Bohan Lin, Kuo Yang, Zelin Tan, Yingchuan Lai, Chen Zhang, Guibin Zhang, Xinlei Yu, Miao Yu, Xu Wang, Yudong Zhang, Yang Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Due to the profound impact of air pollution on human health, livelihoods, and economic development, air quality forecasting is of paramount significance. Initially, we employ the causal graph method to scrutinize the constraints of existing research in comprehensively modeling the causal relationships between the air quality index (AQI) and meteorological features. In order to enhance prediction accuracy, we introduce a novel air quality forecasting model, AirCade, which incorporates a causal decoupling approach. AirCade leverages a spatiotemporal module in conjunction with knowledge embedding techniques to capture the internal dynamics of AQI. Subsequently, a causal decoupling module is proposed to disentangle synchronous causality from past AQI and meteorological features, followed by the dissemination of acquired knowledge to future time steps to enhance performance. Additionally, we introduce a causal intervention mechanism to explicitly represent the uncertainty of future meteorological features, thereby bolstering the model’s robustness. Our evaluation of AirCade on an open-source air quality dataset demonstrates over 20% relative improvement over state-of-the-art models. Our source code is available at https://github.com/PoorOtterBob/AirCade.

Long-term time series forecasting (LTSF) has traditionally relied on large parameters to capture extended temporal dependencies, resulting in substantial computational costs and inefficiencies in both memory usage and processing time. However, time series data, unlike high-dimensional images or text, often exhibit temporal pattern similarity and low-rank structures, especially in long-term horizons. By leveraging this structure, models can be guided to focus on more essential, concise temporal data, improving both accuracy and computational efficiency. In this paper, we introduce TimeBase, an ultra-lightweight network to harness the power of minimalism in LTSF. TimeBase 1) extracts core basis temporal components and 2) transforms traditional point-level forecasting into efficient segment-level forecasting, achieving optimal utilization of both data and parameters. Extensive experiments on diverse real-world datasets show that TimeBase achieves remarkable efficiency and secures competitive forecasting performance. Additionally, TimeBase can also serve as a very effective plug-and-play complexity reducer for any patch-based forecasting models. Code is available at https://github.com/hqh0728/TimeBase.

Discovering regularities from spatiotemporal systems can benefit various scientific and social planning. Current spatiotemporal learners usually train an independent model from a specific source data that leads to limited transferability among sources, where even correlated tasks requires new design and training. The key towards increasing cross-domain knowledge is to enable collective intelligence and model evolution. In this paper, inspired by neuroscience theories, we theoretically derive the increased information boundary via learning cross-domain collective intelligence and propose a Synaptic EVOlutional spatiotemporal network, SynEVO, where SynEVO breaks the model independence and enables cross-domain knowledge to be shared and aggregated. Specifically, we first re-order the sample groups to imitate the human curriculum learning, and devise two complementary learners, elastic common container and task-independent extractor to allow model growth and task-wise commonality and personality disentanglement. Then an adaptive dynamic coupler with a new difference metric determines whether the new sample group should be incorporated into common container to achieve model evolution under various domains. Experiments show that SynEVO improves the generalization capacity by at most 42\% under cross-domain scenarios and SynEVO provides a paradigm of NeuroAI for knowledge transfer and adaptation. Code available at [https://github.com/Rodger-Lau/SynEVO](https://github.com/Rodger-Lau/SynEVO).

The out-of-distribution (OOD) generalization challenge is a longstanding problem in graph learning. Through studying the fundamental cause of data distribution shift, i.e., the changes of environments, significant progress has been achieved in addressing this issue. However, we observe that existing works still fail to effectively address complex environment shifts. Existing practices place excessive attention on extracting causal subgraphs, inevitably treating spurious subgraphs as environment variables. While spurious subgraphs are controlled by environments, the space of environment changes encompass more than the scale of spurious subgraphs. Therefore, existing efforts have a limited inference space for environments, leading to failure under severe environment changes. To tackle this issue, we propose a negative inference graph OOD framework (NeGo) to broaden the inference space for environment factors. Inspired by the successful practice of prompt learning in capturing underlying semantics and causal associations in large language models, we design a negative prompt environment inference to extract underlying environment information. We further introduce the environment-enhanced invariant subgraph learning to effectively exploit inferred environment embedding, ensuring the robust extraction of causal subgraph in the environment shifts. Lastly, we conduct a comprehensive evaluation of NeGo on real-world datasets and synthetic datasets across domains. NeGo outperforms baselines on nearly all datasets, which verify the effectiveness of our framework.

Dynamic graphs are ubiquitous in the real world, presenting the temporal evolution of individuals within spatial associations. Recently, dynamic graph learning research is flourishing, striving to more effectively capture evolutionary patterns and spatial correlations. However, existing methods still fail to address the issue of concept shift in dynamic graphs. Concept shift manifests as a distribution shift in the mapping pattern between historical observations and future evolution. The reason is that some environment variables in dynamic graphs exert varying effects on evolution patterns, but these variables are not effectively captured by the models, leading to the intractable concept shift issue. To tackle this issue, we propose a State-driven environment inference framework (Samen) to achieve a dynamic graph learning framework equipped with concept generalization ability. Firstly, we propose a two-stage environment inference and compression strategy. From the perspective of state space, we introduce a prefix-suffix collaborative state learning mechanism to bidirectionally model the spatio-temporal states. A hierarchical state compressor is further designed to refine the state information resulting in concept shift. Secondly, we propose a skip-connection spatio-temporal prediction module, which effectively utilizes the inferred environments to improve the model's generalization capability. Finally, we select seven datasets from different domains to validate the effectiveness of our model. By comparing the performance of different models on samples with concept shift, we verify that our Samen gains generalization capacity that existing methods fail to capture.

Time series forecasting is a critical and complex task, characterized by diverse temporal patterns, varying statistical properties, and different prediction horizons across datasets and domains. Conventional approaches typically rely on a single, unified model architecture to handle all forecasting scenarios. However, such monolithic models struggle to generalize across dynamically evolving time series with shifting patterns. In reality, different types of time series may require distinct modeling strategies. Some benefit from homogeneous multi-scale forecasting awareness, while others rely on more complex and heterogeneous signal perception. Relying on a single model to capture all temporal diversity and structural variations leads to limited performance and poor interpretability. To address this challenge, we propose a Multi-Agent Forecasting System (MAFS) that abandons the one-size-fits-all paradigm. MAFS decomposes the forecasting task into multiple sub-tasks, each handled by a dedicated agent trained on specific temporal perspectives (e.g., different forecasting resolutions or signal characteristics). Furthermore, to achieve holistic forecasting, agents share and refine information through different communication topology, enabling cooperative reasoning across different temporal views. A lightweight voting aggregator then integrates their outputs into consistent final predictions. Extensive experiments across 11 benchmarks demonstrate that MAFS significantly outperforms traditional single-model approaches, yielding more robust and adaptable forecasts.

The effectiveness of Spatiotemporal Graph Neural Networks (STGNNs) critically hinges on the quality of the underlying graph topology. While end-to-end adaptive graph learning methods have demonstrated promising results in capturing latent spatiotemporal dependencies, they often suffer from high computational complexity and limited expressive capacity. In this paper, we propose MAGE for efficient spatiotemporal forecasting. We first conduct a theoretical analysis demonstrating that the ReLU activation function employed in existing methods amplifies edge-level noise during graph topology learning, thereby compromising the fidelity of the learned graph structures. To enhance model expressiveness, we introduce a sparse yet balanced mixture-of-experts strategy, where each expert perceives the unique underlying graph through kernel-based functions and operates with linear complexity relative to the number of nodes. The sparsity mechanism ensures that each node interacts exclusively with compatible experts, while the balancing mechanism promotes uniform activation across all experts, enabling diverse and adaptive graph representations. Furthermore, we theoretically establish that a single graph convolution using the learned graph in MAGE is mathematically equivalent to multiple convolutional steps under conventional graphs. We evaluate MAGE against advanced baselines on multiple real-world spatiotemporal datasets. MAGE achieves competitive performance while maintaining strong computational efficiency.

The World Wide Web thrives on intelligent services that depend heavily on accurate time series forecasting to navigate dynamic and evolving environments. Due to the partially-observed nature of real world, exclusively focusing on the target of interest, so-called endogenous variables, is insufficient for accurate forecasting, especially in web systems that are susceptible to external influences. Thus, utilizing exogenous variables to harness external information, i.e., forecasting with exogenous variable (FEV), is imperative. Nevertheless, as the external environment is complex and ever-evolving, inadequately capturing external influences can even lead to learning spurious correlations and invalid prediction. Fortunately, recent studies have demonstrated that large language models (LLMs) exhibit exceptional recognition capabilities across open real-world systems, including a deep understanding of exogenous environments. However, it is difficult to directly apply LLMs for FEV due to challenges of task activation, exogenous knowledge extraction, and feature space alignment. In this work, we devise ExoLLM, an LLM-driven method to sufficiently utilize Exogenous variables for time series forecasting. We begin by Meta-task Instruction to activate the knowledge transfer of LLM from natural language processing to FEV. To comprehensively understand the intricate and hierarchical influences of exogenous variables, we propose Multi-grained Prompts, encompassing diverse external influences, including natural attributes, trend correlations, and period relationships between two types of variables. Additionally, a Dual TS-Text Attention is devised to bridge the feature gap between text and numeric data in LLM. Evaluation on real-world datasets demonstrates ExoLLM's superiority in exploiting exogenous information for forecasting with open-world language knowledge.

Spatiotemporal (ST) learning has become a crucial technique to enable smart cities and sustainable urban development. Current ST learning models capture the heterogeneity via various spatial convolution and temporal evolution blocks. However, rapid urbanization leads to fluctuating distributions in urban data and city structures, resulting in existing methods suffering generalization and data adaptation issues. Despite efforts, existing methods fail to deal with newly arrived observations, and the limitation of those methods with generalization capacity lies in the repeated training that leads to inconvenience, inefficiency and resource waste. Motivated by complementary learning in neuroscience, we introduce a prompt-based complementary spatiotemporal learning termed ComS2T, to empower the evolution of models for data adaptation. We first disentangle the neural architecture into two disjoint structures, a stable neocortex for consolidating historical memory, and a dynamic hippocampus for new knowledge update. Then we train the dynamic spatial and temporal prompts by characterizing distribution of main observations to enable prompts adaptive to new data. This data-adaptive prompt mechanism, combined with a two-stage training process, facilitates fine-tuning of the neural architecture conditioned on prompts, thereby enabling efficient adaptation during testing. Extensive experiments validate the efficacy of ComS2T in adapting various spatiotemporal out-of-distribution scenarios while maintaining effective inferences.

Subgraph learning has dominated most practices of improving the expressive power of Message Passing Neural Networks (MPNNs). Existing subgraph discovery policies can be classified into node-based and partition-based, which both achieve impressive performance in most scenarios. However, both mainstream solutions still face a subgraph degradation trap. Subgraph degradation is reflected in the phenomenon that the subgraph-level methods fail to offer any benefits over node-level MPNNs. In this work, we empirically investigate the existence of the subgraph degradation issue and introduce a unified perspective, perfect reconstruction, to provide insights for improving two lines of methods. We further propose a subgraph learning strategy guided by the principle of perfect reconstruction. To achieve this, two major issues should be well-addressed, i.e.,(i) how to ensure the subgraphs to possess with ‘perfect’ information? (ii) how to guarantee the ‘reconstruction’ power of obtained subgraphs?First, we propose a subgraph partition strategyRayleigh-resistanceto extract non-overlap subgraphs by leveraging the graph spectral theory. Second, we put forward aQuerymechanism to achieve subgraph-level equivariant learning, which guarantees subgraph reconstruction ability. These two parts,perfect subgraph partitionandequivariant subgraph learningare seamlessly unified as a novelRayleigh-resistanceEquivariantSubgraph learningarchitecture (RayE-Sub). Comprehensive experiments on both synthetic and real datasets demonstrate that our approach can consistently outperform previous subgraph learning architectures.

Spatiotemporal learning plays a crucial role in mobile computing techniques to empower smart cites. While existing research has made great efforts to achieve accurate predictions on the overall dataset, they still neglect the significant performance heterogeneity across samples. In this work, we designate the performance heterogeneity as the reason for unfair spatiotemporal learning, which not only degrades the practical functions of models, but also brings serious potential risks to real-world urban applications. To fix this gap, we propose a model-independent Fairness-aware framework for SpatioTemporal Graph learning (FairSTG), which inherits the idea of exploiting advantages of well-learned samples to challenging ones with collaborative mix-up. Specifically, FairSTG consists of a spatiotemporal feature extractor for model initialization, a collaborative representation enhancement for knowledge transfer between well-learned samples and challenging ones, and fairness objectives for immediately suppressing sample-level performance heterogeneity. Experiments on four spatiotemporal datasets demonstrate that our FairSTG significantly improves the fairness quality while maintaining comparable forecasting accuracy. Case studies show FairSTG can counter both spatial and temporal performance heterogeneity by our sample-level retrieval and compensation, and our work can potentially alleviate the risks on spatiotemporal resource allocation for underrepresented urban regions.

Data-driven molecular computing has become an increasingly popular topic in AI for molecular and bioinformatic science. Current molecular modeling exploits the Graph Neural Networks (GNNs) to achieve the representation but they mostly fail to generalize to out-of-distribution (OOD) samples. Even though recent advances on OOD-oriented graph learning discovered the invariant rationale on graphs, they still ignore two important issues, i.e., 1) the increasing number and types of molecules expand patterns of environments on graphs, resulting in failures of invariant rationale based models, 2) the associations between discovered molecular subgraphs and corresponding properties are complex where causal substructures cannot fully interpret the labels. To this end, we propose an environment causal learning framework, EMoNet, to tackle the unresolved OOD challenge in molecular science. Specifically, we model the graph environments via bypassing invariant subgraphs. We first incorporate chemistry principle into our graph growth generator to imitate environment growth, and then devise an environment-GIB to squash out environment and finally introduce a cross-attention causal aggregation, allowing dynamic interactions between environments and invariances. We perform experiments on seven datasets and extensive experiments demonstrate strong generalization ability of EMoNet.

Out-of-distribution (OOD) generalization issue is a well-known challenge within deep learning tasks. In dynamic graphs, the change of temporal environments is regarded as the main cause of data distribution shift. While numerous OOD studies focusing on environment factors have achieved remarkable performance, they still fail to systematically solve the two issue of environment inference and utilization. In this work, we propose a novel dynamic graph learning model named EpoD based on prompt learning and structural causal model to comprehensively enhance both environment inference and utilization. Inspired by the superior performance of prompt learning in understanding underlying semantic and causal associations, we first design a self-prompted learning mechanism to infer unseen environment factors. We then rethink the role of environment variable within spatio-temporal causal structure model, and introduce a novel causal pathway where dynamic subgraphs serve as mediating variables. The extracted dynamic subgraph can effectively capture the data distribution shift by incorporating the inferred environment variables into the node-wise dependencies. Theoretical discussions and intuitive analysis support the generalizability and interpretability of EpoD. Extensive experiments on seven real-world datasets across domains showcase the superiority of EpoD against baselines, and toy example experiments further verify the powerful interpretability and rationality of our EpoD.

Human mobility forecasting is the key to energizing considerable mobile computing services. However, we find that the collective mobility suffers the spatiotemporal heterogeneity issue and therefore leads to inferior performances of conventional homogeneous aggregations. Given two fundamental factors, i.e., data and objectives in machine learning, we propose to counter such heterogeneity by improving data utilization and optimization objectives. 1) From data utilization perspective, we discover that such heterogeneity is inherently induced by mobility-related context factors and thus these factors can be exploited to learn heterogeneous mobility patterns. 2) From the optimization perspective, the dependencies among output elements, which give another prior to learning, can extract heterogeneous correlations within output sequences. Specifically, we propose a novel Context-Directional SpatioTemporal Graph Network (CD-STGNet), which tackles the above-mentioned heterogeneity, for achieving accurate mobility predictions. Firstly, we improve data utilization by inputting the encoded context-wise interactions to a direction field learner, which realizes directional spatial aggregations. Secondly, regarding series learning and optimization objectives, a context-trend highway is designed to enable context-aware temporal learning while two regularization objectives are proposed to keep the correlations among predicted elements consistent with the ground-truth. Experiments demonstrate that CD-STGNet surpasses competitive baselines by 13% to 22% and boosts the interpretability of context-directional learning.

Subgraph learning has received considerable attention in its capacity of interpreting important structural information for predictions. Existing subgraph learning usually exploits statistics on predefined structures e.g., node degrees, occurrence frequency, to extract subgraphs, or refine the contents via only capturing label-relevant information with node-level sampling. Given diverse subgraph patterns, and mutual independence with local correlations on graphs, current solutions on subgraph learning still have two limitations in extraction and refinement stages. 1) The universality of extracting substructure patterns across domains is still lacking, 2) node-level sampling in refinement will distort the original local topology and none explicit guidance eliminating redundant information contribute to inefficiency issue. In this paper, we propose a unified subgraph learning scheme, Poly-Pivot Graph Neural Network (P2GNN) where we designate the centric node of each subgraph as the pivot. In the extraction stage, we present a general subgraph extraction principle, i.e., Local; Asymmetry between the centric and affiliated nodes. To this end, we asymmetrically model the similarity between each pair of nodes with random walk and quantify mutual affiliations in Affinity Propagation architecture, to extract subgraph structures. In the refinement, we devise a subgraph-level exclusion regularization to squash the target-independent information by considering mutual relations across subgraphs, cooperatively preserving a support set of subgraphs and facilitating the refinement process for graph representation. Empirical experiments on diverse web and biological graphs reveal 1.1%~7.3% improvements against best baselines, and visualized case studies prove the universality and interpretability of our P2GNN.

Spatiotemporal (ST) learning has become a crucial technique for urban digitalization. Due to expansions and dynamics of cities, current spatiotemporal models are inclined to suffer distribution shifts between training and testing sets, leading to the OOD delimma. However, few studies focus on such OOD problem in temporal regressions, let alone spatiotemporal learning. Spatiotemporal data usually reveals segment-level heterogeneity within periodicity and complex spatial dependencies, posing challenges to invariance extraction. In this paper, we find that ST relations make sense for generalization and devise a Causal ST learning framework, CauSTG, which enables invariant relation transferred to OOD scenarios. Specifically, we take temporal steps as environments, and transform spatial-temporal relations into learnable parameters. To tackle heterogeneity in periodicity, we partition temporal steps into sub-environments by identifying distinctive trend patterns, enabling re-organized samples trained separately. To extract invariance within ST observations, we propose a spatiotemporal consistency learner and a hierarchical invariance explorer to jointly filter out stable relations. Our spatiotemporal learner quantifies bi-directional spatial consistency and extracts disentangled seasonal-trend patterns via trainable parameters. Further, the hierarchical invariance explorer constructs variation-based filter to achieve both local and global invariances. Experiments reveal that CauSTG can increase at most 10.26% performance against best baselines, and visualized invariant relations can well interpret the physical rationales. The appendix and codes can be available in our Github repository.

Spatiotemporal forecasting is an imperative topic in data science due to its critical applications in smart cities. Existing works mostly perform consecutive predictions of following steps with observations continuously obtained, where nearest observations can be exploited as the key knowledge for status estimation. However, the practical issues of early activity planning and sensor failures elicit a new task, non-consecutive forecasting. In this paper, we define spatiotemporal learning systems with missing observations as Grey Spatiotemporal Systems (G2S) and propose a Factor-Decoupled learning framework for G2S to hierarchically decouple multi-level factors, and enable flexible aggregations with uncertainty estimations. We especially select representative sequences to capture periodicity and instantaneous variations, and infer the non-consecutive future statuses under expected exogenous factors, compensating the missing observations. Given the inherent incompleteness and critical applications of G2S, a DisEntangled Uncertainty Quantification is put forward, to identify two types of uncertainty for model interpretations and robustness promotions. Experiments demonstrate that our solution can promote the performance by at least 8.50% on early planning and 2.01%-18.00% on sensor failures. The appendix of this paper can be found at https://github.com/zzyy0929/SDM-G2S.