The resilience of heterogeneous multi-agent systems (HMAS) is critical for maintaining effective task execution under disruptive events. Although extensive studies have explored HMAS resilience on structured collaboration among agents with heterogeneous functional roles, they primarily evaluate system performance through collaboration path efficiency or other statistical indicators. However, a direct structural characterization linking collaborative structures to system capability is still lacking. In this work, we propose a spectral capability metric that quantifies collaborative performance directly from heterogeneous collaborative structures. Building on this representation, we introduce a spectral resilience framework that captures resistance to disruption and recoverability of collaborative structures. We apply the proposed spectral capability metric and resilience framework to HMAS incorporating evolutionary game-based reconfiguration, where system recovery is modeled as adaptive reconfiguration of task-executable collaborative paths. Comparative analyses across benchmark metrics and multiple scenarios validate the discriminative effectiveness of the proposed approach. Finally, extensive numerical experiments examine the influence of network topology, reconfiguration strategy, evolutionary parameters, and agent composition. Results show that balanced role distributions maximize collaborative capability, while resilience remains high under skewed compositions. Additionally, we also find that ER and WS networks achieve higher collaborative performance, while regular and BA networks exhibit stronger resilience.

Objective: Dynamic allocation of emergency medical resources is a critical task in the prevention and control of respiratory infectious diseases (RIDs). This study aims to address the challenge of a “run on the healthcare system” by proposing an effective resource allocation strategy to curb the spread of RIDs. Methods: Considering the infection severity of RIDs, the capacity of emergency medical resources (hospitalization rate), and vaccination status, we construct an SVInR dynamic model of RIDs that considers vaccine failure. Under the constraint of emergency medical resources and with the goal of minimizing the basic reproduction number, we propose a dynamic allocation strategy for distributing emergency medical resources among different types of infected individuals. Results: Simulation results demonstrate that improving the hospitalization efficiency of emergency medical resources significantly contributes to the effective control of RIDs. The model shows that targeted dynamic allocation helps reduce disease transmission. Conclusions: Validation using real-world data confirms that the model is effective and practical. It offers theoretical guidance for dynamically allocating emergency medical resources and supports informed decision-making in response to major emerging RIDs.

The recurrent outbreaks of global infectious diseases have highlighted the urgent need for a unified spatiotemporal modeling framework that can capture multiscale processes, mechanistic interactions, and data heterogeneity. Previous reviews have mainly addressed either mechanistic or data-based approaches, leaving a gap in understanding how these two perspectives can be systematically integrated. To address this gap, this paper traces the evolution of epidemic modeling from mechanism-driven frameworks grounded in epidemiological theory to data-driven paradigms enabled by machine learning. It first follows the progression from classical homogeneous mixing models to those incorporating heterogeneity in age, space, contact structure, and pathogen variation. It then outlines the rise of data-based approaches, beginning with statistical models and extending to modern machine learning methods. By outlining this methodological continuum, the review provides an integrated perspective on the co-evolution of mechanistic and data-driven approaches, offering a unified theoretical and computational foundation for the spatiotemporal analysis of epidemic dynamics and for intelligent, data-informed public health interventions.

Accurately modeling human mobility is essential for understanding traffic congestion, disease transmission, commodity flow and social dynamics. Traditional models such as the gravity model and intervening opportunity model provide foundational insights but often overlook fine-grained geographic factors. In this study, we propose a novel hybrid deep learning framework, FlowSynthNet (FSN) model, to improve the accuracy of human mobility prediction. Our model integrates one-dimensional convolutional neural networks, Transformer-based attention mechanisms, and LightGBM to capture local spatial features, global dependencies and non-linear interactions. Evaluated on mobility datasets from Beijing, New York and Los Angeles, FSN model consistently outperforms classical models and machine learning baselines. Furthermore, feature interpretation reveals how specific geographic elements influence mobility flows. Our results highlight the effectiveness of combining geographic semantics with deep learning for more precise and interpretable human mobility modeling.

Accurate identification, quantification, and continual monitoring of carbon emissions constitute pivotal elements for proactive climate interventions. Conventional methodologies like direct measurement, including point-source assessments or remote sensing, often face challenges related to high costs or limited accuracy. Especially in low- and middle-income nations experiencing escalating emissions, carbon monitoring heavily relies on the administrative capabilities of local governments, which frequently lack adequate monitoring infrastructures. Addressing this predicament, our study introduces a computational framework to forecast CO 2 emissions by leveraging comprehensive observable human activity data from third-party sources. Our findings elucidate a robust correlation between multi-origin CO 2 emissions and human mobility ( r = 0.89). Notably, machine learning models adeptly predict these emissions by integrating characteristics extracted from temporally aggregated, anonymized mobility networks (R 2 ≈1.0). We demonstrate that the model effectively captures the notable reduction in CO 2 emissions during the COVID-19 lockdown, with both human mobility and CO 2 emissions in China decreasing by 56.97% and 32.45%, respectively. The prediction accuracy remains high for countries with varying social economic development, such as the U.S., Italy and Mexico. This study presents an inexpensive, real-time, and robust method of quantifying CO 2 emissions on a large scale with high precision, and it could facilitate tailored CO 2 emission reduction strategies, grounded in robust scientific evidence derived from the dynamics of human mobility.

Accurate forecasting of realized volatility (RV) is crucial for risk management, asset pricing, and policy making in financial markets. This study introduces the Multi-scale Decomposition and Fusion Network (MDFNet), a novel fore casting model designed to enhance the knowledge integration of multi-scale volatility patterns. The core concept of MDFNet centers on decomposition and fusion. It utilizes Variational Mode Decomposition (VMD) to break down the RV series into intrinsic mode functions (IMFs) and applies one-dimensional convolutional kernels for extracting detailed knowledge repre sentations from multi-scale patterns. The model integrates these representations through a local-global fusion mechanism that combines a local progressive strategy with a global approach, enhancing the synthesis of information across feature, temporal, and scale dimensions. Extensive experiments on RV series from six representative stocks across emerging and developed markets, as well as nine major international indices, show that MDFNet consistently outperforms baseline models. It achieves superior accuracy across both daily RV series derived from Level-1 transaction data and higher-frequency RV series constructed from Level-2 order book data. Codes and datasets are available at https://github.com/omx-Maisie/MDFNet.

Modernization of social governance is an important component of modernizing the national governance system and governance capacity.With the integration and development of location-aware technologies such as mobile communications and satellite positioning have generated massive mobile big data with spatiotemporal identifiers,providing an important opportunity for high-precision,real-time,and scientific transformation of social governance.This paper summarized the 5V+5C characteristics and main sources of mobile big data,reviewed research progress in mobile big data-driven human mobility pattern mining and mobile network model construction,and systematically examined its typical applications and frontier advances in key social governance scenarios,including poverty identification,economic assessment,epidemic prevention and control,and emergency response.On this basis,this paper further analyzed the limitations of mobile big data in terms of representativeness,data bias,privacy protection,and usage compliance,and also discussed future research prospects in multi-source data fusion and spatiotemporal semantic alignment,privacy-preserving computation and real-time response in disaster scenarios,artificial society generation,and large model-driven digital simulation.

Unlike warm-start bundle recommendation scenario where models can rely on both user-bundle and user-item interaction data, the cold-start bundle recommendation scenario presents a challenge due to the absence of historical interactions for new bundles. Existing cold-start bundle recommendation methods have made some progresses by using only user-item interactions to predict user preferences for new bundles. However, these methods fail to finely differentiate new bundles using the amount of overlapped items and overlook the distinct interaction patterns between user-bundle and user-item interactions. To address the above limitations, we propose the Heterogeneous Information Network-based Cold-start Bundle Recommendation (HINCBR) framework, which introduces attribute information into user-item interactions to enhance the learned representations. Specifically, we first expand the user-item interactions into a HIN and design a simplified graph neural network to encode diverse interactions within it. Then, we devise a personalized semantic fusion module to learn users' and bundles' representations by adaptively aggregating interaction information. At last, we adopt the contrastive learning approach to further improve the quality of learned representations by aligning user-bundle and user-item interaction view. The experimental results on three real-world datasets validate the superiority of HINCBR over existing state-of-the-art models, achieving up to a 0.0938 and 0.0739 absolute gain in Recall@20 and NDCG@20, respectively. Furthermore, the subsequent ablation study and meta-path analysis demonstrate the effectiveness of the framework design and the utilization of the HIN.

BACKGROUND:Respondent-driven sampling (RDS) is a social network sampling technique used to study hidden behaviours. We used web-based RDS (webRDS) to estimate the prevalence of self-managed abortion (SMA) outside the formal healthcare system in Argentina where abortion was legalised in 2020, but access remains uneven. METHODS:A cross-sectional web survey (February-May 2024) among individuals aged 16-49 years, ever pregnant, and residing in Argentina. Our primary outcome was the proportion of SMA occurring outside the formal healthcare system. Estimates were generated using the RDS II estimator. RESULTS:Seven recruitment chains generated 2437 participants (mean of 19.8 recruitment waves, the longest being 51). We filtered for suspected repeat and ineligible participation and generated RDS estimates for the remaining 1340 participants. The estimated personal network size was 4.9; participants knew an average of 2.7 peers with abortion experience. An estimated 17.1% reported ≥1 abortion, ever in life. Among these, an estimated 20.7% (95% CI 14.2 to 28.0) reported an SMA outside the formal healthcare system; 65.3% before and 24.7% after the legal reform. An estimated 42.2% completed the SMA alone. Reported advantages included autonomy in timing and setting, and support person choice. Disadvantages included concerns about pill quality and uncertainties around the process. CONCLUSIONS:A substantial proportion of women in Argentina with abortion experience have had an SMA outside the formal healthcare system, including post-legalisation. Our findings highlight the need to better address the preferences and needs of those facing unintended pregnancy and the potential of webRDS to study SMA.

Poverty mapping is increasingly important for monitoring Sustainable Development Goal 1 (SDG 1) of the United Nations 2030 Agenda, which aims to end poverty in all its forms everywhere. Yet timely and fine-resolution poverty estimation remains difficult because conventional census- and survey-based approaches are costly, infrequent, and often sparse precisely where deprivation is most severe. As poverty emerges from complex socioeconomic systems shaped by human mobility, social interactions, infrastructure, and economic activities, emerging computational methods and nontraditional data sources have created new opportunities for poverty estimation and mapping. At the intersection of statistical physics, complex systems science, and data science, these approaches enable poverty estimation at finer spatial and temporal resolutions. This review summarizes the main concepts of poverty and the principal frameworks used to measure it, and examines recent advances on poverty estimation and mapping using satellite imagery, mobile phone data, social media data, and multisource data fusion. The review also discusses persistent challenges related to representativeness, transferability across regions, interpretability, and uncertainty quantification. Finally, the review clarifies both the analytical promise and the practical limits of contemporary poverty mapping.

Load balancing-the allocation of work across parallel resources to reduce delay, energy and cost-is a pervasive challenge in science and engineering, from large-scale simulation and data processing to cloud and manufacturing operations. Motivated by the emerging bottleneck in large language model (LLM) serving, we study a particularly stringent regime of load balancing that arises in barrier-synchronized, stateful systems: work cannot be freely migrated and progress is gated by the slowest participant at each step, so heterogeneity and temporal drift in workloads create persistent stragglers and substantial idle time. LLM serving under data-parallel decoding provides a prominent modern instance: in production traces, barrier-induced idle can exceed 40% of compute time per decode step. Here we develop a universal load-balancing principle, which admits a step-wise finite-horizon integer-optimization formulation and yields worst-case guarantees: across LLM decode models and a broader class of non-decreasing workload drift processes, it reduces long-run imbalance by a factor that grows with batch size and system scale. Extensive experiments corroborate the theory, showing substantial improvements in throughput and latency together with reductions in energy consumption. These results provide a general, theoretically grounded framework for load balancing, with immediate implications for sustainable LLM serving and broad relevance to other synchronization-gated resource-allocation problems.

Accurately modeling human mobility is fundamental for understanding and managing urban systems. Among various approaches, the gravity model, derived from Zipf’s law and long used to describe human movement between two locations, is one of the most widely adopted frameworks to predict human movement between areas. However, when applied to intra-city mobility, the model’s parameters show strong sensitivity to the resolution of units, limiting its accuracy and generalizability in real urban environments, which consist of irregular and unequal regions rather than idealized uniform grids. In this study, we systematically examine how spatial scale, the side of spatial unit based on grid partitioning of cities, shapes model behavior and identify two mechanisms: the scale-combination benefit and the scale-disparity penalty. Empirical analyses show that each 1 km increase in the sum of grid sizes improves CPC by 0.012, whereas each 1 km increase in their absolute difference reduces it by 0.020. Further analysis reveals a strong positive linear relationship between model parameters and the logarithm of the product of origin and destination scales. Extending these findings, we develop an area-adaptive gravity model that adapts to regions of varying sizes and is particularly effective for capturing intra-city mobility. Evaluation across grid, irregular, and sub-district partitions using five performance metrics demonstrates that the proposed model achieves superior adaptability and predictive power.

Presently, video communities such as YouTube, bilibili and TikTok have emerged as core fields for information dissemination and public opinion generation. Their embedded user dynamic interaction data support research on public cognitive behavior and content dissemination laws. This study used web crawling technology to construct a complete dataset including 367 video metadata and 2.39 million comment records from Luo Xiang Speaks on Criminal Law-a prominent legal popularization account on the bilibili platform-and systematically explored the temporal evolution patterns of comment interactions in video communities. By establishing a four-dimensional feature system alongside the k-means++ clustering algorithm, this study successfully identified three distinct comment growth patterns (p < 0.001): the burst-decay, the multi-wave oscillation, and the delayed peak. The results of non-parametric tests showed that these three patterns have significant differences in core features (e.g., peak delay time, skewness) and are systematically related to user grade structure, content interaction depth, and release timing. In addition, the user interaction networks of different videos demonstrate significant structural heterogeneity and disassortative mixing, characterized by a highly active minority dominating the discourse, while peripheral nodes gravitate toward high-profile hubs. These findings offer researchers deeper insights into the micro-mechanisms of information dissemination.

Traditional node-based dismantling strategies, which remove nodes along with their associated edges, tend to incur high costs. In contrast, edge-based strategies are more cost-effective but often suffer from low efficiency due to the large number of edges in most networks. To address these challenges, we propose a divide-and-conquer framework that reinterprets network-level dismantling as cluster-level dismantling. Specifically, we integrate community detection with explosive percolation to develop the Community-based Edge Percolation (CEP) algorithm, which targets critical edges whose removal effectively breaks the network into subcritical components, thereby optimizing dismantling efficiency while minimizing costs. Experiments on 38 synthetic networks derived from four different models, as well as on nine empirical networks, show that CEP consistently outperforms state-of-the-art (SOTA) algorithms across nearly all datasets, yielding improvements of up to 30.611 % in fc and 67.108 % in Schneider R. Further analysis indicates that the sets of removed edges identified by CEP have a low correlation with those identified by other benchmarks, underlining its novelty and superior capability in identifying critical edges. Overall, we propose a universal and efficient edge dismantling framework that exhibits substantial advantages in large-scale empirical networks, offering valuable insights into network robustness.

The so-called trust game models asymmetric interactions between trustors, who may offer a certain amount, and trustees who respond by keeping either the entire amount or just a portion of it. Although payoffs depend on these interactions, the members of both groups learn strategy behavior from similar peers. The separation between gaming and learning can be modeled by a bilayer framework, in which individuals learn from neighbors within their own layer but play the game with individuals from the other layer. Realistically, trustee behaviours are heterogeneous. Even if they cooperate and return the donation, its portion could be significantly different. Such diversity can sustain trust even at relatively low levels of trustworthiness. This mechanism is driven by a percolation phenomenon where large clusters of trustworthiness promote trust formation inside the clusters but impede outward diffusion. In contrast, an intermixed distribution facilitates the spread of trust while hindering its local emergence. Consequently, a moderate degree of spatial clustering of trustworthiness maximizes overall trust by balancing connectivity and nucleation. These findings provide delicate insights into the self-organization of spatial and dynamical mechanisms in large-scale populations characterized by asymmetric interactions.

Predicting how single cells respond to perturbations is a central problem in computational biology, with potential relevance to emerging artificial intelligence virtual cell (AIVC) research and drug-discovery efforts. However, substantial variation in perturbation responses across biological contexts and the limited generalizability of current models make prediction across cell types, patients, species, and other contexts particularly challenging. To address this challenge, we present single-cell perturbation inference via latent optimal transport (scPILOT), a query-conditioned framework for transferring responses to previously observed perturbations across biological contexts. scPILOT learns a generative latent representation through discriminator-assisted training and separates perturbation inference into cell-level response estimation from observed contexts and query-specific response transfer using latent optimal transport. Across held-out cell-type, patient, and species benchmarks, scPILOT achieved context-averaged R2 mean/MMD2 values of 0.945/0.137, 0.598/0.025, and 0.853/0.287, respectively. It also maintained strong population-average accuracy in a held-out cell-line benchmark, while complementary analyses indicated that performance was associated with dataset learnability and query-context match. With the continued expansion of single-cell perturbation datasets, scPILOT may provide a practical framework for transferring responses to previously observed perturbations across increasingly diverse biological contexts.

Flow matching (FM) based on Ordinary Differential Equations (ODEs) has achieved significant success in generative tasks. However, it faces several inherent limitations, including an inability to model trajectory intersections, capture delay dynamics, and handle transfer between heterogeneous distributions. These limitations often result in a significant mismatch between the modeled transfer process and real-world phenomena, particularly when key coupling or inherent structural information between distributions must be preserved. To address these issues, we propose Delay Flow Matching (DFM), a new FM framework based on Delay Differential Equations (DDEs). Theoretically, we show that DFM possesses universal approximation capability for continuous transfer maps. By incorporating delay terms into the vector field, DFM enables trajectory intersections and better captures delay dynamics. Moreover, by designing appropriate initial functions, DFM ensures accurate transfer between heterogeneous distributions. Consequently, our framework preserves essential coupling relationships and achieves more flexible distribution transfer strategies. We validate DFM's effectiveness across synthetic datasets, single-cell data, and image-generation tasks.

Background: With the growing emphasis on full-process disease management, efficient post-discharge care has become increasingly critical. Although prior studies have examined follow-up services, resource allocation, and facility location in primary healthcare, model-based optimization of collaborative frameworks between comprehensive hospitals and primary care systems remains limited. Methods: We study a cooperative community hospital selection problem involving contractual cooperation, patient engagement, and follow-up resource allocation. A multi-objective mixed-integer programming model is developed to maximize patient accessibility and minimize total hospital costs, and an NSGA-II-based heuristic is proposed for solution generation. A real-world case study using data from a comprehensive hospital in Chengdu, China, is conducted. Results: The proposed approach produces a Pareto set that quantifies the accessibility-cost trade-off and reveals a knee region with diminishing returns: moderate expansion of cooperating providers substantially improves accessibility, whereas further expansion yields limited additional gains while increasing hospital cost. Sensitivity analyses indicate that cost-related parameters and follow-up frequencies are key drivers of the trade-off. Conclusions: The proposed optimization framework serves as an implementable decision aid for designing hospital-primary care collaboration for post-discharge follow-up: it supports partner selection and capacity planning and indicates levers to improve performance.