Anomaly detection is a critical component for ensuring data quality in data management; however, as continuously collected data introduces unseen normal and anomalous classes, the performance of traditional methods often deteriorates markedly. While some approaches attempt to mitigate this challenge by simulating unseen anomaly distributions, they are constrained by the quality of the generated pseudo-anomalies and fail to solve the core problem of misidentifying unseen normal instances as anomalous. We address these limitations from a novel perspective of normal invariant representation learning by proposing WAlign, which introduces a misclassification-aware weighting mechanism for the normal distribution alignment process. This mechanism mitigates the detrimental influence of misclassified instances and unlabeled anomalies on representation learning for normal instances. As a plug-and-play module, WAlign can be seamlessly integrated into two well-established anomaly detection paradigms. For each paradigm, we instantiate a lightweight base model and conduct extensive experiments on five real-world datasets. Experimental results demonstrate that integrating WAlign improves the AUC-PR by up to 3.7
Incremental Named Entity Recognition (INER) aims to learn new entity types over time while preserving performance on previously seen types, addressing the challenges of catastrophic forgetting and semantic shift. Existing methods mostly combine pseudo-labeling and knowledge distillation to protect old knowledge, achieving promising results. However, they typically randomly initialize the classification head for new entity types, which creates a significant misalignment between the new classifier and features extracted by the backbone model. This misalignment leads to highly unstable gradient updates during early training, slowing convergence and, more critically, increasing the risk of distorting the shared feature space-ultimately harming performance. To address this issue, we propose New Classifier Pre-IniTialization (NCPT), a strategy that generates well-informed initial parameters for new classifiers before formal incremental training begins. Instead of random initialization, NCPT learns a structured transformation from the weights of old classifiers to initialize new classifiers. This ensures that the new classifier starts from a point that is semantically coherent with the existing feature space, enabling smoother optimization and better integration of new knowledge. Furthermore, we introduce a cross-step class similarity mechanism that leverages semantic relatedness computed from old prototypes and new entity types to guide the transformation. These allow the model to balance stability and plasticity. Experiments on three datasets demonstrate that NCPT significantly enhances performance and training stability across incremental steps when integrated into state-of-the-art methods, offering a principled solution to classifier initialization.
Spatio-temporal forecasting plays a vital role in various domains such as transportation, meteorology, and urban management. However, existing approaches often struggle to model complex spatio-temporal dependencies effectively while maintaining computational efficiency, particularly when handling full spatio-temporal sequences. The quadratic time-space complexity of the traditional self-attention mechanism significantly limits its scalability to long-sequence tasks. To address these challenges, we propose LOSTFormer (Linear Orthogonal Spatio-Temporal Transformer) for efficient and accurate spatio-temporal forecasting. LOSTFormer constructs hierarchical embeddings across temporal, spatial, and joint spatio-temporal dimensions to enhance representation learning. Linear attention framework, a novel Learnable Orthogonal Rotation Attention (Lor-Attention) is introduced, which employs the Cayley transform to learn orthogonal rotation matrices and adaptively refine the mapping of Positive Orthogonal Random Features. Our approach can achieve linear computational complexity and effectively capture long-range and cross-dimensional dependencies in spatio-temporal data. Extensive experiments on spatio-temporal datasets demonstrate that LOSTFormer consistently outperforms state-of-the-art methods in prediction accuracy while maintaining efficiency. The code can be found on our GitHub repository( https://github.com/QAQYYC/LOSTFormer .).
Detecting sparse anomalies in long time series is a fundamental challenge in time series anomaly detection, where anomalous events are extremely rare, subtle, and occupy only a minute fraction of the sequence. Excessive false positives in such sparse settings can severely undermine system usability and erode user trust. Most deep learning models are constrained by fixed window scopes and single-mode decision mechanisms, lacking explicit connections to historical windows and sufficient anomaly evaluation criteria, which leads to frequent false alarms. To tackle these limitations, we propose a self-supervised method named C ^3 DR (Cross-window Contrast and Cross-level Discriminative Reconstruction), which unifies representation quality and reconstruction deviation as joint multi-mode evaluation criteria within a single algorithm. Specifically, C ^3 DR models long-range cross-window consistency during encoding, and introduces a discriminative bidirectional cross-attention mechanism during decoding to mitigate overfitting on anomalous regions. On the high-quality UCR benchmark, four-protocol testing shows C ^3 DR reliably outperforms leading deep learning and distance-based detectors, achieving superior event-level accuracy, robust regional detection, and reduced false alarms.
Complex event processing faces significant challenges in real-world scenarios where noisy event streams containing missing events, misordered sequences, and erroneous attribute values are prevalent. Existing approaches address only specific aspects of this problem, with approximate matching techniques handling missing events and out-of-order processing methods addressing sequence misalignment, but none provide a comprehensive solution. This paper introduces K-NFA, a novel fault-tolerant automaton that systematically handles all major error types in complex event matching. Our approach features a compact automaton structure with specialized transitions for event deletion and swap operations, coupled with a dynamic matching algorithm that efficiently processes noisy streams while respecting temporal and attribute constraints. Experimental results show our method improves fault-tolerant matching capacity by 32.91
Accurate causal discovery of railway delay event sequences is crucial for reliable operations in modern complex railway networks, but presents significant challenges due to pronounced non-stationarity and complex network topological dependencies. To address these challenges, we propose the Non-Stationary Spatio-Temporal Self-Attentive Hawkes Process (NSSTHP), a novel framework for learning non-stationary causal structures in railway delay event networks. First, we employ likelihood-based change point detection (PELT) to automatically partition long event sequences into approximately stationary segments. Within each segment, a general spatio-temporally-aware self-attentive Hawkes intensity function is employed for different delay scenarios, which jointly incorporates the spatial topology and temporal dependencies of railway networks under different event delay states. To capture global causal structures, we propose the Symmetric-Balance Thresholding (SBT) method, which adaptively determines the optimal threshold for conversion of real-valued causal matrices to Boolean graphs across segments. Extensive experiments on both synthetic data and real-world datasets demonstrate that NSSTHP significantly outperforms baseline methods in causal edge recovery, structural stability, and interpretability.
Rapid and accurate disaster classification is critical for effective emergency response and resource allocation. Existing approaches to multimodal crisis classification typically presume that the textual and visual elements of a tweet share the same label, thereby overlooking frequent mismatches between modalities and neglecting the contextual links that span multiple tweets. To address these limitations, we introduce GNN-MCC, a novel Graph Neural Network based framework for Multimodal Crisis Classification. Our method constructs a unified graph representation in which individual tweets are decomposed into separate text and image nodes, allowing the model to capture complex intra- and inter-modal relationships. To further reconcile discrepancies between modalities, we incorporate a contrastive attention mechanism that aligns complementary information across text and image channels. Empirical evaluation on the publicly available CrisisMMD dataset demonstrates that GNN-MCC substantially outperforms state-of-the-art baselines under both label-consistent and label-inconsistent conditions. The source code for this paper is available at https://github.com/kliean/GNN-MCC .
Tabular data contains rich structural semantics and plays a crucial role in organizing and manipulating information. Recent methods employ Multi-modal Large Language Models (MLLMs) to address table-related tasks across various modalities of table representations. However, existing studies mainly focus on exploring the table understanding ability of MLLMs using unimodal representations, which limits further exploration of multi-modal representations to enable more effective table reasoning. To better capture structural semantics from the tabular data, this paper introduces the HybrId-modal Preference oPtimizatiOn (HIPPO) model, which represents tables using both text and image, optimizing MLLMs by learning more comprehensive table information from these multiple modalities. Specifically, HIPPO samples MLLM responses from hybrid-modal table representations and designs a modality-consistent sampling strategy to enhance response diversity and mitigate modality bias during Direct Preference Optimization (DPO) training. Experiments on table question answering and table fact verification tasks demonstrate the effectiveness of HIPPO, achieving a 4 https://github.com/NEUIR/HIPPO .
Irregular Multivariate Time Series (IMTS), commonly observed in domains like healthcare and finance, introduce unique challenges to the forecasting task. Existing IMTS forecasting methods primarily address interval irregularity and channel asynchronism. However, they largely overlook sparsity drift, a mismatch in sparsity patterns between the historical and forecasting windows, which disrupts temporal dependency capturing and reduces prediction accuracy. In this work, we introduce a novel indicator, Temporal Observation Density (TOD), and subsequently build History-forEcast spArsity Drift Smoother (HEADS). Specifically, we first quantify sparsity patterns in IMTS with TOD, and derive TOD bias from historical data to forecasts to reflect severity of sparsity drift. Subsequently, we build HEADS, which utilizes TOD bias for smoothing sparsity drift and acts as a plugin seamlessly attached to IMTS forecasting models. Extensive experiments present HEADS makes up to a 5.02 https://github.com/MatthewHu0723/HEADS .
Multi-label text classification (MLTC) faces significant challenges in real-world scenarios due to complex label correlations and the scarcity of labeled data. Existing MLTC methods typically rely on large amounts of labeled data to learn text representations and label dependencies, severely limiting their few-shot ability. While prompt tuning has recently emerged as an effective few-shot learning strategy, its potential in MLTC remains underexplored due to the inherent gap between masked language modeling and multi-label classification. In this paper, we propose a Dual-Verbalizer framework with Label Correlation modeling (DVLC)(The code is available at: https://anonymous.4open.science/r/DVLC ) for few-shot MLTC, integrating multi-label correlation modeling into the prompt tuning framework. Specifically, we introduce text-specific continuous prefixes that incorporate text features into soft prompts, and develop a GCN-based verbalizer with multi-label modeling, which collaborates with a mask verbalizer for prediction. In addition, we apply supervised contrastive learning based on multi-label similarity to improve generalization and mitigate overfitting. Extensive experiments demonstrate the effectiveness of DVLC in few-shot settings.
Synthetic data generation is a key strategy for addressing few-shot time-series tasks. However, existing methods struggle to synthesize diverse, realistic data and lack the fine-grained output control required for effective application in real-world time-series downstream tasks. In this paper, we propose a multi-agent framework that addresses these challenges by leveraging multimodal analysis to capture real-world data patterns from limited samples, while simultaneously enabling fine-grained control over the final output through natural language. To ensure Large Language Models (LLMs) can effectively comprehend the intricate characteristics of time-series data, we design a multimodal analysis agent to interpret both the numerical features and visual patterns of the raw data, translating these into rich textual descriptions. This textual analysis, fused with objectives and domain knowledge refined by an expert agent, creates precise instructions that direct a code-generating agent to programmatically synthesize the data. A discriminator agent then assesses the synthesized data’s alignment with these instructions, providing feedback that drives the framework’s iterative refinement loop. Our framework significantly outperforms baselines, reducing forecasting task MSE by up to 22.1
Multi-turn text-to-SQL aims to translate context-dependent natural language questions into corresponding SQL queries. The two-stage rewrite-then-parse paradigm has emerged as a promising solution that first decontextualizes conversational queries into self-contained questions and then employs powerful off-the-shelf single-turn text-to-SQL parsers for SQL generation. Existing methods mainly optimize the rewriter in isolation, overlooking parser feedback and causing misalignment between the two modules, which undermines effective collaboration. In this paper, we propose Bridge-SQL, which seamlessly adapts single-turn text-to-SQL parsers to conversational scenarios through a question rewriter aligned with parser preferences. Our approach begins by training a base rewriter on synthetic rewrites generated through a bidirectional data synthesis strategy. Following this, the rewriter self-samples multiple candidate rewrites for each question, which are scored by a parser consensus mechanism to construct preference pairs for further alignment. Finally, direct preference optimization is applied to the rewriter under a curriculum learning strategy that progresses from easy to hard examples, ensuring a stable and effective alignment process. Extensive experiments on real-world datasets demonstrate the effectiveness and broad applicability of our proposed method.
Traditional geometry engines can precisely construct boundary representation (BRep) models from procedural design histories, yet their internal representations are often opaque to data-driven systems and difficult to integrate into data management workflows. This opacity limits large-scale analysis, semantic retrieval, and interoperability across heterogeneous CAD environments. To address this challenge, this work introduces BRep-H, a data-centric framework that reformulates the transformation from construction histories to BRep structures as a structured data translation task. The framework decomposes topology and geometry reconstruction into modular language-model components and employs a vector-quantized autoencoder to discretize complex B-spline geometries into symbolic tokens. This formulation establishes a unified schema connecting symbolic modeling with analytic geometry, enabling interpretable, reconstructable, and queryable BRep data. Experiments on the BRep-History dataset demonstrate high structural validity and parametric fidelity, highlighting the potential of language-based schema translation for integrating geometric reasoning into data-driven CAD systems and for bridging AI modeling with database-level design knowledge management.
Multivariate Time series forecasting is a foundation for data-centric decision making in domains such as traffic, energy, and industrial IoT, where accurately capturing complex dependencies between variables and over time is crucial. Among existing modeling paradigms, multi-scale (MS) approaches have been extensively explored for their strong capability to uncover correlations in time series, and recent studies have begun to extend this idea to multi-dependency (MD) modeling. However, existing approaches that couple MS and MD often repetitively apply multi-scale mechanism across different dependency types or simply treat scales as independent branches, which leads to redundant representations and fails to realize true hierarchical interaction across scales. To address this, we propose PyraTSG, a hierarchical scale transition framework that unifies multi-scale and multi-dependency modeling for multivariate time series forecasting. PyraTSG progressively aggregates fine-grained representations into coarse-grained ones through learnable scale transitions, enabling distinct dependencies to be captured at different hierarchical levels, expanding feature propagation ranges while maintaining efficiency. Extensive experiments on multiple benchmark datasets demonstrate that PyraTSG consistently outperforms state-of-the-art baselines, validating the effectiveness of hierarchical coupling for long-term forecasting.
Spatial data classification often faces severe class imbalance and insufficient modeling of spatial dependencies. To address these challenges, we propose Spatially-Aware Generation and Explanation via Large Language Models (SAGE-LLM), a unified framework that integrates LLM-based semantic reasoning with spatial feature construction for imbalanced data. SAGE-LLM performs boundary-aware reasoning on spatially adjacent samples with divergent labels to derive interpretable and discriminative features. It further employs neighborhood-guided aggregation to capture regional context and semantic-driven augmentation to enhance minority-class recognition. Experiments on real-world land-parcel datasets demonstrate that the proposed method consistently improves both F1-score and AUC over strong baselines. These results confirm that integrating spatial reasoning and LLM-based feature generation offers a practical, interpretable, and robust solution for large-scale imbalanced data learning.
Capturing spatio-temporal dependencies is the key challenge for applications such as intelligent transportation and energy management. Contrastive learning-based methods have recently emerged as a promising paradigm to address this challenge for their ability to encode discriminant spatio-temporal dependencies. However, most of these methods assign binary labels to contrastive pairs according to their static relationship, which neglect the dynamic dependencies prevalent in data and fundamentally limit their forecasting ability. To address these limitations, we propose a novel Dynamic Soft Contrastive Learning (DynSCon) framework for spatio-temporal forecasting. Specifically, we design a dynamic dependency aware similarity metric through dynamic graph construction and clustering regularization to capture the variation of dependency patterns across time and space dimensions. Contrastive learning is then conducted on node representations across nodes and time supervised by dynamic soft labels derived from this metric to supplement forecasting loss. Thus the learned representation will be able to reveal the underlying dynamic spatio-temporal dependencies and forecasting quality can be enhanced. Extensive experiments on three datasets spanning from traffic flow forecasting to power generation forecasting demonstrates that DynSCon achieves state-of-the-art performance.
In recent years, the widespread dissemination of fake news across various domains has drawn significant academic attention to multi-domain fake news detection. Among different approaches, single-modal text modification-based detection remains the predominant method in this field. However, existing methods still face two fundamental challenges: (1) the varying semantic meanings of the same word across different domains, known as the domain knowledge shift problem, and (2) the model bias caused by imbalanced data distribution across domains, referred to as the domain distribution imbalance problem. To address these challenges, we proposes the Domain Adaptation Network with Dual-Encoder for Fake News Detection framework(DADE). The framework incorporates a domain adaptation expert network to effectively extract and process domain-specific information, thereby mitigating the domain knowledge shift issue. Additionally, it employs expert-level contrastive learning to explicitly optimize the weighting of features from different domains, resolving the domain distribution imbalance problem. Furthermore, DADE utilizes a dual-encoder architecture to jointly extract textual features, enhancing the richness of input representations. Experimental results on multiple real-world datasets demonstrate that DADE significantly outperforms state-of-the-art methods in detection performance.
Accurate traffic flow prediction plays a vital role in intelligent transportation systems, yet existing deep learning approaches still face challenges in modeling complex long-range spatio-temporal dependencies and handling the inherent non-stationarity of traffic data. To address these issues, we propose CTMamba, a hybrid Transformer–Mamba architecture that integrates cross-temporal fusion, memory-based spatial attention, and selective state space modeling for efficient and adaptive traffic flow prediction. Specifically, the Cross-Temporal Fusion Embedding Module (CTFEM) introduces a temporal smoothing mechanism by combining the moving average of recent time windows with the latest observation, providing a stable reference that mitigates temporal non-stationarity. The Spatial Memory Attention (SMA) module employs a learnable memory matrix to efficiently capture dynamic spatial dependencies among nodes with linear computational complexity. Furthermore, the TMamba module leverages selective state space modeling combined with temporal convolution to capture long-range temporal dependencies while maintaining scalability. Extensive experiments on four real-world traffic datasets demonstrate that CTMamba consistently outperforms state-of-the-art baselines in both prediction accuracy and efficiency. Ablation analyses further confirm that the proposed CTMamba design effectively alleviate data non-stationarity and enhance spatio-temporal representation learning.
Time-series forecasting with long-range, multi-scale dependencies remains challenging. Large language models (LLMs) are promising for sequence reasoning; however, direct prompting often struggles with rapidly varying dynamics and underutilizes frequency-domain structure, while full fine-tuning is costly. We present tsLLM-SG, a fine-tuning framework that combines Dynamic Soft Prompts (DSP) with Gated Frequency Transformation Adapters (GFTA). DSP uses lightweight, learnable embeddings to capture statistical cues and frequency-domain distributions from the input sequence and injects these prompts into the token stream, providing the LLM with critical task-specific context. GFTA acts as a residual spectral filter that adaptively modulates token representations in the frequency domain. On long-horizon and few-shot forecasting benchmarks, tsLLM-SG attains higher accuracy with substantially fewer trainable parameters than recent LLM-based baselines. Ablations and visualizations indicate that DSP and GFTA deliver complementary gains, strengthening temporal representations. Overall, tsLLM-SG offers a scalable, computationally efficient way to leverage LLMs for time-series forecasting.
Time series forecasting has rapidly advanced across diverse domains, and the effective utilization of exogenous variables has become a critical factor in improving predictive performance. In air pollution forecasting, satellite-based spatiotemporal datasets such as those obtained from the Geostationary Environment Monitoring Spectrometer (GEMS) have emerged as essential sources of exogenous information. However, such satellite data are difficult to effectively integrate into time series forecasting models due to issues such as spatial mismatches between satellite grids and ground-based observations. To address this limitation, this study proposes Target Variance Minimization Clustering (TVMC) that represents GEMS spatiotemporal data as informative exogenous variables for ground-level air pollution forecasting. Specifically, a 3D Convolutional AutoEncoder (3D-CAE) is employed to encode GEMS spatiotemporal patches into latent vectors. Our proposed method constructs clusters such that the intra-cluster variance of the forecasting targets is minimized for each latent representation. We theoretically show that the clusters generated by our method exhibit greater mutual information between the cluster assignments and the target variable compared to those formed by conventional clustering methods. Experimental results show that the proposed TVMC-based representation consistently improves performance across diverse time series forecasting models, including state-of-the-art methods.
