Modern cities generate vast streams of urban dynamics data reflecting mobility demand, environmental conditions, and traffic patterns. The value of these data lies not only in individual modalities but in their integration—urban signals are highly interdependent, with changes in one modality often influencing others. Consequently, predicting any single urban dynamic requires information from multiple interrelated sources. Although numerous methods—ranging from deep learning models to recent LLM-based approaches—have been proposed, most are limited in scope. They either focus on single-modality prediction, rely on rigid model designs that lack flexibility, or overlook inter-modal dependencies. As a result, they struggle to adapt to dynamic urban conditions and suffer from degraded predictive performance across modalities. In this paper, we propose UniLLM, a unified large language model for multi-modal urban dynamics prediction. At its core, UniLLM introduces a Unified Cross-Modal Alignment Module that transforms heterogeneous urban data into latent representations while preserving modality-specific patterns and capturing cross-modal correlations through a contrastive learning objective. To support dynamic adaptation across tasks and modalities, we design a Routing-Aware Prompting Mechanism that learns soft prompts based on task context and modality semantics. Furthermore, a Multi-Modal Memory-Guided Adaptive Algorithm employs replay-based gradient coordination and Frank–Wolfe optimization to mitigate cross-modal catastrophic forgetting during fine-tuning. Extensive experiments across multiple cities and urban modalities demonstrate that UniLLM consistently outperforms state-of-the-art baselines. These results highlight UniLLM's potential as a flexible and robust forecasting model for real-world, multi-modal urban environments.

This paper introduces a novel approach to fine-tuning transformer-based models for various spatial-temporal downstream tasks. While fine-tuning approaches have shown remarkable success in fields like natural language processing, their efficacy in human-generated spatial-temporal data is often hindered by the complex spatial-temporal correlation and extensive reliance on labeled data. We introduce Knowledge Graph-Guided Spatial-Temporal Cross-task Fine-Tuning method, i.e., KG-STFT, a cross-task fine-tuning approach designed for adapting to data-scarce spatial-temporal tasks by leveraging rich knowledge embedded in similar downstream tasks and pre-trained model. Our KG-STFT framework utilizes i) a non-linear knowledge ensembler to capture and integrate knowledge embedded in different transformer blocks, and ii) constructs a task knowledge graph to "transfer" knowledge from data-rich to data-scarce tasks. Empirical experiments on real-world taxi trajectory data show that KG-STFT outperforms baselines, especially in data-scarce tasks, by leveraging task commonalities to improve fine-tuning.

Score-based Generative Models (SGMs) have demonstrated remarkable generalization abilities, e.g. generating unseen, but natural data. However, the greater the generalization power, the more likely the unintended generalization, and the more dangerous the abuse. Research on moderated generalization in SGMs remains limited. To fill this gap, we first examine the current 'gold standard' in Machine Unlearning (MU), i.e., re-training the model after removing the undesirable training data, and find it does not work in SGMs. Further analysis of score functions reveals that the MU 'gold standard' does not alter the original score function, which explains its ineffectiveness. Based on this insight, we propose the first Moderated Score-based Generative Model (MSGM), which introduces a novel score adjustment strategy that redirects the score function away from undesirable data during the continuous-time stochastic differential equation process. Extensive experimental results demonstrate that MSGM significantly reduces the likelihood of generating undesirable content while preserving high visual quality for normal image generation. Albeit designed for SGMs, MSGM is a general and flexible MU framework that is compatible with diverse diffusion architectures (SGM and DDPM) and training strategies (re-training and fine-tuning), and enables zero-shot transfer of the pre-trained models to downstream tasks, e.g. image inpainting and reconstruction. The code will be shared upon acceptance.

The development of sensing technologies has broad-ened the scope of urban dynamics research. However, existing methods primarily focus on using isolated aspects of urban data, limiting their ability to capture the complex spatial-temporal dependency among different urban dynamics and generalizability across applications. Addressing these shortcomings requires a more comprehensive model capable of integrating multifaceted data, generating generalized representations adaptable to diverse applications and scenarios. In this paper, we introduce the Multifaceted SpAtial-TempoRal ContrastivE Learning framework, i.e., MARCEL, an innovative approach designed to learn robust, universally applicable, and adaptable representations of multifaceted urban dynamics through contrastive learning. MARCEL employs pretrained preliminary representation learning modules to extract distinct spatial-temporal dependencies inherent to each urban dynamic independently. It then features a Spatial-Temporal Contrastive Learning strategy to capture unified spatial-temporal patterns, including asynchronous, conflicting, and complementary behaviors across multifaceted urban dynamics. Additionally, MARCEL integrates a Multifaceted Knowledge Transfer mechanism to capture inter-dependencies among different urban dynamics and facilitate knowledge sharing. The learned representations are highly generalizable and can be applied effectively to various downstream tasks. Extensive experiments on real-world urban datasets demonstrate that MARCEL is effective and significantly outperforms state-of-the-art baselines.

Understanding and predicting urban dynamics is crucial for managing transportation systems, optimizing urban planning, and enhancing public services. While neural network-based approaches have achieved success, they often rely on task-specific architectures and large volumes of data, limiting their ability to generalize across diverse urban scenarios. Meanwhile, Large Language Models (LLMs) offer strong reasoning and generalization capabilities, yet their application to spatial-temporal urban dynamics remains underexplored. Existing LLM-based methods struggle to effectively integrate multifaceted spatial-temporal data and fail to address distributional shifts between training and testing data, limiting their predictive reliability in real-world applications. To bridge this gap, we propose UrbanMind, a novel spatial-temporal LLM framework for multifaceted urban dynamics prediction that ensures both accurate forecasting and robust generalization. At its core, UrbanMind introduces Muffin-MAE, a multifaceted fusion masked autoencoder with specialized masking strategies that capture intricate spatial-temporal dependencies and intercorrelations among multifaceted urban dynamics. Additionally, we design a semantic-aware prompting and fine-tuning strategy that encodes spatial-temporal contextual details into prompts, enhancing LLMs' ability to reason over spatial-temporal patterns. To further improve generalization, we introduce a test time adaptation mechanism with a test data reconstructor, enabling UrbanMind to dynamically adjust to unseen test data by reconstructing LLM-generated embeddings. Extensive experiments on real-world urban dynamics datasets from multiple cities demonstrate the effectiveness of UrbanMind. The results consistently show that UrbanMind outperforms state-of-the-art baselines, achieving superior accuracy and strong generalization, even in zero-shot scenarios with no prior data.

The problem of traffic dynamics prediction, aiming to capture the complicated patterns of urban dynamics and forecast short-term future traffic status, is essential for managing transportation systems, reducing congestion, enhancing safety, improving commuter efficiency, and supporting urban planning and infrastructure development. Current approaches using ma-chine learning and deep neural networks have advanced traffic prediction but often focus on individual urban dynamic aspects and rely on auto-regressive methods for consecutive predictions, which can be inaccurate and computationally expensive. In this work, we propose the Spatial- Temporal Graph LAtent DIffusion ModeL (STGAIL) to address these limitations. STGAIL views geographical regions as graphs with various traffic features, capturing their interconnections. Operating in a pre-trained latent space, STGAIL uses latent diffusion processes and inno-vative spatial-temporal graph layers for accurate and efficient multi-step predictions. Fine-tuning with temporal binary masks further enhances its performance, avoiding error accumulation and reducing computational costs. Experiments on real-world datasets demonstrate STGAIL's superior accuracy and efficiency over state-of-the-art methods. We also make our code and dataset available, contributing to ongoing research in traffic dynamics prediction.

Enhancing diverse human decision-making processes in an urban environment is a critical issue across various applications, including ride-sharing vehicle dispatching, public transportation management, and autonomous driving. Offline reinforcement learning (RL) is a promising approach to learn and optimize human urban strategies (or policies) from pre-collected human-generated spatial-temporal urban data. However, standard offline RL faces two significant challenges: (1) data scarcity and data heterogeneity, and (2) distributional shift. In this paper, we introduce MODA - a Multi-Task Offline Reinforcement Learning with Contrastive Data Sharing approach. MODA addresses the challenges of data scarcity and heterogeneity in a multi-task urban setting through Contrastive Data Sharing among tasks. This technique involves extracting latent representations of human behaviors by contrasting positive and negative data pairs. It then shares data presenting similar representations with the target task, facilitating data augmentation for each task. Moreover, MODA develops a novel model-based multi-task offline RL algorithm. This algorithm constructs a robust Markov Decision Process (MDP) by integrating a dynamics model with a Generative Adversarial Network (GAN). Once the robust MDP is established, any online RL or planning algorithm can be applied. Extensive experiments conducted in a real-world multi-task urban setting validate the effectiveness of MODA. The results demonstrate that MODA exhibits significant improvements compared to state-of-the-art baselines, showcasing its capability in advancing urban decision-making processes. We also made our code available to the research community.

Human-generated Spatial-Temporal Data (HSTD), represented as trajectory sequences, has undergone a data revolution, thanks to advances in mobile sensing, data mining, and AI. Previous studies have revealed the effectiveness of employing attention mechanisms to analyze massive HSTD. However, traditional attention models face challenges when managing lengthy and noisy trajectories as their computation comes with large memory overheads. Furthermore, attention scores within HSTD trajectories are sparse (i.e., most of the scores are zeros), and clustered with varying lengths (i.e., consecutive tokens clustered with similar scores). To address these challenges, we introduce an innovative strategy named Memory-efficient Trajectory Attention (MeTA). We leverage complicated spatial-temporal features (e.g., traffic speed, proximity to PoIs) and design an innovative feature-based trajectory partition technique to shrink trajectory length. Additionally, we present a learnable dynamic sorting mechanism, with which attention is only computed between sub-trajectories that have prominent correlations. Empirical validations using real-world HSTD demonstrate that our approach not only yields competitive results but also significantly lowers memory usage compared with state-of-the-art methods. Our approach presents innovative solutions for memory-efficient trajectory attention, offering valuable insights for handling HSTD efficiently.

Rapid advances in perception, planning, and decision-making areas for self-driving vehicles have led to great improvements in their function and capabilities and enabled several prototypes to be driving on the roads and streets, such as Waymo Driver, TuSimple, Nuro, etc. Among various applications of self-driving vehicles, a promising one is the ride service as it has the potential to improve service quality and productivity and to provide service to anyone at any time. Extensive studies have been conducted on self-driving planning and safety, but few works focus on self-driving ride service decision-making and routing. In this work, we take the lead to study self-driving ride service planning and decision-making problem leveraging human-generated spatial-temporal data, and propose the data-driven Conservative Actor-Critic approach – CAC – based on offline reinforcement learning. Our CAC is able to make conservative decisions in a complicated environment with multiple goal states, and avoid dangerous and overly optimistic behaviors by exploiting human decisions. Extensive experiments with real-world data demonstrate that our CAC-learned policies are able to improve taxi service operation efficiency and quality drastically in terms of shortening passenger waiting time and improving service revenue.

Advancements in mobile sensing, data mining, and artificial intelligence have revolutionized the collection and analysis of Human-generated Spatial-Temporal Data (HSTD), paving the way for diverse applications across multiple domains. However, previous works have primarily focused on designing task-specific models for different problems, which lack transferability and generalizability when confronted with diverse HSTD. Additionally, these models often require a large amount of labeled data for optimal performance. While pre-trained models in Natural Language Processing (NLP) and Computer Vision (CV) domains have showcased impressive transferability and generalizability, similar efforts in the spatial-temporal data domain have been limited. In this paper, we take the lead and introduce the Spatial-Temporal Pre-Training model, $i.e$., STPT, which is connected with a self-supervised learning task, to address these limitations. STPT enables the creation of robust and versatile representations of HSTD. We validate our framework using real-world data and demonstrate its efficacy through two downstream tasks, $i.e$., trajectory classification and driving activity identification $(e.g$., identifying seeking $vs$. serving behaviors in taxi trajectories). Our results achieve an accuracy of 83.125% (16.2% higher than the average baseline) for human mobility identification and an accuracy of 77.88% (13.0% higher than the average baseline) for the human activity identification task. These outcomes underscore the potential of our pre-trained model for diverse downstream applications within the spatial-temporal data domain.

The two leading solution paradigms for imitation learning (IL), BC and GAIL, each suffers from notable drawbacks. BC, a supervised learning approach to mimic expert actions, is vulnerable to covariate shift. GAIL applies adversarial training to minimize the discrepancy between expert and learner behaviors, which is prone to unstable training and mode collapse. In this work, we propose DC – Distributional Cloning – a novel IL approach for addressing the covariate shift and mode collapse problems simultaneously. DC directly maximizes the likelihood of observed expert and learner demonstrations, and gradually encourages the learner to evolve towards expert behaviors based on an averaging effect. The DC solution framework contains two stages in each training loop, where in stage one the mixed expert and learner state distribution is estimated via SoftFlow, and in stage two the learner policy is trained to match both the expert’s policy and state distribution via ADMM. Experimental evaluation of DC compared with several baselines in 10 different physics-based control tasks reveal superior results in learner policy performance, training stability, and mode distribution preservation.

The Human Mobility Signature Identification (HuMID) problem aims at determining whether the incoming trajectories were generated by a claimed agent from the historical movement trajectories of a set of individual human agents such as pedestrians and taxi drivers. The HuMID problem is significant, and its solutions have a wide range of real-world applications, such as criminal identification for police departments, risk assessment for auto insurance providers, driver verification in ride-sharing services, and so on. Though Deep neural networks (DNN) based HuMID models on spatial-temporal mobility fingerprint similarity demonstrate remarkable performance in effectively identifying human agents' mobility signatures, it is vulnerable to adversarial attacks as other DNN-based models. Therefore, in this paper, we propose a Spatial-Temporal iterative Fast Gradient Sign Method with L0 regularization - ST-iFGSM - to detect the vulnerability and enhance the robustness of HuMID models. Extensive experiments with real-world taxi trajectory data demonstrate the efficiency and effectiveness of our ST-iFGSM algorithm. We tested our method on both the ST-SiameseNet and an LSTM-based HuMID classification model. It shows that ST-iFGSM can generate successful attacks to fool the HuMID models with only a few steps of attack in a small portion of the trajectories. The generated attacks can be used as augmented data to update and improve the HuMID model accuracy significantly from 47.36% to 76.18% on testing samples after the attack(86.25% on the original testing samples).

A key challenge with supervised learning (e.g., image classification) is the shift of data distribution and domain from training to testing datasets, so-called “domain shift” (or “distribution shift”), which usually leads to a reduction of model accuracy. Various meta-learning approaches have been proposed to prevent the accuracy loss by learning an adaptable model with training data, and adapting it to test time data from a new data domain. However, when the domain shift occurs in multiple domain dimensions (e.g., images may be transformed by rotations, transitions, and expansions), the average predictive power of the adapted model will deteriorate. To tackle this problem, we propose a domain disentangled meta-learning (DDML) framework. DDML disentangles the data domain by dimensions, learns the representations of domain dimensions independently, and adapts to the domain of test time data. We evaluate our DDML on image classification problems using three datasets with distribution shifts over multiple domain dimensions. Comparing to various baselines in meta-learning and empirical risk minimization, our DDML approach achieves consistently higher classification accuracy with the test time data. These results demonstrate that domain disentanglement reduces the complexity of the model adaptation, thus increases the model generalizability, and prevents it from overfitting.

Lightweight neural networks refer to deep networks with small numbers of parameters, which can be deployed in resource-limited hardware such as embedded systems. To learn such lightweight networks effectively and efficiently, in this paper we propose a novel convolutional layer, namely Channel-Split Recurrent Convolution (CSR-Conv), where we split the output channels to generate data sequences with length T as the input to the recurrent layers with shared weights. As a consequence, we can construct lightweight convolutional networks by simply replacing (some) linear convolutional layers with CSR-Conv layers. We prove that under mild conditions the model size decreases with the rate of $O\left( {\frac{1}{{{T^2}}}} \right)$. Empirically we demonstrate the state-of-the-art performance using VGG-16, ResNet-50, ResNet-56, ResNet-110, DenseNet-40, MobileNet, and EfficientNet as backbone networks on CIFAR-10 and ImageNet. Codes can be found on https://github.com/tuaxon/CSR_Conv.

Smart passenger-seeking strategies employed by taxi drivers contribute not only to drivers’ incomes, but also higher quality of service passengers received. Therefore, understanding taxi drivers’ behaviors and learning the good passenger-seeking strategies are crucial to boost taxi drivers’ well-being and public transportation quality of service. However, we observe that drivers’ preferences of choosing which area to find the next passenger are diverse and dynamic across locations and drivers. It is hard to learn the location-dependent preferences given the partial data (i.e., an individual driver’s trajectory may not cover all locations). In this article, we make the first attempt to develop conditional generative adversarial imitation learning (cGAIL) model, as a unifying collective inverse reinforcement learning framework that learns the driver’s decision-making preferences and policies by transferring knowledge across taxi driver agents and across locations. Our evaluation results on three months of taxi GPS trajectory data in Shenzhen, China, demonstrate that the driver’s preferences and policies learned from cGAIL are on average 36.2 percent more accurate than those learned from other state-of-the-art baseline approaches.

Thanks to the rapid development of mobile sensing techniques, massive human-generated spatial-temporal data (HSTD) are generated from the urban areas, e.g., passenger-seeking trajectories from taxi drivers, and public transit trips from urban dwellers. These HSTD record sequential decisions made by human agents. Studying human behavior from HSTD provides benefits to many aspects, for example, studying passenger-seeking strategies from experienced taxi drivers can help improve the operation efficiencies of those new drivers. One common method to analyze human behavior from HSTD is Imitation Learning (IL). Existing IL approaches rely on data collected from experts. However, human agents who generate HSTD may have diverse expertise levels across geographical regions, i.e., with good policies in some regions and poor policies in less experienced regions. The problem of how to infer the optimal policy for agents in their unfamiliar or less-experienced regions remains open. In this paper, we propose the novel Generative Adversarial Imitation Learning for Non-experts (NEXT-GAIL) framework to first disentangle expert knowledge, which is irrelevant to spatial-temporal regions, from the demonstration data. Then, such knowledge can be transferred to regions, where the agent does not possess an expert policy. We take the real-world taxi trajectory data as an example to evaluate the performance of our proposed framework. The comparison results illustrate that our proposed NEXT-GAIL outperforms existing state-of-the-art approaches regarding the accuracy of the inferred optimal policy for non-experts.

Given the underlying road network of an urban area, the problem of urban dynamics prediction aims to capture the patterns of urban dynamics and to forecast short-term urban traffic status continuously from the historical observations. This problem is of fundamental importance to urban traffic management, planning, and various business services. However, predicting urban dynamics is challenging due to the highly dynamic (i.e., varying across geographical locations and evolving over time) and uncertain (i.e., affected by unexpected factors) nature of urban traffic systems. Recent works adopt meta-learning approaches to capture irregular and rare patterns but make unrealistic assumptions such as single-domain uncertainties and explicit temporal task segmentation. In this paper, we solve the urban dynamics prediction problem from the Bayesian meta-learning perspective and propose a novel domain adaptable continuous meta-learning approach (DAC-ML) that does not require task segmentation. Trained on a sequence of spatial-temporal urban dynamics data, DAC-ML aims to detect and infer unobserved latent variations (from task and domain levels) and generalize well in a sequential prediction setting, where the underlying data generating process varies over time. Experimental results on three real-world datasets demonstrate that DAC-ML can outperform baselines in urban dynamics prediction, especially when obvious urban dynamics and temporal uncertainties are present.

Mobile sensing and information technology have enabled us to collect a large amount of mobility data from human decision-makers, for example, GPS trajectories from taxis, Uber cars, and passenger trip data of taking buses and trains. Understanding and learning human decision-making strategies from such data can potentially promote individual's well-being and improve the transportation service quality. Existing works on human strategy learning, such as inverse reinforcement learning, all model the decision-making process as a Markov decision process, thus assuming the Markov property. In this work, we show that such Markov property does not hold in real-world human decision-making processes. To tackle this challenge, we develop a Trajectory Generative Adversarial Imitation Learning (TrajGAIL) framework. It captures the long-term decision dependency by modeling the human decision processes as variable length Markov decision processes (VLMDPs), and designs a deep-neural-network-based framework to inversely learn the decision-making strategy from the human agent's historical dataset. We validate our framework using two real world human-generated spatial-temporal datasets including taxi driver passenger-seeking decision data and public transit trip data. Results demonstrate significant accuracy improvement in learning human decision-making strategies, when comparing to baselines with Markov property assumptions.

Imitation learning (IL) aims to learn a policy from expert demonstrations that minimizes the discrepancy between the learner and expert behaviors. Various imitation learning algorithms have been proposed with different pre-determined divergences to quantify the discrepancy. This naturally gives rise to the following question: Given a set of expert demonstrations, which divergence can recover the expert policy more accurately with higher data efficiency? In this work, we propose f-GAIL – a new generative adversarial imitation learning model – that automatically learns a discrepancy measure from the f-divergence family as well as a policy capable of producing expert-like behaviors. Compared with IL baselines with various predefined divergence measures, f-GAIL learns better policies with higher data efficiency in six physics-based control tasks.

Smart passenger-seeking strategies employed by taxi drivers contribute not only to drivers' incomes, but also higher quality of service passengers received. Therefore, understanding taxi drivers' behaviors and learning the good passenger-seeking strategies are crucial to boost taxi drivers' well-being and public transportation quality of service. However, we observe that drivers' preferences of choosing which area to find the next passenger are diverse and dynamic across locations and drivers. It is hard to learn the location-dependent preferences given the partial data (i.e., an individual driver's trajectory may not cover all locations). In this paper, we make the first attempt to develop conditional generative adversarial imitation learning (cGAIL) model, as a unifying collective inverse reinforcement learning framework that learns the driver's decision-making preferences and policies by transferring knowledge across taxi driver agents and across locations. Our evaluation results on three months of taxi GPS trajectory data in Shenzhen, China, demonstrate that the driver's preferences and policies learned from cGAIL are on average 34.7% more accurate than those learned from other state-of-the-art baseline approaches.