Taming the Long Tail in Human Mobility Prediction

TL;DR

LoTNext tackles the long-tail challenge in next POI prediction by integrating a Long-Tail Graph Adjustment to prune sparse, high-noise edges and a Long-Tailed Loss Adjustment to reweight learning toward tail POIs, complemented by an auxiliary time-prediction task. The approach combines a Transformer-based trajectory encoder with a Spatial Contextual Attention layer and denoised POI embeddings derived from a denoised User-POI interaction graph and a Global Transition graph. Empirical results on Gowalla and Foursquare show consistent improvements over 10 baselines, with ablations confirming the contributions from LTGA, LTLA, and the auxiliary task. This work promises practical gains in personalized navigation and urban mobility applications by improving prediction for less-visited POIs without requiring extra data sources, while also highlighting privacy considerations for trajectory data.

Abstract

With the popularity of location-based services, human mobility prediction plays a key role in enhancing personalized navigation, optimizing recommendation systems, and facilitating urban mobility and planning. This involves predicting a user's next POI (point-of-interest) visit using their past visit history. However, the uneven distribution of visitations over time and space, namely the long-tail problem in spatial distribution, makes it difficult for AI models to predict those POIs that are less visited by humans. In light of this issue, we propose the Long-Tail Adjusted Next POI Prediction (LoTNext) framework for mobility prediction, combining a Long-Tailed Graph Adjustment module to reduce the impact of the long-tailed nodes in the user-POI interaction graph and a novel Long-Tailed Loss Adjustment module to adjust loss by logit score and sample weight adjustment strategy. Also, we employ the auxiliary prediction task to enhance generalization and accuracy. Our experiments with two real-world trajectory datasets demonstrate that LoTNext significantly surpasses existing state-of-the-art works.

Figures (6)

• Figure 1: The long-tailed distribution for POI check-in frequency from the Gowalla dataset.
• Figure 2: The Architecture of Long-Tail Adjusted Network for Next POI Prediction (LoTNext).
• Figure 3: The performance comparison of the long-tailed and head POIs between LoTNext and Graph-Flashback on Gowalla dataset.
• Figure 4: The visualization of tail POIs on Gowalla dataset. The color represents the POI frequency.
• Figure 5: Sample prediction from Gowalla dataset with Graph-Flashback and LoTNext.

Theorems & Definitions (1)

• Definition 1: User Next POI Prediction