LingoTrip: Spatiotemporal context prompt driven large language model for individual trip prediction

Journal of Public Transportation, Journal Year: 2025, Volume and Issue: 27, P. 100117 - 100117

Published: Jan. 1, 2025

Language: Английский

---

Context-aware inverse reinforcement learning for modeling individuals’ daily activity schedules

Engineering Applications of Artificial Intelligence, Journal Year: 2025, Volume and Issue: 146, P. 110279 - 110279

Published: Feb. 17, 2025

Language: Английский

---

Multi-modal contrastive learning of urban space representations from POI data

Computers Environment and Urban Systems, Journal Year: 2025, Volume and Issue: 120, P. 102299 - 102299

Published: April 30, 2025

Language: Английский

---

A causal intervention framework for synthesizing mobility data and evaluating predictive neural networks

Transportation Research Interdisciplinary Perspectives, Journal Year: 2025, Volume and Issue: 31, P. 101398 - 101398

Published: May 1, 2025

Language: Английский

---

diffIRM: A Diffusion-Augmented Invariant Risk Minimization Framework for Spatiotemporal Prediction over Graphs

Transportation Science, Journal Year: 2025, Volume and Issue: unknown

Published: April 8, 2025

Spatiotemporal prediction over graphs (STPG) is challenging because real-world data suffer from the out-of-distribution (OOD) generalization problem, where test follow different distributions training ones. To address this issue, invariant risk minimization (IRM) has emerged as a promising approach for learning representations across environments. However, IRM and its variants are originally designed Euclidean data, such images, may not generalize well to graph-structure spatiotemporal graphs, of spatial correlations in graphs. overcome challenge posed by existing graph OOD methods adhere principles invariance existence (i.e., there exist features that consistently relate label various environments) or environment diversity diversifying environments increases likelihood align with ones). little research combines both STPG problem. A combination two crucial efficiently distinguishing between spurious In study, we fill gap propose diffusion-augmented (diffIRM) framework these Our diffIRM contains processes: (1) augmentation, (2) learning. augmentation process, causal mask generator identifies features, graph-based diffusion model acts an augmentor generate augmented data. penalty using then serves regularizer model. We provide theoretical evidence supporting diffIRM’s ability identify features. The effectiveness further demonstrated through experiments on numerical generated known structural (SCM), our proposed successfully true experiment uses three human mobility sets, is, SafeGraph, PeMS04, PeMS08. outperforms baselines. Furthermore, demonstrates interpretability discerning while making predictions. History: This paper been accepted Transportation Science Special Issue Machine Learning Methods Urban Mobility. Funding: work was supported National Foundation [Grant 2218809].

Language: Английский

---

A dual-layer hypergraph model integrating urban context and courier profiles for route prediction

Transportmetrica A Transport Science, Journal Year: 2025, Volume and Issue: unknown

Published: April 25, 2025

Language: Английский

---

Evaluating geospatial context information for travel mode detection

Journal of Transport Geography, Journal Year: 2023, Volume and Issue: 113, P. 103736 - 103736

Published: Nov. 6, 2023

Detecting travel modes from global navigation satellite system (GNSS) trajectories is essential for understanding individual behavior and a prerequisite achieving sustainable transport systems. While studies have acknowledged the benefits of incorporating geospatial context information into mode detection models, few summarized modeling approaches analyzed significance these features, hindering development an efficient model. Here, we identify representations related work propose analytical pipeline to assess contribution based on random forest model SHapley Additive exPlanation (SHAP) method. Through experiments large-scale GNSS tracking dataset, report that features describing relationships with infrastructure networks, such as distance railway or road network, significantly contribute model's prediction. Moreover, point entities help public travel, but most land-use land-cover barely task. We finally reveal contexts distinct contributions in identifying different modes, providing insights selecting appropriate approaches. The results this study enhance our relationship between movement guide implementation effective models.

Language: Английский

---

HLFSRNN-MIL: A Hybrid Multi-Instance Learning Model for 3D CT Image Classification

Applied Sciences, Journal Year: 2024, Volume and Issue: 14(14), P. 6186 - 6186

Published: July 16, 2024

At present, many diseases are diagnosed by computer tomography (CT) image technology, which affects the health of lives millions people. In process disease confrontation, it is very important for patients to detect in early stage deep learning 3D CT images. The paper offers a hybrid multi-instance model (HLFSRNN-MIL), hybridizes high-low frequency feature fusion (HLFFF) with sequential recurrent neural network (SRNN) classification tasks. Firstly, uses Resnet-50 as feature. main HLFSRNN-MIL lies its ability make full use advantages HLFFF and SRNN methods up their own weakness; i.e., can extract more targeted information avoid problem excessive gradient fluctuation during training, used time-related sequences before classification. experimental study on two public datasets, namely, Cancer Imaging Archive (TCIA) dataset lung cancer China Consortium Chest Image Investigation (CC-CCII) pneumonia. results show that exhibits better performance accuracy. On TCIA dataset, Residual Network (ResNet) extractor achieves an accuracy (ACC) 0.992 area under curve (AUC) 0.997. CC-CCII ACC 0.994 AUC Finally, compared existing methods, has obvious all aspects. These demonstrate effectively solve field

Language: Английский

---

HRNet: Differentially Private Hierarchical and Multi-Resolution Network for Human Mobility Data Synthesization

Proceedings of the VLDB Endowment, Journal Year: 2024, Volume and Issue: 17(11), P. 3058 - 3071

Published: July 1, 2024

Human mobility data offers valuable insights for many applications such as urban planning and pandemic response, but its use also raises privacy concerns. In this paper, we introduce the Hierarchical Multi-Resolution Network (HRNet), a novel deep generative model specifically designed to synthesize realistic human while guaranteeing differential privacy. We first identify key difficulties inherent in learning under response these challenges, HRNet integrates three components: hierarchical location encoding mechanism, multi-task across multiple resolutions, private pre-training. These elements collectively enhance model's ability constraints of Through extensive comparative experiments utilizing real-world dataset, demonstrates marked improvement over existing methods balancing utility-privacy trade-off.

Language: Английский

---

Personalized and On-device Trajectory Mobility Prediction

Published: Oct. 29, 2024

Language: Английский

---