Diffusion Transformers (DiTs) achieve remarkable performance within image generation. Conventionally, DiTs are constructed by stacking serial isotropic transformers, which face significant quadratic computational cost. However, through empirical analysis, we find that DiTs do not rely as heavily on long-distance information as previously believed. In fact, most layers exhibit significant redundancy in long-distance computation. Additionally, conventional attention mechanisms suffer from low-frequency inertia, limiting their efficiency. To address these issues, we propose Pseudo Shifted Window Attention (PSWA), which fundamentally mitigates long-distance attention redundancy. PSWA achieves moderate global-local information through Static Window Attention. It further utilizes a high-frequency bridging branch to enrich the high-frequency information and strengthen inter-window connections. Furthermore, we pioneer the concept of Kth-order Attention and propose the Progressive Coverage Channel Allocation (PCCA) strategy that captures high-order attention by reallocating the existing channel budget. Based on these innovations, we propose a series of Pseudo Progressive Diffusion Transformer (PiT). Extensive experiments show superior performance of PiT; for example, PiT-L achieves 54% FID improvement over DiT-XL/2 with less computation.
Inductive kriging estimates values at unobserved locations from sparse sensor data, enabling continuous field reconstruction when dense deployment is impractical. However, common 2 & times; 2 and 2 & times; 3 evaluation protocols can leak spatial information through model selection and obscure true out-of-distribution (OOD) behavior. We propose a leakage-free 3 & times; 3 partition that separates training, validation, and testing in both space and time, so that model fitting, checkpoint selection, and final reporting are performed on distinct spatio-temporal domains. Under this stricter setting, we introduce DRIK (Distribution-Robust Inductive Kriging), a framework with three task-specific mechanisms: Spatial Continuity Regularization (SCR) perturbs coordinates to reduce dependence on one discretized graph; Masked Flow Disambiguation (MFD) prunes ambiguous propagation from zero-padded masked nodes; and Structural Domain Expansion (SDE) uses validation-node topology without labels to reduce train-inference structural mismatch. Experiments on six spatio-temporal datasets show that DRIK consistently outperforms state-of-the-art baselines, reducing MAE by up to 12.48% and achieving lower test-to-validation MAE ratios under leakage-free evaluation. These results indicate that robust inductive kriging requires both leakage-free evaluation and mechanisms that explicitly address the structural shifts introduced by unseen nodes.
Optimal scale combination (OSC) selection is crucial for knowledge extraction in multi-scale decision systems (MDSs). However, existing multi-scale rough set models typically represent the three regions as unions of equivalence classes, losing the structural information of equivalence classes and hindering rule extraction. Moreover, most OSC selection algorithms adopt a fine-to-coarse backward search strategy, which not only contradicts the human coarse-to-fine cognitive mode but also is inefficient on large-scale datasets due to scale combination explosion. To address these issues, a granular sequential three-way decision model for MDSs (GrS3WD-M) is developed, and then an efficient forward optimal scale combination selection algorithm (FOSC) is designed. First, a two-step trisecting strategy is proposed to accelerate boundary region reduction, and then GrS3WD-M based on granular rough sets is developed to preserve the structural information of equivalence classes. As a result, rules can be directly extracted from equivalence classes in the positive region, while redundant computation during granulation is avoided. Second, the core scale combination is defined, an equivalent computational formula is derived, and an efficient algorithm based on GrS3WD-M and attribute grouping is proposed for its computation. Finally, the outer significance of attributes in MDSs is defined, and an efficient method for computing it is presented. Furthermore, the FOSC algorithm is proposed, which starts from the core scale combination and uses the outer significance as a heuristic to guide a coarse-to-fine forward search. Experimental results show that, compared with the state-of-the-art OSC selection algorithms, FOSC improves the average computational efficiency by more than 50 times.
Although many handwritten text recognition systems have been successfully developed, open challenges remain, particularly in accurately grading descriptive answers. This is because the handwriting of descriptive answers may contain unreadable words called malicious text, including various types of struck-out words, overwriting, scribbling, deformation, and degradation. Such words are often used when students are uncertain about their answers and to add more detail to meet the requirements of the question. Recognizing such words accurately is important and an open challenge for the present OCRs. Restoring actual information in such malicious words can help improve the automatic grading of descriptive answers. This paper presents a new technique to improve recognition performance through the restoration of malicious words. We take advantage of the high-quality image generation capability of diffusion models and propose a Multi-Layer Diffusion Network (MLDN), which comprises three modules considering the characteristics of malicious text. Module 1 combines U-Net and GAN to generate noisy images and to perform denoising. Module 2 employs a diffusion-Transformer to restore the structure of characters using feature maps derived from the input and generated samples. Module 3 introduces a pretrained Tr-OCR that utilizes dual decoders for recognizing the text. The performance of the MLDN is validated through experiments on our datasets and benchmark datasets. A comparative analysis with state-of-the-art models shows that the proposed method outperforms state-of-the-art models.
Over the past few decades, feature extraction and pattern recognition in nonlinear time series and complex systems have remained a prominent research focus. Among various analytical approaches, entropy-based measures have gained widespread adoption for their ability to quantify the complexity and irregularity inherent in temporal data. This paper proposes a novel entropy measure termed Angle-Weighted Visibility-based Phase Attention Entropy (AWVP-AttEn). The method enhances traditional attention entropy by integrating a visibility-graph framework, incorporating phase information from the signal, and introducing an angle-based weighting scheme for the visibility graph. These innovations enable more effective identification and extraction of topological and rhythmic information from time series. Experiments are conducted to evaluate AWVP-AttEn on both simulated signals and real-world datasets for identification and classification tasks. The performance is compared against conventional entropy methods as well as several ablated variants of the proposed approach. Results demonstrate that AWVP-AttEn exhibits superior pattern recognition and classification capabilities compared to existing techniques, which provides a new tool and a valuable reference for advancing research in complex system analysis and nonlinear time series pattern recognition.
