Selected work

Representative Papers

CASE-NET: Deep Spatio-Temporal Representation Learning via Causal Attention and Channel Recalibration for Multivariate Time Series Classification

May 21, 2026

This work addresses representation distortion and noise sensitivity in multivariate time series classification, which arise from non-causal modeling and channel independence assumptions. To this end, the authors propose a Structured Manifold Preconditioning Network that jointly integrates causal self-attention, causal convolutions, and an adaptive channel recalibration mechanism—enabling the first unified modeling of temporal causal structure and channel-wise information bottlenecks. This approach effectively mitigates temporal confounding under non-stationary dynamics and suppresses noise in latent representations. The method achieves new state-of-the-art performance across four tasks spanning six heterogeneous domains, attaining 98.6% accuracy on the AWR dataset and demonstrating exceptional robustness in non-stationary scenarios.

0 citationsRead paper

L-Drive: Beyond a Single Mapping-Latent Context Drives Time Series Forecasting

May 17, 2026

Existing time series forecasting methods often exhibit delayed responses under distribution shifts or regime changes, leading to error accumulation and reduced reliability around turning points. To address this limitation, this work proposes the L-Drive framework, which explicitly models dynamic system evolution by integrating time-varying implicit context with gated incremental representations for the first time. Furthermore, L-Drive introduces piecewise-shared relative positional basis functions to enhance intra-segment structural modeling and mitigate overfitting caused by reliance on absolute positional encoding. Experimental results demonstrate that L-Drive achieves significantly improved prediction accuracy across multiple datasets while maintaining computational efficiency.

0 citationsRead paper

MHMamba: Multi-Head Mamba for 3D Brain Tumor Segmentation

May 15, 2026

This work addresses the limitations of convolutional neural networks in capturing long-range dependencies and the high computational complexity of Transformers in 3D brain tumor segmentation by proposing a U-Net-based multi-head State Space Model (Mamba) approach. The method parallelizes SSM heads along the channel dimension and integrates them via residual aggregation, augmented with a channel-spatial calibration module and an adaptive skip connection fusion mechanism. This design maintains linear computational complexity while enhancing multi-modal training stability, boundary consistency, and sensitivity to small lesions. Evaluated on the BraTS2021 and BraTS2023 datasets, the proposed model achieves substantial improvements in overall segmentation accuracy, particularly excelling in delineating the tumor core and detecting subtle enhancing regions.

0 citationsRead paper

PESD-TSF: A Period-Aware and Explicit Structured Decomposition Framework for Long-Term Time Series Forecasting

May 14, 2026

This work addresses key limitations in deep time series forecasting models, which often suffer from degraded periodicity awareness, entangled trend and noise components, and disrupted dynamic inter-variable relationships due to the channel independence assumption when network depth increases. To overcome these issues, the authors propose a structured decomposition framework that preserves deep periodic structures through multiplicative periodic gating, explicitly disentangles trends from high-frequency components via multi-scale decoupled encoding, and captures global topological dependencies among variables using cross-scale collaborative attention (CSCA) enhanced with RLC regularization—incorporating orthogonality and physics-informed consistency constraints. The method achieves state-of-the-art performance across multiple benchmark datasets, demonstrating substantial accuracy gains in complex, tightly coupled multivariate long-term forecasting scenarios and confirming its superior structural modeling capacity and generalization ability.

0 citationsRead paper

What If We Let Forecasting Forget? A Sparse Bottleneck for Cross-Variable Dependencies

May 08, 2026

This work addresses the challenge of modeling inter-variable dependencies in multivariate time series forecasting, where dense cross-variable interactions often introduce spurious correlations and lead to oversmoothed representations. To overcome this, the authors propose MS-FLOW, a novel framework that, for the first time, formulates cross-variable interaction as a bandwidth-constrained information flow. By employing a sparse routing mechanism under a strict communication budget, MS-FLOW transmits only the most critical dependency signals, thereby realizing an “effective rather than excessive” interaction paradigm. Integrating selective sparse routing with capacity-limited information flow, the method constructs a lightweight yet efficient dependency learning architecture. Evaluated on twelve real-world benchmarks, MS-FLOW achieves state-of-the-art forecasting accuracy while uncovering fewer but more reliable inter-variable dependencies.

0 citationsRead paper

Recent publications

Latest Papers

CASE-NET: Deep Spatio-Temporal Representation Learning via Causal Attention and Channel Recalibration for Multivariate Time Series Classification

May 21, 2026

This work addresses representation distortion and noise sensitivity in multivariate time series classification, which arise from non-causal modeling and channel independence assumptions. To this end, the authors propose a Structured Manifold Preconditioning Network that jointly integrates causal self-attention, causal convolutions, and an adaptive channel recalibration mechanism—enabling the first unified modeling of temporal causal structure and channel-wise information bottlenecks. This approach effectively mitigates temporal confounding under non-stationary dynamics and suppresses noise in latent representations. The method achieves new state-of-the-art performance across four tasks spanning six heterogeneous domains, attaining 98.6% accuracy on the AWR dataset and demonstrating exceptional robustness in non-stationary scenarios.

0 citationsRead paper

L-Drive: Beyond a Single Mapping-Latent Context Drives Time Series Forecasting

May 17, 2026

Existing time series forecasting methods often exhibit delayed responses under distribution shifts or regime changes, leading to error accumulation and reduced reliability around turning points. To address this limitation, this work proposes the L-Drive framework, which explicitly models dynamic system evolution by integrating time-varying implicit context with gated incremental representations for the first time. Furthermore, L-Drive introduces piecewise-shared relative positional basis functions to enhance intra-segment structural modeling and mitigate overfitting caused by reliance on absolute positional encoding. Experimental results demonstrate that L-Drive achieves significantly improved prediction accuracy across multiple datasets while maintaining computational efficiency.

0 citationsRead paper

MHMamba: Multi-Head Mamba for 3D Brain Tumor Segmentation

May 15, 2026

This work addresses the limitations of convolutional neural networks in capturing long-range dependencies and the high computational complexity of Transformers in 3D brain tumor segmentation by proposing a U-Net-based multi-head State Space Model (Mamba) approach. The method parallelizes SSM heads along the channel dimension and integrates them via residual aggregation, augmented with a channel-spatial calibration module and an adaptive skip connection fusion mechanism. This design maintains linear computational complexity while enhancing multi-modal training stability, boundary consistency, and sensitivity to small lesions. Evaluated on the BraTS2021 and BraTS2023 datasets, the proposed model achieves substantial improvements in overall segmentation accuracy, particularly excelling in delineating the tumor core and detecting subtle enhancing regions.

0 citationsRead paper

PESD-TSF: A Period-Aware and Explicit Structured Decomposition Framework for Long-Term Time Series Forecasting

May 14, 2026

This work addresses key limitations in deep time series forecasting models, which often suffer from degraded periodicity awareness, entangled trend and noise components, and disrupted dynamic inter-variable relationships due to the channel independence assumption when network depth increases. To overcome these issues, the authors propose a structured decomposition framework that preserves deep periodic structures through multiplicative periodic gating, explicitly disentangles trends from high-frequency components via multi-scale decoupled encoding, and captures global topological dependencies among variables using cross-scale collaborative attention (CSCA) enhanced with RLC regularization—incorporating orthogonality and physics-informed consistency constraints. The method achieves state-of-the-art performance across multiple benchmark datasets, demonstrating substantial accuracy gains in complex, tightly coupled multivariate long-term forecasting scenarios and confirming its superior structural modeling capacity and generalization ability.

0 citationsRead paper

What If We Let Forecasting Forget? A Sparse Bottleneck for Cross-Variable Dependencies

May 08, 2026

This work addresses the challenge of modeling inter-variable dependencies in multivariate time series forecasting, where dense cross-variable interactions often introduce spurious correlations and lead to oversmoothed representations. To overcome this, the authors propose MS-FLOW, a novel framework that, for the first time, formulates cross-variable interaction as a bandwidth-constrained information flow. By employing a sparse routing mechanism under a strict communication budget, MS-FLOW transmits only the most critical dependency signals, thereby realizing an “effective rather than excessive” interaction paradigm. Integrating selective sparse routing with capacity-limited information flow, the method constructs a lightweight yet efficient dependency learning architecture. Evaluated on twelve real-world benchmarks, MS-FLOW achieves state-of-the-art forecasting accuracy while uncovering fewer but more reliable inter-variable dependencies.

0 citationsRead paper