Integrating complementary observational information into a unified representation remains a core challenge for robust knowledge discovery systems. In practical scenarios, observational data often exhibit multi-scale heterogeneity, data source redundancy, and dynamic variability, posing three key challenges for existing fusion methods: first, selecting a single attribute scale for all samples fails to accommodate real-world attribute differences across samples; second, redundant and highly correlated data sources tend to cause information duplication and a surge in computational overhead; third, the lack of effective mechanisms for handling dynamic changes in data sources makes it difficult to ensure the accuracy of fusion results. To tackle these problems, this paper proposes a dynamically adaptive multi-source clustering and entropy-guided consistent-scale fusion model based on neighborhood rough set theory. The model first uses three-way clustering to adaptively group data sources and assign differentiated weights, addressing the issues of information duplication and inefficient computation caused by multi-source redundancy. It then employs conditional entropy to select optimal scales within each data source and uses a consistency strength measure to automatically choose appropriate scales for each object, thereby resolving the problem of granularity mismatch in sample-level multi-scale selection. A lightweight dynamic source management strategy is designed that efficiently adapts to the addition or removal of data sources through online updates of cluster centroids and re-clustering mechanisms, ensuring the stability of fusion results. Experimental comparisons with various traditional fusion techniques demonstrate significant improvements across all evaluation metrics, confirming the effectiveness of the proposed model.
Federated Learning (FL) has emerged as a pivotal paradigm for privacy-preserving distributed machine learning. Decentralized FL (DFL), by removing reliance on a single global coordinator, offers enhanced robustness, scalability, and autonomy. However, this architecture introduces challenges in communication, optimization, and security. This survey provides a structured synthesis of DFL, spanning its theoretical foundations, communication and optimization mechanisms, privacy and trust protection, application paradigms, and implementation platforms. We first establish the theoretical framework of DFL through algebraic graph theory and decentralized learning principles. Next, we investigate optimization strategies for data heterogeneity and communication bottlenecks through methodological categorization. Additionally, we analyze privacy threats in decentralized systems and present solutions ranging from differential privacy to cryptographic techniques. We investigate each dimension by reviewing established techniques and discussing their pros and cons, as well as the unresolved challenges. Moreover, we introduce the application scenarios and software frameworks of DFL. Finally, we outline future research directions. In a nutshell, this paper provides a comprehensive overview of the fundamentals of DFL, its applications in edge computing and multi-agent systems, and the existing challenges and open issues with insights for future in-depth research.
Evaluating the similarity between generated and reference text in natural language generation (NLG) remains a significant challenge. Existing automatic metrics often fail to capture nuanced semantic relationships and do not consistently align with human judgment. To address this limitation, we propose an iterative multi-agent framework for text similarity evaluation. The framework comprises three stages: a Task Performer (Stage 1) for initial assessment, a Multi-Agent Analysis stage (Stage 2) in which specialized agents iteratively reassess the output, and a Convergence Checker (Stage 3) that determines whether consensus or majority agreement has been reached. When convergence fails, an Arbiter produces the final decision. We implement four architectural variants: Architecture One incorporates all three stages with an Arbiter; Architecture Two uses only Stages 2 and 3; Architecture Three omits the Arbiter, returning the Task Performer's judgment upon non-convergence; and Architecture Four extends Stage 2 with an Evidence Card, requiring each agent to provide a label, supporting points, conflicting points, and missing information. We evaluate the framework on six tasks spanning open-QA, summarization, and medical, legal, financial, and biomedical QA domains. Results demonstrate that the proposed framework generally improves on the standalone LLM baseline and more closely aligns with human evaluation across the evaluated tasks and configurations.
Vision transformers (ViTs) have achieved significant success in various computer vision tasks, enabling their deployment on privacy-protected devices. These devices generally have privacy protection requirements, which motivates the development of model inversion techniques such as Sparse Model Inversion (SMI) specifically designed for ViTs. SMI is a promising approach for model inversion, enabling the reconstruction of original data in data-free compression scenarios. Although SMI improves inversion efficiency via background patch pruning, it provides limited treatment of the accuracy dimension. Therefore, for privacy-protected devices, it is crucial to develop accurate and efficient model inversion strategies. In this paper, we present SMI V2 as an inversion-specific extension of SMI for efficient and accurate data-free compression of ViTs. SMI V2 retains SMI's iterative sparse-inversion framework and adapts established bipartite matching and weighted aggregation to reduce redundancy among the tokens that remain active. Matching is recomputed as the synthetic image evolves, and fusion is placed before self-attention so that each enabled block processes a shorter sequence. For the label-conditioned setting, a compensation objective refines the target-versus-non-target logit margin. Experiments on data-free quantization, knowledge transfer, and sparse pruning evaluate the resulting efficiency-accuracy trade-off.
The proliferation of high-velocity big data streams from contemporary technologies, such as the internet of things, social media platforms, wireless sensor networks, blockchain systems, etc., has intensified the need for scalable methods to model and analyze evolving big data representations. Advanced evolutionary representation learning techniques, employing deep learning, graph neural networks, and transformer architectures, generate time-varying feature vectors for diverse entities, including users, sensor nodes, and cryptocurrency wallets. Unsupervised learning over these evolving representations enables the discovery of latent structure, emerging patterns, and anomalies, which are critical for characterizing concept drift. Existing evolutionary clustering approaches, however, often assume fixed data points and a constant number of clusters across timestamps, or lack the scalability required for large-scale, real-world applications. Furthermore, most methods are unable to account for essential dynamics, such as cluster splitting and merging, which limits their ability to capture complex drift behaviors. To overcome these limitations, this paper presents evolVAT, a fast and scalable evolutionary clustering algorithm that incrementally updates clustering results by leveraging previously inferred structure. The proposed method accommodates multiple data-point transitions between consecutive snapshots and supports the addition and removal of entities over time. Experimental evaluations on a broad set of synthetic and real-world datasets demonstrate the effectiveness, robustness, and adaptability of evolVAT in diverse application domains.
