Dual-grained human mobility learning for location-aware trip recommendation with spatial–temporal graph knowledge fusion

Abstract

Trip recommendation is a popular and significant location-aware service that can help visitors make more accurate travel plans. Its principal purpose is to provide a sequence of points of interest (POIs) for a visitor who has a specific demand, such as the number of attractions, start location, end location, etc. Previous studies either regard the trip recommendation task as a simplistic orienteering problem or endeavor to maximize the users’ visiting preferences and transitional regularities in a multi-round iteration manner. However, significant contexts such as long-term transitional dependencies and spatial–temporal correlations are under-explored. Although the emerging deep recursive models (e.g., recurrent neural networks) enable us to relieve the above limitations, there still exist three major challenges: (1) most conventional offline-homogeneous knowledge distillation approaches for key entities (e.g, POI and visiting time) could lead to the inadequacy of capturing the inherent heterogeneous interactions among the different entities and even result in the deviation of understanding human real individual preferences; (2) data sparsity in transitional regularity learning impedes the ability to comprehend human diverse mobility patterns; and (3) the lack of considering contextual facts such as attractive bias would degrade the generalization ability of the model. To remedy the above issues, this work presents a novel framework based on spatial–temporal graph representation learning, namely GraphTrip. Specifically, we first introduce a location-aware information fusion after building the spatial–temporal graph (ST-Graph). Then, a dual-grained human mobility learning module is proposed to address the sparsity of periodic regularity. In addition, we fuse the explicit information (e.g., POI popularity) behind the trip data as prior knowledge to facilitate the performance of trip inference. In the end, the experimental results conducted on five real-world trip datasets demonstrate our proposed GraphTrip achieves promising gains against several cutting-edge baselines, e.g., up to 5.75% and 4.45% improvements on Toronto regarding F1 and pairs-F1.