction. In contrast to some other proposals, NextRoute provides efficient noise tolerance strategy that loosened the similarity measure when matching the current trajectory of vehicle with historical data. An extensive experimental evaluation was conducted with real-world and synthetic datasets providing quite encouraging results.