Current methods for document key information extraction (DKIE) rely heavily on labeled data with high annotation costs. To mitigate this issue, the semi-supervised learning (SSL) paradigm, which utilizes unlabeled document samples, has gained broad attention in DKIE. However, existing SSL methods require labeled and unlabeled data to share an identical label space, which is impractical in many DKIE tasks (i.e., some unlabeled samples do not belong to any known classes in the labeled set). In this paper, we formulate this problem as Class-Missing Semi-supervised (CMSS) DKIE. In DKIE, unknown classes usually belong to minority and fine-grained categories, intensifying the misconnections between known and unknown classes and making CMSS more challenging. To address this issue, we propose Synergistic Refinement Estimation (SRE), a progressive prototype estimation scheme that alleviates the unknown classes bias to the majority known classes on long-tailed unlabeled data. Furthermore, dynamic threshold hash rectification and structural calibration mechanisms are proposed to correct connections between fine-grained classes. Extensive experimental results demonstrate that SRE surpasses existing state-of-the-art methods on several DKIE benchmarks. Code is available at https://github.com/anonymoulink/SRE_DKIE .

Learning directed acyclic graphs (DAGs) from observational data that involve a set of variables carrying intrinsic noise is a crucial yet challenging task. Recent approaches frame the DAG learning task as minimizing a reconstruction-based objective function, i.e., reconstructing observed data by learning a DAG, while adhering to an acyclic constraint. However, optimizing this objective does not always guarantee the correctness of the learned graphs. One reason for this is that the intrinsic noise entangled with the variables is inadvertently absorbed in the reconstruction process at the expense of inferring incorrect DAG structures. To address this issue, we propose a novel DAG learner that first injects artificial noise into observational variables that are contaminated by fixed intrinsic noise. The next step involves reconstructing these perturbed variables using a weighted structure estimator and a weighted noise estimator, instead of reconstructing the observational variables solely with the fixed intrinsic noise. This strategy effectively reduces the sensitivity of the structure estimator to the fixed intrinsic noise. Additionally, we observe a strong similarity between the proposed DAG learner and diffusion models. This similarity motivates us to replicate the well-known denoising capabilities of diffusion models in our DAG learner. We reformulate and adapt the denoising process in denoising diffusion probabilistic models (DDPMs), which allows us to derive a specific weight schedule for the weighted structure and noise estimators of our DAG learner. Extensive experiments conducted on synthetic and real datasets with varying scales demonstrate the outstanding performance of our proposed method.

Exploiting multi-modal magnetic resonance imaging complementary information for brain tumor segmentation is still a challenging task. Existing methods are usually inclined to learn the joint representation of all tumor regions indiscriminately, thus salient sub-region or healthy tissue would be dominant during the training procedure, which leads to a biased and limited representation performance. In this study, a novel transformer-based multi-modal brain tumor segmentation approach is developed by decoupling and coupling strategy. First, Anatomy-induced Region Decoupler decouples the representation of the tumor scattered in different semantic sub-regions following anatomical view, which forces the model to fully learn intra-region representation separately with multiple modalities context. Additionally, we introduce the collaborative decoupling of the corresponding sub-region edge to serve auxiliary cues. We then design the Edge-supported Intra-region Coupler to separately couple edge and object learning within each anatomical sub-region structure. Lastly, the Mutual Cross-region Coupler is further applied to implement mutual improvement by coupling complementary gains among the above decoupled sub-regions. Extensive experiments clearly demonstrate that our method outperforms current state-of-the-arts for brain tumor segmentation on BRATS2018, BRATS2020, MSD, and BRATS2021 benchmarks while retaining high efficiency in the learning procedure. The code is available at https://github.com/mathwrx/Decouple-and-Couple_Learning_in_Multi-Modal_Brain_Tumor_Segmentation.

Recently, the expansion of Variance Reduction (VR) to Riemannian stochastic non-convex optimization has attracted increasing interest. Inspired by recursive momentum, we first introduce Stochastic Recursive Variance Reduced Gradient (SRVRG) algorithm and further present Stochastic Recursive Gradient Estimator (SRGE) in Euclidean spaces, which unifies the prevailing variance reduction estimators. We then extend SRGE to Riemannian spaces, resulting in a unified Stochastic rEcursive vaRiance reducEd gradieNt frAmework (SERENA) for Riemannian non-convex optimization. This framework includes the proposed R-SRVRG, R-SVRRM, and R-Hybrid-SGD methods, as well as other existing Riemannian VR methods. Furthermore, we establish a unified theoretical analysis for Riemannian non-convex optimization under retraction and vector transport. The IFO complexity of our proposed R-SRVRG and R-SVRRM to converge to $\varepsilon$-accurate solution is $\mathcal{O}\left(\min \{n^{1/2}{\varepsilon^{-2}}, \varepsilon^{-3}\}\right)$ in the finite-sum setting and ${\mathcal{O}\left( \varepsilon^{-3}\right)}$ for the online case, both of which align with the lower IFO complexity bound. Experimental results indicate that the proposed algorithms surpass other existing Riemannian optimization methods.

In this paper, we propose AdaMSS, an adaptive multi-subspace approach for parameter-efficient fine-tuning of large models. Unlike traditional parameter-efficient fine-tuning methods that operate within a large single subspace of the network weights, AdaMSS leverages subspace segmentation to obtain multiple smaller subspaces and adaptively reduces the number of trainable parameters during training, ultimately updating only those associated with a small subset of subspaces most relevant to the target downstream task. By using the lowest-rank representation, AdaMSS achieves more compact expressiveness and finer tuning of the model parameters. Theoretical analyses demonstrate that AdaMSS has better generalization guarantee than LoRA, PiSSA, and other single-subspace low-rank-based methods. Extensive experiments across image classification, natural language understanding, and natural language generation tasks show that AdaMSS achieves comparable performance to full fine-tuning and outperforms other parameter-efficient fine-tuning methods in most cases, all while requiring fewer trainable parameters. Notably, on the ViT-Large model, AdaMSS achieves 4.7\% higher average accuracy than LoRA across seven tasks, using just 15.4\% of the trainable parameters. On RoBERTa-Large, AdaMSS outperforms PiSSA by 7\% in average accuracy across six tasks while reducing the number of trainable parameters by approximately 94.4\%. These results demonstrate the effectiveness of AdaMSS in parameter-efficient fine-tuning. The code for AdaMSS is available at https://github.com/jzheng20/AdaMSS.

Identifying unknown protein functional modules, such as protein complexes and biological pathways, from protein-protein interaction (PPI) networks, provides biologists with an opportunity to efficiently understand cellular function and organization. Finding complex nonlinear relationships in underlying functional modules may involve a long-chain of PPI and pose great challenges in a PPI network with an unevenly sparse and dense node distribution. To overcome these challenges, we propose AdaPPI, an adaptive convolution graph network in PPI networks to predict protein functional modules. We first suggest an attributed graph node presentation algorithm. It can effectively integrate protein gene ontology attributes and network topology, and adaptively aggregates low- or high-order graph structural information according to the node distribution by considering graph node smoothness. Based on the obtained node representations, core cliques and expansion algorithms are applied to find functional modules in PPI networks. Comprehensive performance evaluations and case studies indicate that the framework significantly outperforms state-of-the-art methods. We also presented potential functional modules based on their confidence.

Molecular optimization, which transforms a given input molecule X into another Y with desired properties, is essential in molecular drug discovery. The traditional approaches either suffer from sample-inefficient learning or ignore information that can be captured with the supervised learning of optimized molecule pairs. In this study, we present a novel molecular optimization paradigm, Graph Polish. In this paradigm, with the guidance of the source and target molecule pairs of the desired properties, a heuristic optimization solution can be derived: given an input molecule, we first predict which atom can be viewed as the optimization center, and then the nearby regions are optimized around this center. We then propose an effective and efficient learning framework, Teacher and Student polish, to capture the dependencies in the optimization steps. A teacher component automatically identifies and annotates the optimization centers and the preservation, removal, and addition of some parts of the molecules; a student component learns these knowledges and applies them to a new molecule. The proposed paradigm can offer an intuitive interpretation for the molecular optimization result. Experiments with multiple optimization tasks are conducted on several benchmark datasets. The proposed approach achieves a significant advantage over the six state-of-the-art baseline methods. Also, extensive studies are conducted to validate the effectiveness, explainability, and time savings of the novel optimization paradigm.

Clustering techniques attempt to group objects with similar properties into a cluster. Clustering the nodes of an attributed graph, in which each node is associated with a set of feature attributes, has attracted significant attention. Graph convolutional networks (GCNs) represent an effective approach for integrating the two complementary factors of node attributes and structural information for attributed graph clustering. Smoothness is an indicator for assessing the degree of similarity of feature representations among nearby nodes in a graph. Oversmoothing in GCNs, caused by unnecessarily high orders of graph convolution, produces indistinguishable representations of nodes, such that the nodes in a graph tend to be grouped into fewer clusters, and pose a challenge due to the resulting performance drop. In this study, we propose a smoothness sensor for attributed graph clustering based on adaptive smoothness-transition graph convolutions, which senses the smoothness of a graph and adaptively terminates the current convolution once the smoothness is saturated to prevent oversmoothing. Furthermore, as an alternative to graph-level smoothness, a novel fine-grained nodewise-level assessment of smoothness is proposed, in which smoothness is computed in accordance with the neighborhood conditions of a given node at a certain order of graph convolution. In addition, a self-supervision criterion is designed considering both the tightness within clusters and the separation between clusters to guide the entire neural network training process. The experiments show that the proposed methods significantly outperform 13 other state-of-the-art baselines in terms of different metrics across five benchmark datasets. In addition, an extensive study reveals the reasons for their effectiveness and efficiency.

The coronavirus disease 2019 (COVID-19) pandemic is spreading worldwide. Considering the limited clinicians and resources and the evidence that computed tomography (CT) analysis can achieve comparable sensitivity, specificity, and accuracy with reverse-transcription polymerase chain reaction, the automatic segmentation of lung infection from CT scans supplies a rapid and effective strategy for COVID-19 diagnosis, treatment, and follow-up. It is challenging because the infection appearance has high intraclass variation and interclass indistinction in CT slices. Therefore, a new context-aware neural network is proposed for lung infection segmentation. Specifically, the autofocus and panorama modules are designed for extracting fine details and semantic knowledge and capturing the long-range dependencies of the context from both peer level and cross level. Also, a novel structure consistency rectification is proposed for calibration by depicting the structural relationship between foreground and background. Experimental results on multiclass and single-class COVID-19 CT images demonstrate the effectiveness of our work. In particular, our method obtains the mean intersection over union (mIoU) score of 64.8%, 65.2%, and 73.8% on three benchmark datasets for COVID-19 infection segmentation.

The ongoing coronavirus disease 2019 (COVID-19) pandemic spread throughout China and worldwide since it was reported in Wuhan city, China in December 2019. 4 589 526 confirmed cases have been caused by the pandemic of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), by May 18, 2020. At the early stage of the pandemic, the large-scale mobility of humans accelerated the spread of the pandemic. Rapidly and accurately tracking the population inflow from Wuhan and other cities in Hubei province is especially critical to assess the potential for sustained pandemic transmission in new areas. In this study, we first analyze the impact of related multisource urban data (such as local temperature, relative humidity, air quality, and inflow rate from Hubei province) on daily new confirmed cases at the early stage of the local pandemic transmission. The results show that the early trend of COVID-19 can be explained well by human mobility from Hubei province around the Chinese Lunar New Year. Different from the commonly-used pandemic models based on transmission dynamics, we propose a simple but effective short-term prediction model for COVID-19 cases, considering the human mobility from Hubei province to the target cities. The performance of our proposed model is validated by several major cities in Guangdong province. For cities like Shenzhen and Guangzhou with frequent population flow per day, the values of [Formula: see text] of daily prediction achieve 0.988 and 0.985. The proposed model has provided a reference for decision support of pandemic prevention and control in Shenzhen.