Traffic flow prediction is a typical spatio-temporal prediction problem and has a wide range of applications. The core challenge lies in modeling the underlying complex spatio-temporal dependencies. Various methods have been proposed, and recent studies show that the modeling of dynamics is useful to meet the core challenge. While handling spatial dependencies and temporal dependencies using separate base model structures may hinder the modeling of spatio-temporal correlations, the modeling of dynamics can bridge this gap. Incorporating spatio-temporal heterogeneity also advances the main goal, since it can extend the parameter space and allow more flexibility. Despite these advances, two limitations persist: 1) the modeling of dynamics is often limited to the dynamics of spatial topology (e.g., adjacency matrix changes), which, however, can be extended to a broader scope; 2) the modeling of heterogeneity is often separated for spatial and temporal dimensions, but this gap can also be bridged by the modeling of dynamics. To address the above limitations, we propose a novel framework for traffic prediction, called Meta Dynamic Graph (MetaDG). MetaDG leverages dynamic graph structures of node representations to explicitly model spatio-temporal dynamics. This generates both dynamic adjacency matrices and meta-parameters, extending dynamic modeling beyond topology while unifying the capture of spatio-temporal heterogeneity into a single dimension. Extensive experiments on four real-world datasets validate the effectiveness of MetaDG.

Human mobility prediction is a fundamental technique for many urban applications, e.g., location-based recommendation, traffic scheduling, and travel demand prediction. Over the past decades, many methods, e.g., Markov Model, RNN, Transformer, have been leveraged to tackle the problem. However, existing approaches mainly train a supervised model based on an offline training dataset, which overlooks the phenomenon that the mobility behaviors of humans vary across time, and the trained models may not achieve ideal performance when applied to the testing data. To tackle this challenge, in this paper, we propose AdaMove, an efficient Test-Time Adaptive (TTA) model for human mobility prediction. AdaMove has a Preference-aware Test-Time Adaptation module called PTTA, which can adjust the parameters of a trained model based on the input test trajectory such that the model can generalize to the test distribution. In addition, to address the issue of reduced inference efficiency caused by parameter adjustment during the testing phase, AdaMove is equipped with a Lightweight human Mobility prediction model called LightMob, which only requires the recent trajectory as input to accelerate the inference. It is enhanced by historical trajectory knowledge via contrastive learning during the training time, so it has competitive performance compared with existing models. Extensive experiments on three real-world human mobility datasets demonstrate that AdaMove outperforms the best baseline by 9.3% on average in accuracy, and accelerates the inference speed by 28.5% on average compared with the original TTA - based inference.

Inferring the actual road segment purely based on one positioning point, known as single-point map matching (SMM), is vital for many urban applications, e.g., ride-hailing and geo-tagging. However, it is challenging due to inherent positioning errors and extrinsic heterogeneous environments. Existing methods either overlook the heterogeneity of different regions, or do not exploit the commonality of different matching tasks. In this paper, we treat each region as an individual SMM task to tackle the heterogeneity, and propose Spatial Hierarchical Meta-Learning for SMM (SHSMM) to learn the shared knowledge across tasks. SHSMM is equipped with a Dual-view Map Matcher to perform the matching, which can perceive the knowledge of road segments globally. To learn the task-specific model parameters, SHSMM modulates initial parameters and scales the local update learning rate based on hierarchical geographical and semantic knowledge about spatial tasks. A local update learning rate scheduling strategy is further proposed to facilitate the meta-training. Extensive experiments as well as case studies based on two real-world datasets demonstrate the effectiveness of the proposed method.

Air pollution remains one of the most critical environmental challenges globally, posing severe threats to public health, ecological sustainability, and climate governance. While existing physics-based and data-driven models have made progress in air quality forecasting, they often struggle to jointly capture the complex spatiotemporal dynamics and ensure spatial continuity of pollutant distributions. In this study, we introduce CTENet, a novel chemical transport deep learning model that embeds the Advection-Diffusion-Reaction equation into a Physics-Informed Neural Network (PINN) framework using an Eulerian representation to model the spatiotemporal evolution of pollutants. Extensive experiments on two real-world datasets demonstrate that CTENet consistently outperforms state-of-the-art (SOTA) baselines, achieving a remarkable RMSE improvement of 45.8% on the USA dataset and 21.0% on the China dataset.

Large Language models (LLMs) usually rely on extensive training datasets. In the financial domain, creating numerical reasoning datasets that include a mix of tables and long text often involves substantial manual annotation expenses. To address the limited data resources and reduce the annotation cost, we introduce FinLLMs, a method for generating financial question-answering (QA) data based on common financial formulas using LLMs. First, we compile a list of common financial formulas and construct a graph based on the variables these formulas employ. We then augment the formula set by combining those that share identical variables as new elements. Specifically, we explore formulas obtained by manual annotation and merge those formulas with shared variables by traversing the constructed graph. Finally, utilizing LLMs, we generate financial QA data that encompasses both tabular information and long textual content, building on the collected formula set. Our experiments demonstrate that the synthetic data generated by FinLLMs effectively enhances the performance of various numerical reasoning models in the financial domain, including both pre-trained language models (PLMs) and fine-tuned LLMs. This performance surpasses that of two established benchmark financial QA datasets.

Intelligent logistics relies on accurately predicting the service time, which is a part of time cost in the last-mile delivery. However, service time prediction (STP) is non-trivial given complex delivery circumstances, location heterogeneity, and skewed observations in space, which are not well-handled by existing solutions. In our prior work, we treat STP at each location as a learning task to keep the location heterogeneity, propose a prior knowledge-enhanced meta-learning to tackle skewed observations, and introduce a Transformer-based representation module to encode complex delivery circumstances. Maintaining the design principles of prior work, in this extended paper, we propose MetaSTP+. In addition to fusing the prior knowledge after the meta-learning process, MetaSTP+also injects the prior knowledge before and during the meta-learning process to better tackle skewed observations. More specifically, MetaSTP+completes the support set of tasks with scarce samples from other tasks based on prior knowledge and is equipped with a prior knowledge-aware historical observation encoding module to achieve those purposes accordingly. Experiments show MetaSTP+outperforms the best baseline by 11.2% and 8.4% on two real-world datasets. Finally, an intelligent waybill assignment system based on MetaSTP+is deployed in JD Logistics.

Cross-domain sequential recommendation (CDSR) utilizes data from multiple domains to recommend the user’s next interaction based on his latest interaction sequence. Currently, many cross-domain sequential recommendation algorithms have been proven to achieve good recommendation performance. However, these algorithms overlook the influence of users’ long-term behavioral patterns and general interests when extracting their current preferences. In this article, we propose a Hierarchical Gating Network for Cross-Domain Sequential Recommendation (HGNCDSR). Specifically, we simultaneously train single-domain and cross-domain interaction sequences, utilizing a hierarchical gating network to capture user interest representations in single-domain and cross-domain, respectively. A feature gating and an instance gating are applied respectively to extract user interests at item feature level and instance level. While learning current preferences from behavior sequences, user representations that reflect behavioral patterns and general interests are simultaneously learned and strengthened. Additionally, we employ the item–item product to model the relationships between candidate items and those in the interaction sequence. Both current interests and item relevance are considered simultaneously, integrating single-domain and cross-domain user preferences to predict the user’s next interaction. We design extensive experiments to show that HGNCDSR has better recommendation performance than other state-of-the-art models.

Large language models (LLMs) require vast and high-quality training datasets, which are often difficult to annotate, particularly in domains such as finance. In finance, the datasets not only encompass large volumes of textual data but also involve intricate numerical reasoning, frequently presented in the form of tables. One significant challenge in this domain is the processing of exceptionally long financial statements and complex tables, which can hinder the reasoning capacity of LLMs and negatively impact their accuracy. To mitigate the adverse effects of lengthy text passages and complex tables, this study focuses on identifying and extracting key information (gold facts) that are essential for solving financial problems, with the aim of enhancing both reasoning quality and data efficiency. We investigate the disparities between various financial datasets and examine different model extraction techniques. Our experimental findings demonstrate that the precise extraction of these key facts from both textual and tabular data substantially improves the accuracy of LLMs-based question answering, while simultaneously reducing the amount of data necessary for effective model performance.

Urban sensing aims to sense the status of the city, e.g., air quality, noise level, concentration of viruses, which can be completed by spatial crowdsourcing. Multi-destination people, who have many intermediate locations to visit before the final destination, e.g., couriers and tourists, are ideal recruitment candidates to conduct sensing tasks since they spend more time outside and have a wide spatio-temporal distribution. However, existing spatial crowdsourcing methods are only designed for workers who have single destinations, e.g., commuters, which are not applicable to recruit the multiple-destination people. Therefore, in this paper, we generalize the urban crowdsensing problem to the multi-destination scenario, namely, Urban Sensing for Multi-Destination Workers (USMDW). We prove its NP-hardness, and propose a framework Urban Sensing for Multi-destination Workers via Deep REinforcement learning, i.e., SMORE, to solve it effectively and efficiently. SMORE is composed of two steps: 1) candidate assignment initialization, which initializes all feasible sensing task-worker assignment pairs by a pre-trained reinforcement learning-based working route planning solver; and 2) reinforcement learning-based iterative selection, which iteratively selects a sensing task-worker pair to the current assignment via a novel policy network, i.e., Two-stage Assignment Selection Network (TASNet). Extensive experiments on three real-world datasets show SMORE outperforms the best baseline in data coverage by 5.2% on average with high efficiency.

The updated residential-level fine-grained digital map is essential for last-mile delivery. However, many of those low-level roads are not recorded in maps due to the high mapping costs. With the digitization of the logistics industry, couriers’ trajectories become a promising data source to complete missing roads in maps. Existing trajectory-based map updating work rely on heavy parameter tuning to overcome the positioning error issue due to their unsupervised nature, and are not able to handle issues of unreliable road indicators and skewed data distributions. To tackle those challenges, in this paper, we propose a framework DelvMap to complete missing roads in maps based on couriers’ trajectories and satellite images. DelvMap first leverages a dual signal fusion network (DSFNet) to extract an inferred map from both data modalities, which fully exploits the positive and negative signals in the satellite images to fuse with roads indicated from trajectories, then employs a map completion algorithm to complete the existing map with the inferred map, the connection strategy of which is adaptive to the number of traversing trajectories. Experiments show DelvMap outperforms baselines by at least 11.0% in TOPO F1 on the real-world dataset. Finally, we demonstrate a multi-modal map updating system based on DelvMap.