In exploratory search scenarios, users frequently turn to Knowledge Graphs (KGs) as rich sources of structured knowledge. However, user queries are often imprecise, incomplete, or ambiguous, and KGs themselves are inherently incomplete. Moreover, when KGs are constructed from crowd-sourced content or by integrating heterogeneous data sources of varying quality, the need for mechanisms to assess information reliability becomes critical. This work introduces Context-Augmented Knowledge Graphs (CoaKG), an enriched KG model designed to support context-aware exploration for decision-making and task learning. CoaKG incorporates both explicit contextual mappings and inference rules to derive implicit context, enabling more nuanced query answering. We also present the CoaKG Query Engine, a query processing framework to retrieve contextually enriched answers tailored to users' goals and context.
Investigating narratives in specific text-as-data applications has been extensively implemented using hybrid approaches, combining several NLP solutions and statistical techniques. While this direction provides a substantial understanding of key semantic patterns, it remains a challenging task not applicable to all languages. Narratives, especially in social sciences, are often regarded as the study of causality between specific terms, to uncover embedded directed or indirected links for explanation and further hypotheses testing. This paper establishes CausaMap, as a semi-supervised method to unfold causality between predefined terms of interests. It is built on dependency parsers and the use of tree-based structural equation models to quantify direct and indirect causal links. Thus, testing terms' causality in a directed network of parsed, lemmatized words is equivalent to uncovering informative inner structures with preserved semantic properties, without external information. Applied to two different corpora in English and German languages, CausaMap yielded mixtures of directed acyclic graphs and bidirectional edges, forming a narrative picture centered around monetary objectives and macroeconomic goals reported in the Federal Reserve governors' speeches (1996-2024). The German political parties' manifestos (1990-2021) demonstrated a dominance of policy-oriented terms when studying economic and employment debates, where the relatively high number of uncovered connecting nodes explains the causality buildup of pre-defined terms as well as languagespecific morphologies.
Temporal knowledge graphs (TKGs) have emerged as a critical component in modern artificial intelligence systems, enabling machines to reason over dynamic information. However, existing methods for TKG reasoning use massive amounts of information and computational resources, failing to capture minimal essential temporal knowledge with their dynamics while eliminating irrelevant or noisy information for a more precise reasoning process. To this end, we introduce the Forest Fire Contrastive Approach (FFCA), a contrastive learning architecture based on forest fire sampling that presents a preferential attachment mechanism for the extrapolation of TKG. This allows high-degree nodes to attract new connections, improving the pipeline's predictive capability while keeping it compact, leading to an efficient learning and inference process. This approach introduces a degree-biased burn probability that gathers a minimal but highly correlated subgraph relevant to the query as the global view. Simultaneously, a sufficient number of the most recent snapshots were gathered as the local view, preserving task-relevant graph information and removing noise while reducing the computational complexity and memory requirements, improving the sustainability of TKG reasoning models. Experiments on three publicly available benchmark datasets widely used for TKG extrapolation tasks demonstrate that the proposed approach achieves competitive predictive performance to current state-of-the-art methods while demonstrating substantial improvements in memory efficiency, a critical consideration for scalability to large-scale temporal knowledge graph datasets.
The identification of anomalies and outliers through machine-learning models constitutes a critical and extensively studied domain. However, lack of training data constitutes a significant concern in many application domains. Considering the realm of network intrusion detection - detecting previously unseen attacks, and/or attacks on devices for which training data has not been collected - is essential for implementing effective security measures. To address this issue, we propose an innovative transfer-learning approach. Our method generates new labels for source datasets by leveraging insights from the target dataset, thereby facilitating the accurate identification of normal data within the source datasets. Subsequently, we train an unsupervised model using these normal data to infer unseen anomalies. Furthermore, we employed state-of-the-art explainable artificial intelligence techniques to extract significant features from source datasets, thereby enhancing our model's performance. Our comprehensive evaluations clearly demonstrate the efficacy of our method across both generic (non-IoT) and IoT datasets, consistently surpassing existing transfer learning approaches. For example, our method achieved a remarkable accuracy of 0.99 in detecting anomalous traffic from UDP flooding attacks, whereas the well-known transfer learning model, SSkNNO, achieved an accuracy of only 0.83.
The Visual Question Answering (VQA) task requires not only accurate answers but also interpretable reasoning processes, particularly in real-world applications where transparency is critical. To reduce annotation and computational costs while maintaining interpretability, the Few-shot Multimodal Explainable VQA (FS-MEVQA) task has been introduced, which aims to generate explanations with limited supervision. In this work, we propose OPeMer (One-shot Prompting and Execution-driven Multimodal Explainable Reasoning), a code-based framework that leverages large language models (LLMs) to generate executable Python programs for multimodal reasoning in a oneshot setting. These programs interact with a lightweight Python API to process visual inputs, capture intermediate reasoning artifacts—such as object crops and spatial relations—and optionally call external tools for open-world visual understanding. The resulting execution traces are serialized and provided to the LLM via a secondary prompt, enabling the generation of coherent multimodal explanations grounded in both visual and textual evidence. Designed without reliance on handcrafted rules or large-scale supervision, OPeMer offers an efficient and extensible approach to explainable multimodal reasoning. Experimental results on the SME dataset demonstrate that OPeMer achieves strong answer accuracy and explanation quality, even when using cost-effective LLMs under limited supervision, suggesting its potential for scalable and interpretable VQA.
Knowledge graphs are structured and interlinked formations that can enhance data expressiveness, as crucial data can be illustrated in a meaningful way that highlights the most significant and impactful affinities among them. In this work, we propose optimized AI-based construction of domain-specific and generic knowledge graphs, including datasets of specific unexplored domains, such as SDGs, and we work on various datasets to prove the generality of our findings, that outperform existent methods. CNN variations and various LLM models improve the performance and reduce training times for the constructed generic and domain-specific knowledge graphs and, hence, facilitate the work of practitioners, who collect and evaluate the data.
Polarization poses a growing threat to democratic discourse, public trust, and societal stability. To better understand its structure and evolution, we conceptualize polarization as a multi-level phenomenon spanning entities, groups, and topics. We present PRISM, a framework that models polarization using a typed, weighted, and directed structure known as the Polarization Knowledge Graph (PKG). PRISM introduces a suite of analytical methods for multi-level analysis: i) identifying key actors and categorizing them as protagonists or antagonists based on their contribution to conflict, ii) measuring group cohesiveness through ideological alignment and topic-level agreement, and iii) ranking topics by their polarization intensity. We validate PRISM through a case study on U.S. COVID-19 media discourse, uncovering polarization patterns that align with established findings and highlight the politicization of the pandemic.
Due to the inherent complexity of video data, video action recognition faces significant challenges in modeling spatial-temporal dynamics and handling diverse scene contexts. Although scene graph-based methods can effectively model interactions between entities, most existing approaches overlook the rich semantic information embedded within scene graphs. Additionally, integrating large language models (LLMs) for semantic enhancement often suffers from hallucination problems, potentially introducing incorrect reasoning that misleads action recognition. To address these limitations, we propose SSGR-AR, a novel framework that structurally represents videos through scene graphs and constrain LLM reasoning using structured semantic paths derived from scene graph knowledge, ensuring controllable and reliable semantic enrichment. Moreover, we formulate entity alignment as a link prediction task and leverage a graph transformer to model the dynamic evolution of actions, thereby enhancing the model's capacity for long-term temporal reasoning. Experimental results on three widely used benchmark datasets show that our method outperforms state-of-the-art methods in terms of action recognition accuracy and generalization robustness.
Event Sequence Anomaly Detection (ESAD) plays a crucial role in domains such as healthcare, DevOps, and information security, where identifying deviations from normal patterns in event sequences is essential for ensuring stability and mitigating risks. Despite notable progress, existing ESAD methods often struggle when handling continuous-time event streams. Statistical models such as Poisson or Hawkes processes offer efficiency but fail to capture nonlinear temporal dependencies, while deep learning methods demand large-scale labeled data and frequently suffer from poor interpretability. These limitations hinder their deployment in high-stakes applications where reliability and transparency are critical. To address these challenges, we introduce TPP-LLMAD, a novel framework that integrates Temporal Point Processes (TPPs) with Large Language Models (LLMs) for interpretable anomaly detection. In TPP-LLMAD, neural TPPs model event sequences and estimate intensity functions, which are then transformed into structured triplets of timestamps, event marks, and intensities. These representations are embedded into tailored prompts that guide an LLM to assess deviations, assign anomaly labels, and generate humanreadable explanations. By combining the quantitative rigor of TPPs with the interpretive capacity of LLMs, the framework bridges the gap between mathematical modeling and naturallanguage reasoning. Extensive experiments on real-world datasets demonstrate that TPP-LLMAD achieves performance comparable to or exceeding state-of-the-art baselines while providing explanations that enhance interpretability and usability. This work represents the first systematic integration of TPP intensity modeling with LLM-based reasoning for ESAD, advancing the frontier of interpretable event sequence analysis.
Retrieval-Augmented Generation (RAG) addresses the limitations of Large Language Models (LLMs) when providing responses to domain-specific questions. Graph-based RAG approaches, such as GraphRAG, enhance retrieval by capturing semantic relationships within knowledge graphs (KGs). While the FAIR principles (Findability, Accessibility, Interoperability, and Reusability) are becoming prevalent for scientific data management, especially in complex domains such as medicine, existing RAG approaches lack a structured FAIRification of the underlying knowledge resources. This lack limits their potential for FAIR information retrieval in these domains. To address this gap, we introduce FAIR GraphRAG, a novel framework that integrates FAIR Digital Objects (FDOs) as the fundamental units of a graph-based retrieval system. Each graph node represents an FDO that incorporates core data, metadata, persistent identifiers, and semantic links. We leverage LLMs to support schema construction and automated extraction of content and metadata from data sources. The framework was co-designed by physicians and computer scientists to ensure technical and clinical relevance. We apply FAIR GraphRAG to a biomedical dataset in gastroenterology, demonstrating its applicability to single-cell data. Beyond ensuring adherence to the FAIR principles, FAIR GraphRAG significantly improves question answering accuracy, coverage, and explainability, particularly for complex queries involving metadata and ontology links. This work shows the feasibility of combining FAIR data practices with graph-based retrieval techniques. We see potential for applying our approach to other specialized fields such as education and business.
In cybersecurity analysis, existing intrusion detection techniques can identify attack behaviors but fail to provide deep contextual information such as exploited vulnerabilities, attack techniques, and defense strategies. Traditional entity recognition and relation extraction methods are difficult to apply directly due to the high sparsity and specialization of cybersecurity domain entity relationships, limiting deep analysis capabilities. To address this problem, this paper proposes a large language model-based cybersecurity knowledge graph construction method. The method adopts a three-stage architecture. The first stage constructs a cybersecurity domain ontology framework and establishes a large model-driven entity relation extraction method, utilizing knowledge extraction capabilities, chain-of-thought learning, and in-context learning of large language models, combined with Retrieval-Augmented Generation (RAG) to achieve precise extraction of specialized cybersecurity entity relationships. The second stage stores extracted triplet relationships into a graph database to construct a structured cybersecurity knowledge graph, enabling efficient management of security data. The third stage designs a front-end and back-end separated visualization system that displays the knowledge graph through a graphical interface and supports interactive query and analysis. Experimental results demonstrate that the proposed method can effectively extract cybersecurity entity relationships, and the constructed knowledge graph provides comprehensive support for security personnel in attack attribution, vulnerability correlation, and defense strategy formulation.
Federated learning (FL) is a promising solution for healthcare Artificial Intelligence (AI), striking a balance between patient privacy and the need for diverse datasets. FL enables collaborative model training across institutions, preserving confidentiality and advancing clinical tasks such as diagnosis and treatment planning. However, a key challenge in this setting is the inherent heterogeneity of medical datasets acquired in different institutions, which can undermine the generalizability and performance of the model. This issue is particularly pronounced in neuroimaging applications, such as Magnetic resonance imaging (MRI), where site-specific biases arise from variations in scanner hardware, acquisition protocols, and preprocessing pipelines. These differences introduce non-biological variability that can jeopardize downstream analyses and model training. To remove the effect of sites, harmonization techniques are essential tools to improve robustness and reliability. Harmonization techniques are usually applied at the feature level; however, given the limited access to the data possessed by FL schemes, feature-level harmonization may not be enough to remove site effects. In this work, we propose two complementary harmonization strategies within the FL framework: (1) the traditional feature harmonization, by applying ComBat to directly correct the MRI-derived features; and (2) gradient harmonization, which aligns local model updates, particularly the gradients of fully connected layers, across sites to mitigate inter-site distributional shifts before global aggregation. Together, these approaches aim to improve cross-site consistency and improve the model's overall performance in federated medical imaging tasks.
Existing knowledge graph construction (KGC) methods often rely on extensive manual annotation, confined to specific domains and limiting their adaptability to diverse tasks. Language models (LMs) have recently demonstrated exceptional potential for KGC owing to their superior text comprehension and generalization capabilities, addressing the limitations of traditional methods. We propose a topic-enhanced instruction tuning approach, where an instruction-tuned LM extracts structured knowledge from unstructured text. Text topics predicted by deep learning models are incorporated into instructions to guide the LM in identifying domain-relevant entities and relations. Experimental results demonstrate that our approach significantly improves triplet extraction F1-score by 1.23% compared to topic-agnostic baselines, validating the efficacy of topic-enhanced instruction tuning for automatic KGC.
The detection of financial fraud with Graph Neural Networks (GNNs) on Knowledge Graphs (KGs) is a state-of-the-art approach. However, the performance of these models in production is threatened by data drift, where the statistical properties of the data change over time. This study provides a foundational analysis of the robustness of different GNN architectures (convolutional and attention-based) when exposed to natural drift in a KG built from the IEEE-CIS Fraud Detection dataset (Kaggle). We trained three models, Relational Graph Convolutional Network (R-GCN), Heterogeneous Graph Transformer (HGT) and Heterogeneous Attention Network (HAN) -on a historical dataset and evaluated their performance degradation in 50 subsequent temporal windows. Our results quantify a significant and volatile performance drop over time for all models, with the F1-score of HGT falling from a baseline of 0.747 to a minimum of 0.455. In contrast, the simpler convolutional architecture of R-GCN proved to be more resistant to drift, achieving a high average F1-score of 0.868 across the monitored windows. We show that architectural choice directly affects the degradation profile, revealing a trade-off between peak performance and temporal stability. This provides strong evidence that drift analysis is a crucial step in the lifecycle of the GNN model for financial applications, establishing a vital benchmark for future work on drift-aware models.
Adding the next improvement increment to metalforming manufacturing systems is a challenging undertaking nevertheless motivated by possible economic and ecological advantages. Data-driven approaches are expected to be the enabler for this latest increment, allowing, e.g., the implementation of early warning systems for tool wear and its possible impact on product quality. The consideration of domain expert knowledge can guide the implementation of such approaches towards more robust and efficient solutions. Therefore, many efforts have been undertaken to formalize human knowledge resulting, among others, in cause-effect graphs (CEG). However, leveraging this technique in an interactive knowledge base is often limited through their specialized technical realization, which is detached from traditional everyday engineering tools. Thus in this paper, we establish the Cause-Effect Learning System (CELS) as a framework for machine operators, simulation engineers, and data analysts to collaboratively embed their knowledge in a CEG structure. By defining analysis contexts in the CELS, partial CEGs serve as a blueprint for the development of correlation models. The intricate requirements of a data-driven metal-forming system project are elicited and the resulting CELS implementation is discussed based on the domain experts' feedback.
Due to the widespread presence of time series data across various domains, accurate time series forecasting holds significant importance. However, existing complex forecasting models are still affected by issues such as computational efficiency and information bottlenecks. In this paper, we propose a novel deep learning backbone model called IRDNet. Additionally, we introduce a method for decomposing time series based on frequency domain characteristics, which IRDNet utilizes to achieve effective information aggregation and modeling of long-term dependencies. We also conduct a comprehensive evaluation of IRDNet's performance across six datasets, demonstrating that IRDNet is an efficient approach in terms of both accuracy and speed, outperforming state-of-the-art deep learning and statistical methods.
The field of recommendation systems aims to address the problem of information overload by employing classifier and ensemble-based strategies. Early ensemble models explored approaches that made use of user meta-data to create a performance predictor. This predictor then guided the ensemble model in assigning weights to the base recommenders. However, such approaches are not feasible in real-world scenarios where user meta-data is unavailable, such as in ad hoc searching. In this study, we introduce a novel ensemble model called CRESRec, which is both user-agnostic and instance-aware. CRESRec harnesses the power of contrastive learning to acquire representations of the base recommenders from historical interactions. These representations enable CRESRec to capture patterns and characteristics within the recommenders' outputs. Experimental results demonstrate that CRESRec surpasses state-of-the-art recommender ensembles, achieving improvements of up to 38%. This study emphasizes the importance of constructing representations for base recommenders in ad-hoc inquiries and sheds light on the potential of ensemble learning in sessionbased recommendation systems.
Despite their impressive capabilities, Large Language Models (LLMs) struggle to access knowledge not encoded during pre-training. In-context learning (ICL) addresses this limitation by embedding relevant information directly in the prompt, enabling LLMs to use external knowledge without updating their parameters. Recent research has explored integrating knowledge from knowledge graphs (KGs) into prompts, as KGs offer structured and factual representations of concepts and their relationships. A common strategy involves identifying key concepts in the task, grounding them in the KG, extracting subgraphs that connect question and answer concepts, and incorporating the corresponding statements into the prompts. However, a major challenge in ICL is selecting appropriate knowledge, that is, information that supports the model's reasoning and improves performance, while minimizing irrelevant or noisy content that can reduce accuracy. This study investigates how the representation of knowledge in prompts and the relevance and scope of task-grounded KG knowledge affect LLM performance on multiple-choice question answering (MCQA) tasks. We compare triple-based versus path-based representations, assess relevance filtering strategies, and evaluate different knowledge processing approaches. Our findings show that path-based representations outperform triple-based approaches but are more sensitive to noise. While KG-based knowledge can enhance LLM reasoning, imprecise selection of relevant knowledge can degrade performance below the zero-shot baseline (where no additional context is provided), highlighting the challenge of integrating KG knowledge without task-specific selection strategies. Expanding the scope of extracted subgraphs increases recall (retrieving more relevant information) but reduces precision (as more noise is also included). These findings underscore the critical importance of balancing informativeness with noise reduction in KG-enhanced LLM systems. Source code is publicly available at https://github.com/maryam-ghanbari/KGSweetSpot.
Large Language Models (LLMs) face significant challenges when attempting to utilize knowledge not encoded during pre-training. In-context learning (ICL) addresses this limitation by embedding relevant information directly into prompts, enabling LLMs to leverage external knowledge without updating their parameters. Recent advances have explored integrating knowledge graph (KG) information into prompts, taking advantage of the structured and factual representations of entities and relationships that KGs provide. A common approach involves identifying key concepts in a task, grounding them to KG nodes, extracting ego-subgraphs centered on these concepts, and incorporating them into prompts. However, not all subgraphs are equally relevant to the task, and selecting appropriate knowledge remains a critical challenge, as irrelevant or noisy subgraphs can reduce model accuracy. This study investigates how KG-based information related to task-specific concepts influences LLM performance on multiple-choice question answering (MCQA) tasks. We implement a four-stage pipeline that (1) identifies key concepts from questions and grounds them in the KG, (2) extracts individual ego-subgraphs for each concept, (3) integrates these subgraphs into prompts, and (4) evaluates their impact on LLM reasoning via probability scoring. Our analysis highlights which concept-specific subgraphs enhance performance, which introduce misleading information, and which have neutral effects. We show that carefully selected KG subgraphs can substantially outperform others in semantic relevance. Furthermore, we examine how the size and connectivity of selected KG subgraphs influence model performance. Overall, this work deepens understanding of KG-based knowledge selection in ICL and informs the design of more effective, targeted prompting strategies. This project's source code is publicly available at https://github.com/maryam-ghanbari/InformativenessOfKGs.
Recent years have witnessed the increasing research attention on the topic of data-to-text generation(it is also called table-to-text generation), due to its wide range of applications, such as generating textual summaries from structured knowledge graphs. Existing methods usually require a large amount of labeled data to achieve satisfying performance. However, it is very expensive and time consuming to collect enough labeled data in a specific domain to train a model. Though it is possible to pre-train a general model through multiple source domain data and then use the target domain data to fine-tune it, there may exist the gap between each other. To this end, we propose a novel data-to-text generation model, named SMSTL, with utilizing multiple sources in a transfer learning setting. The core of SMSTL is a special designed self-disentangling mechanism to disentangle the domain-specific representations and domainagnostic representations tailored for data-to-text generation. We conduct extensive experimental evaluation to demonstrate that our method can significantly outperform the state-of-the-art methods.
