Huijie Zhang

Implicit neural representations (INRs) have been widely adopted for volume data modeling in scientific visualization due to their continuous representation capability and flexibility. However, existing INR‐based methods still struggle to accurately capture high‐frequency details due to spectral bias, which can degrade reconstruction quality and affect downstream analysis. To address this issue, we propose a frequency‐decomposed residual learning framework for neural volume representation. Specifically, we decompose the volume data into low‐frequency components and high‐frequency residuals, and jointly model them with a shared INR by reformulating volume representation as a multi‐task learning (MTL) problem. To alleviate potential learning conflicts between different frequency branches while preserving model compactness, we introduce a lightweight feature gating mechanism on top of the shared backbone to adaptively reweight shared features for the two branches, enabling more effective modeling of low‐ and high‐frequency information. Experimental results on multiple scientific datasets demonstrate that our method consistently outperforms existing approaches in reconstruction, super‐resolution, and compression tasks, while maintaining comparable model size and computational efficiency. Further ablation studies validate the effectiveness of the proposed frequency‐decomposition and feature gating mechanism. Source code is publicly available at https://github.com/LiuLan2001/FDNR .

Balancing scalability and structural integrity in multi-view multi-label feature selection remains a significant challenge. Current anchor-based frameworks typically rely on isolated pairwise graphs. This inherently separates structural modeling from semantic fitting, failing to integrally couple cross-view high-order correlations with sample-label dependencies. To break this bottleneck, we propose a unified scalable framework which strategically migrates high-order structural modeling from the full-sample space to a compact anchor space. Specifically, we first introduce a mask-guided joint learning paradigm with bidirectional information propagation, enabling pseudo-labels to simultaneously incorporate precise supervision and topological structures. Furthermore, a tensor-coupled consensus mechanism employs low-rank constraints to extract latent pure structures and explicitly capture global high-order correlations across views. Building upon this cross-view consensus, a novel scalable hypergraph-bipartite regularization factorizes the hypergraph Laplacian to seamlessly bridge the multi-view structures with the multi-label semantics. Consequently, it captures nonlinear manifolds with linear computational complexity with respect to the sample size, providing reliable structural guidance for accurate feature evaluation. Extensive experiments demonstrate the superiority of our method, validating its effectiveness in characterizing high-order correlations for discriminative feature selection.

Selecting appropriate temporal intervals for analyzing dynamic graphs is a critical but non‐trivial task. Poorly chosen intervals can obscure key structural changes and lead to flawed interpretations, while existing tools often lack effective guidance, forcing users into a cycle of tedious manual adjustments and trial‐and‐error. We present Multi2‐Vis, a visual analytics system that reframes temporal segmentation as a human‐in‐the‐loop interactive refinement process. Multi2‐Vis employs an Evaluate‐Recommend‐Refine loop: it uses structure‐aware metrics to automatically identify suboptimal segments and then presents optimized alternatives in a multi‐scale visual interface, guiding users toward informed decisions. Through two case studies, a quantitative experiment and a controlled user study, we demonstrate that our guided workflow significantly improves analysts' analytical efficiency and conclusion quality compared to traditional methods. By transforming temporal segmentation from a rigid prerequisite into a flexible, interactive dialogue, Multi2‐Vis provides a more interpretable solution for dynamic graph analysis.

Implicit neural representations (INRs) have become a common surrogate representation for scientific data due to their continuous representation and flexibility. Since INRs are lossy representations that introduce errors in data reconstruction, it is crucial to convey the uncertainty of the results to scientists to prevent them from being misled. However, current uncertainty quantification schemes generally face the challenge of compromising data reconstruction accuracy while maintaining the compactness of INRs. To address this issue, we formulate data reconstruction and uncertainty quantification as a multi‐task learning (MTL) problem and treat uncertainty quantification as an auxiliary task to the reconstruction task. Unlike conventional auxiliary tasks which aim to improve the performance of the primary task, our objective is to obtain highly credible uncertainty estimates while minimizing the negative impact on the primary reconstruction task. To this end, we address this MTL problem with a Multi‐gate Mixture‐of‐Experts (MMoE) architecture that partitions the network into multiple sub‐networks and employs distinct gates to weight their outputs for each task. Furthermore, we propose a virtual expert mechanism to mitigate the trade‐off between model flexibility and sub‐network performance. To ensure that both tasks are fully optimized, we propose a loss weighting method combining dynamic and static weights. The dynamic weights adapt based on task convergence, while the static weights scale the losses and balance their relative importance. We comprehensively evaluated existing uncertainty quantification methods and compared them with our method. The results indicate that our method achieves higher reconstruction accuracy and more reliable uncertainty than existing methods. Source code is publicly available at https://github.com/LiuLan2001/MTL‐SV .

Multi-View Multi-Label Learning is pivotal for visual content understanding yet faces two fundamental bottle-necks: (i) Structural Decoupling, where prevalent matrix flattening severs intrinsic high-order correlations between visual features and semantic concepts; and (ii) Semantic Uniformity Bias, where the neglect of label topology heterogeneity leads to overfitting on ambiguous annotations. To address these, we propose the Adaptive HypergrAph-Induced Dual-Structure Tensor Embedding (AHATE) framework. Drawing conceptual inspiration from tensor spectral graph theory, we introduce a novel Dual-Structure Tensor Nuclear Norm (DS-TNN). Unlike loose additive regularization, this mechanism functions as a high-order spectral filter in the Fourier domain, explicitly coupling feature manifolds with label topology to rigorously enforce consensus while suppressing structural noise. Furthermore, grounded in network centrality theory, we design an Adaptive Hypergraph-Weighted Loss. By leveraging PageRank to quantify topological importance, this strategy prioritizes high-confidence structural anchors over ambiguous instances. Finally, a joint sparse projection mechanism is incorporated to eliminate redundancy. Extensive experiments demonstrate that AHATE achieves statistically significant superiority. Specifically, compared to the second-best methods, it improves Average Precision by up to 5.52% and reduces Hamming Loss by up to 4.40%, while achieving the lowest Hamming Loss of 0.0137 on Corel5k. The full source code, including preprocessing and configuration scripts, is available at https://github.com/hpinty/AHATE.git.