Multimodal data is typically collected through heterogeneous sensors and processing pipelines. However, due to variations in acquisition environments, device capabilities, and feature extraction methods, such data often suffers from incompleteness and inconsistent quality across modalities. To address these challenges, prior studies have explored modality selection and data completion strategies to improve information fusion. Nevertheless, these approaches face two main limitations: (1) they struggle to simultaneously ensure computational efficiency for large-scale graph data and maintain structural and semantic consistency across heterogeneous modality graphs; and (2) most of them operate at the modality level and fail to capture fine-grained, sample-specific quality variations. To overcome these issues, we propose a novel clustering framework, Sample Weighted Incomplete Multimodal Clustering Based on Graph Coarsening Label Extraction (IMC-GCSW). The proposed method introduces a graph coarsening-based label extraction strategy. It significantly reduces the computational cost of multimodal graph processing, while preserving key node information and local topological structures. Furthermore, a quality-aware sample weighting strategy is designed to enable fine-grained modeling of modality-specific data quality, allowing the model to dynamically suppress the influence of low-quality modalities on individual samples. Experiments on both general-purpose datasets and the Fructus Aurantii Disease and Pest Datasets demonstrate that the proposed method exhibits superior performance and strong adaptability in handling multimodal data with incompleteness and quality inconsistency.

Composite federated learning provides a comprehensive framework for addressing machine learning tasks that incorporate regularization terms. Nevertheless, numerous established methods within this framework face challenges stemming from data heterogeneity and the high variance induced by stochastic gradients mapping. These issues cause significant deviations in clients' local models, resulting in unstable and slow convergence of the global model and increased communication costs. To address this, we propose a novel algorithm named FedSAGA, designed to solve composite federated optimization problems that include a nonsmooth global regularization term. In FedSAGA, we construct a combination of the latest and historical local model updates without incurring additional communication overhead. This estimate serves as a control variate on the client side to reduce the variance introduced by stochastic gradient mapping and client sampling. To further accelerate convergence, we leverage the inertial extrapolation step locally at the clients. Theoretical analysis demonstrates that FedSAGA attains a sublinear convergence rate under general convex settings and a linear convergence rate under strongly convex settings. Numerical experiments based on both real and synthetic datasets effectively demonstrate the convergence of the proposed algorithm under partial client participation.

Momentum methods based on the inertial technique have been widely adopted in federated learning (FL). However, a unified framework for FL acceleration methods needs further exploration. The further, gradient descent is the main training algorithms in FL, which, although powerful, is not universally feasible or best choice. Motivated by this, we provide a unified framework to bridge the gap between practice and theory by the distributed version of general inertial Krasnosel’ski ${ }^{\text {c }}$ i-Mann (DG-IKM) iteration and consider operator splitting methods applicable to nonsmooth optimization. By formulating the federated optimization problem as a monotone inclusion problem and instantiating the general framework, we propose a novel accelerated algorithm, FedIDR, which achieves fast convergence by using inertial technique in the parallel Douglas-Rachford splitting method. And then, the inertial Krasnosel’ski ${ }^{\text {-}}$ i-Mann (IKM) theory allows us to reuse the general convergence results by formulating FedIDR as an application of the nonexpansive operator. Finally, we derive the convergence rate and numerical experiments based on real and synthetic datasets are considered to evaluate the proposed algorithm.

There is growing evidence that product introductory information plays a crucial role in online consumer purchase behavior. However, existing researches on the effect of introductory information on individuals’ purchase intention in live-streaming were limited. In this setting, based on the Stimulus-Organism-Response (S-O-R) model, we explored the relationship among introductory information, trust, purchase intention and display uniqueness. It was done by collecting questionnaire survey data from 500 consumers who have participated in live-streaming e-commerce. The results showed that the introductory information in live-streaming e-commerce has a significant positive effect on consumers’ purchase intention through trust. Additionally, the results revealed a significant moderating effect of display uniqueness on the relationship between introductory information and trust. Display uniqueness could significantly and positively moderate the predictive effect of introductory information on trust. Thus, this study will extend the applicability of S-O-R model to further understand the influence mechanism of introductory information on purchase intention and enrich the research in the live-streaming e-commerce. Also, it can enhance the understanding of consumers’ purchase behaviors to optimize the live-streaming system and online marketing tools, thereby helping companies achieve their business goals.

Recently there has been a significant surge in multimodal learning in terms of both image-to-text and text-to-image generation. However, the success is typically limited to English, leaving other languages largely behind. Building a competitive counterpart in other languages is highly challenging due to the low-resource nature of non-English multimodal data (i.e., lack of large-scale, high-quality image-text data). In this work, we propose MPM, an effective training paradigm for training large multimodal models in low-resource languages. MPM demonstrates that Multilingual language models can Pivot zero-shot Multimodal learning across languages. Specifically, based on a strong multilingual large language model, multimodal models pretrained on English-only image-text data can well generalize to other languages in a (quasi)-zero-shot manner, even surpassing models trained on image-text data in native languages. Taking Chinese as a practice of MPM, we build large multimodal models VisCPM in image-to-text and text-to-image generation, which achieve state-of-the-art (open-source) performance in Chinese. To facilitate future research, we open-source codes and model weights at https://github.com/OpenBMB/VisCPM.

Seismic deconvolution aims to improve the vertical seismic resolution by compressing the seismic wavelet and extending the seismic frequency bandwidth. Deconvolution in the frequency domain is normally implemented in three steps: wavelet estimation, deconvolution operator construction, and data filtering. We propose a seismic resolution enhancement approach by spectral shaping using shaping-regularized inversion. Instead of designing a deconvolution operator based on the extracted wavelet, we formulate an inverse problem using the seismic spectrum as the diagonal kernel matrix and a predefined filter as the expected spectrum. Assuming the randomness in the reflectivity and smoothness of the spectral shaping operator, we estimate the spectral shaping operator by inversion via a shaping regularization scheme, which imposes constraints by shaping the estimated spectral shaping operator model to the admissible model space. The role of the shaping regularization in inversion is to ensure the continuity and smoothness of the spectral shaping operator. We use a synthetic model and real land seismic data to demonstrate the effectiveness of the proposed approach in seismic resolution enhancement.

To predict the availability state of a node in a distribution network, its history trace is usually used. Sometimes, some usage behavior patterns cannot be captured precisely from the insufficient trace, which may lead to unreliable predictors. In this paper, to alleviate the data sparseness problem, the nodes with the similar behaviors are clustered, and all history information in a same cluster is seen as another information source for any node in it. For each node, an N-gram model is used to train the predictor by the combination of the new source and the node's own trace. In addition, because it is hard to capture the trace of all nodes in large scale networks, such as P2P networks, a bagging based prediction algorithm is proposed, which can be applied in the distribution environment and relieve the effect of the noisy data. In our experiments, three datasets are evaluated. Results show that the prediction performance of our cluster based N-gram predictor is better than the results of several other predictors. And the bagging based prediction algorithm presents its validity in the distribution environment.