Answers are grounded in this paper and cite their sources.
Next-location prediction is commonly studied using long and repetitive mobility histories, whereas tourist trajectories are short and their destination frequencies are strongly imbalanced. We investigate this sparse, long-tailed setting using two tourist-mobility datasets derived from geo-tagged Weibo posts in Suzhou and Beijing. We present a unified neural framework that combines a mixture-of-experts mobility encoder, retrieval-based trajectory context, and two domain-derived spatio-temporal signals, with auxiliary supervision for location category and visiting hour. Across the two datasets, the framework improves Top-1 accuracy and mean reciprocal rank over recurrent, attention-based, graph-based, and recent representation-learning baselines. Short-history analysis in Suzhou and frequency-stratified results in both cities further show higher reported scores when the observed context is limited or the destination is infrequently visited. Component comparisons show the largest aggregate reductions when the spatio-temporal signals or expert encoder are removed, while retrieval provides its clearest contribution on the tail subset. These findings support the utility of integrating behavioural heterogeneity, trip-level context and population-level trajectory evidence when individual observations are limited.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: