Temporal knowledge graph reasoning (TKGR) aims to predict missing facts or infer future events based on historical information. Recently, language model (LM)-based approaches have shown promising results by incorporating textual representations of historical facts. However, existing LM-based methods still face two key limitations: they do not adequately capture structural dependencies in TKGs, and most approaches formulate reasoning as a discriminative task rather than leveraging the generative capability of modern large language models (LLMs). To address these limitations, we propose SAG, a Structure-Aware Generative framework for temporal knowledge graph reasoning with historical evidence. Specifically, SAG first constructs dual-view historical evidence to better capture structural dependencies in temporal knowledge graphs. The structural representations are then projected into the embedding space of a large language model as soft evidence tokens through a structure-text adapter, enabling the language model to jointly reason over structural evidence and textual query context. Finally, SAG formulates TKGR as an end-to-end generative task through instruction tuning, allowing the model to directly generate the missing entity. Extensive experiments on three widely used benchmarks demonstrate that SAG consistently outperforms strong baselines. In particular, compared with the strongest baseline, SAG achieves Hits@1 improvements of 6.45
In modern enterprise architectures, the persistence of business data often suffers from “Semantic Reduction”—the stripping of multidimensional business contexts into flat storage records, leading to isolated data silos. While Large Language Models (LLMs) offer promising capabilities for automated Knowledge Graph (KG) construction, existing methods struggle with hallucination and lack of structural integrity when processing complex enterprise events without architectural constraints. To address this, we propose the Knowledge Proliferation Engine (KPE), a domain-driven architectural framework that transforms passive data persistence into active, event-driven knowledge creation. We interpret DDD Bounded Contexts as an architectural prior that partitions the semantic space into domain-specific scopes, which may help reduce cross-domain schema interference during LLM-based knowledge generation. The framework features a novel Federated Identity-Role Model—comprising a pure Shared Kernel and Domain Proxies—to support precise entity alignment across heterogeneous domains. We evaluated KPE using a curated dataset of complex business events (RB-50) and a real-world financial log (BPI-2017). In the evaluated RB-50 setting, KPE achieved a Factual Fidelity of 0.934 and Schema Conformance of 86.7 https://anonymous.4open.science/r/KPE-Framework-40B6 .
Modern web-based distributed systems are becoming increasingly complex and scaled up, making real-time anomaly detection more challenging and essential for ensuring trust and quality of services that support modern society. Traditional monitoring approaches lack coordination between components, semantic interpretation and adaptability, leading to high false positive rates and insufficient explanations. These are critical issues in complex microservice environments, which require rapid incident response. This paper introduces a Distributed Agent-based Intelligent Monitoring System (DAIMS) to address these challenges. The framework combines machine learning (ML) with semantic reasoning through a novel four-layer architecture. Intelligent agents monitor service metrics locally and detect anomalies using XGBoost models. These agents collaborate via a message bus and semantically enrich detected events with a lightweight OWL ontology. Simultaneously, SWRL rules are applied by a central reasoner to identify system-wide problems, root causes and propagation chains. Experimental results demonstrate that DAIMS outperforms baseline methods on the TrainTicket benchmark. It achieves an F1-score of 0.91, a detection rate of 94
Cross-domain text classification (CDTC) is applied to improve the performance of models trained on source domains when applied to a target domain, but often rely heavily on domain specificity. Domain generalization (DG) is considered as a good solution. In this work, we explore and construct a unified text DG framework UAM-CFEO to improve the performance and generalization of CDTC. Specifically, UAM-CFEO incorporates a module and a two-stage feature optimization and enhancement strategy. Specifically, CAE (Causal AutoEncoder) module is designed to extract cross-domain stable causal features. The first-stage feature optimization strategy is introduced to combine causal invariant learning with supervised contrastive learning for enhancing both the consistency and discriminability of causal representations across domains. Then, we augment the extracted causal features with a key-value memory augmentation that trained with an uncertainty-aware meta-learning framework in the second enhancement stage, so as to improve the robustness under low-resource settings and large distribution shifts. Meanwhile, Extensive experiments demonstrate that UAM-CFEO consistently outperforms existing methods in both multi-domain leave-one-domain-out settings and cross-dataset evaluations, validating its effectiveness and superiority.
Incomplete Multi-view Clustering (IMC) has attracted increasing attention in recent years. However, most existing IMC approaches may suffer from the following problems: (a) they may neglect to learn partial-consensus knowledge, and this knowledge widely exists in incomplete multi-view data; and (b) they may fail to fully explore the knowledge in views. To address the above issues, we propose a Tensor-constrained Consensus, Partial-consensus and Specificity Components learning framework for Incomplete Multi-view Clustering (TCPS-IMC). TCPS-IMC includes two components: (a) Views Division Learning (VDL): The views are divided into consensus, partial-consensus and specific parts to separately learn incomplete multi-view data; and (b) Knowledge Enhancement Learning (KEL): the view low-rank is guaranteed by imposing the tensor nuclear norm on views, the important features of partial-consensus knowledge are extracted by minimizing the Hadamard product between partial-consensus parts, and the more specific knowledge is learned by imposing the F-norm on specific parts. By TCPS-IMC, incomplete multi-view data are sufficiently learned, and the knowledge in views are explored as much as possible. In the incomplete multi-view clustering, TCPS-IMC achieves higher average ACC than the suboptimal method by 4.64% under the 50 https://github.com/GDUT-zhangjinchao/TCPS-IMC .
