Self-supervised human mobility learning for next location prediction and trajectory classification

Massive digital mobility data are accumulated nowadays due to the proliferation of location-based service (LBS), which provides the opportunity of learning knowledge from human traces that can benefit a range of business and management applications, such as location recommendation, anomaly trajectory detection, crime discrimination, and epidemic tracing. However, human mobility data is usually sporadically updated since people may not frequently access mobile apps or publish the geotagged contents. Consequently, distilling meaningful supervised signals from sparse and noisy human mobility is the main challenge of existing models. This work presents a Self-supervised Mobility Learning (SML) framework to encode human mobility semantics and facilitate the downstream location-based tasks. SML is designed for modeling sparse and noisy human mobility trajectories, focusing on leveraging rich spatio-temporal contexts and augmented traces to improve the trajectory representations. It provides a principled way to characterize the inherent movement correlations while tackling the implicit feedback and weak supervision problems in existing model-based approaches. Besides, contrastive instance discrimination is first introduced for spatio-temporal data training by explicitly distinguishing the real user check-ins from the negative samples that tend to be wrongly predicted. Extensive experiments on two practical applications, i.e., location prediction and trajectory classification, demonstrate that our method can significantly improve the location-based services over the state-of-the-art baselines. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

This study introduces a Self-supervised Mobility Learning framework for sparse and noisy human mobility data, focusing on enhancing trajectory representations with rich spatio-temporal contexts and augmented traces. Contrastive instance discrimination is first introduced to improve model training accuracy by distinguishing real user check-ins from negative samples.