Location-based services are gaining popularity; prediction of future destinations and crowd movement patterns are crucial components of these services. This article presents an attention-based neural network regression model designed to forecast future user locations, a critical aspect of location-based services. Leveraging two real-world mobility datasets, New York City check-in data from the Foursquare API and Porto taxi trajectories from Portugal. The model employs an attention-based encoder–decoder neural network to predict user destination coordinates in latitude and longitude. Beyond its predictive capabilities, this research delves into the intricacies of human mobility patterns, contributing to a deeper understanding of movement behavior and shedding light on challenges in current models for mobility prediction. The study explores the impact of various optimization algorithms on model performance, analyzing their effects on accuracy, with the mean haversine distance error serving as the evaluation metric. Notably, the model achieves remarkable results, giving a mean haversine distance error of 1.3336 for the Porto dataset and 1.6379 for the Foursquare NYC dataset when employing the Adam optimizer. We have extended our study by implementing our model on Universal Transverse Mercator coordinate systems. These findings underscore the model’s superiority over previous approaches, offering valuable insights for developing more precise location-based services and advancing mobility and human behavior analysis.

基于位置的服务越来越受欢迎；对未来目的地和人群流动模式的预测是这些服务的重要组成部分。本文提出了一种基于注意力的神经网络回归模型，旨在预测未来用户位置，这是基于位置的服务的一个关键方面。利用两个现实世界的移动数据集：来自 Foursquare API 的纽约市签到数据和来自葡萄牙的波尔图出租车轨迹。该模型采用基于注意力的编码器-解码器神经网络来预测用户目的地的纬度和经度坐标。除了预测能力之外，这项研究还深入研究了人类流动模式的复杂性，有助于更深入地理解运动行为，并揭示当前流动预测模型中的挑战。该研究探讨了各种优化算法对模型性能的影响，分析了它们对精度的影响，以平均半正弦距离误差作为评价指标。值得注意的是，该模型取得了显着的结果，当使用 Adam 优化器时，Porto 数据集的平均半正弦距离误差为 1.3336，Foursquare NYC 数据集的平均半正矢距离误差为 1.6379。我们通过在通用横轴墨卡托坐标系上实施我们的模型来扩展我们的研究。这些发现强调了该模型相对于以前方法的优越性，为开发更精确的基于位置的服务以及推进移动性和人类行为分析提供了宝贵的见解。