Intention-Aware Out-of-Town Mobility Prediction for Social Network Users

Predicting social network users' out-of-town mobility based on their generated spatio- temporal data has become an impending demand for urban collaborative management. User out-of-town mobility is indeed a “long tail,, event compared with user mobility inside a city, resulting in extremely sparse check-in data generated by users in the out-of-town area. It is difficult for existing studies to utilize the limited cross-city check-in data to model users' out-of-town mobility preference, and afterwards accurately predict users' mobility outside the city. Toward these issues, in this article, a novel intention-aware framework called TIEMPO (short for Intention- award Mobility Preference modeling) for modeling user out-of-town mobility preference is proposed and implemented. Firstly, in order to alleviate the data sparsity problem, abundant trajectories are sampled from the constructed out-of-town location network via random walk, based on which a specific number of user intentions for out-of-town mobility can be discovered through unsupervised clustering. Secondly, the memory network is introduced to refine user intentions from similar users' out-of-town trajectories. Thirdly, following the idea of transfer learning, user check-ins inside a city and user intentions outside a city are interactively modeled to enhance user's out-of-town mobility preference representation. Finally, the probability of a user visiting an out-of-town location is quantified by integrating the user's out-of-town mobility preference representation and the location hidden representation. Extensive experiments based on multiple cross-city check-in datasets are conducted, and empirical results indicate that the proposed TIEMPO framework can effectively predict users' out-of-town mobility in terms of the visited locations, where the prediction accuracy metric Acc@ 10 shows a significant advantage of 12 % -15 % and the ranking reliability metric NDCG@10 achieves 3 % - 5 % advantage compared with the baseline models. Even in cold-start prediction scenarios, TIEMPO framework still has the best performance. © 2024 Science Press. All rights reserved.