Multimodal Knowledge Graphs enhance traditional symbolic knowledge graphs by incorporating complementary data such as text and images, enabling richer semantic inference. However, integrating heterogeneous modalities remains challenging due to their distinct geometric and semantic characteristics. We propose CurvBi (Curvature-aware Bicomplex), a novel framework for multimodal knowledge graph completion that unifies modality alignment and relational modeling under a coherent mathematical foundation. CurvBi leverages bicomplex embeddings to capture symmetric, asymmetric, and cyclic relationships in a four-dimensional algebraic space, and applies Riemannian manifold fusion via Fisher-Rao distance to align modality-specific embeddings within a curvature-aware structure. Unlike prior approaches relying on Euclidean projection or shallow concatenation, CurvBi offers a geometry-sensitive and algebraically expressive representation space. We evaluate CurvBi on four benchmark datasets, where it consistently achieves competitive performance across multiple evaluation metrics, with improvements reaching up to 6
Recommendation systems effectively model user preferences through historical behavior data, but their reliance on such data exposes vulnerabilities to adversarial attacks. Transfer-based attacks craft adversarial examples (i.e., fake data) in surrogate models to manipulate unknown target systems. Existing methods generate fake users located in regions of sharp loss landscapes, causing them to be extremely sensitive to model parameters and difficult to transfer to unknown models. The flatness of the loss landscape is crucial for transferable adversarial attacks because flat regions exhibit small gradients, making adversarial examples less sensitive to parameter variations between different recommendation models. This property enhances the transferability of adversarial examples across diverse target systems. To address this, we propose Recommendation-specific Gradient Penalty (RecGP), which steers adversarial examples towards flat loss regions by regularizing gradient magnitudes during optimization in recommendation systems. Moreover, to enhance the resource efficiency of RecGP, we further develop Recommender Systems Gradient Penalty with Resource-efficient Sampling (RecGP-RS) in discrete neighborhoods. RecGP-RS employs semantic-aware neighborhood sampling to handle discrete recommendation data, approximating second-order gradients via first-order interpolation while preserving semantic consistency in perturbations. Extensive experiments on MovieLens-1 M and Gowalla datasets demonstrate that RecGP achieves average attack success rate improvements of approximately 8
Probabilistic topic models are a conventional standard for discovering hidden semantics from text documents. However, when applied on short texts, existing models work poorly and yield incoherent and repetitive topics because of the limited word co-occurrence information. Several attempts have been made in this regard, where some methods have been proposed to integrate supervision into the learning process, but these models are still unable to infer coherent topics that correlate with human judgment. To address this issue, we propose an embedded supervised topic model that encodes word embeddings and latent topic representations guided by class labels on the hypersphere. Unlike existing supervised topic models that focus on target prediction and neglect topic interpretability, our model integrates pre-trained word embeddings to infer interpretable topics. The resulting spherical representations can be used for prediction tasks, as the inferred topics are indicative of document labels. To further enhance the interpretability aspect of topics, we propose a second framework that integrates knowledge graph embeddings into our probabilistic model, which is crucial to further palliate the data sparsity problem. Experimental results on four benchmark datasets show that our proposed models outperform existing models on topic interpretability, while having competitive label prediction capability.
In the realm of knowledge distillation (KD), a method widely acknowledged for compressing and regularizing models, there has been growing interest in its application to Graph Neural Networks (GNNs), which often suffer from the overfitting issue, significantly limiting their generalizability. Recent KD techniques aim to alleviate this overfitting by enhancing GNN model capabilities. However, conventional KD approaches face challenges in effectively leveraging well-trained GNN teachers for student training, notably due to the oversmoothing phenomenon in GNNs, and conventional self-distillation techniques suffer from the information bottleneck problem. To address the overfitting issue in GNNs, this study introduces a novel memory-augmented self-learning framework that extracts and provides diverse learning sources for adaptive knowledge distillation from the student model itself. A competency-based knowledge source selection mechanism dynamically determines the most pertinent source. Unlike existing self-distillation approaches, our method demonstrates improved learning with enhanced regularization, resulting in a 2.5-6
Several classifiers operate exclusively on discrete variables, requiring continuous attributes to be preprocessed through discretization. Supervised discretization addresses this need by partitioning the domain of continuous variables into intervals that define a discretization policy. Federated learning, in turn, enables the collaborative training of global models without sharing raw data: users train local models and communicate their parameters to a central server, which aggregates them into a global model. A key challenge in this setting is achieving performance comparable to centralized methods while minimizing communication overhead. In this work, we propose a federated discretization framework that computes a global discretization policy in a single communication round. Each user estimates Gaussian mixture models for the continuous variables and transmits their parameters-referred to as representatives-to the server. The server then applies a discretization algorithm to these representatives to derive a global policy, which is subsequently distributed back to the users. The proposed framework accommodates multiple discretization strategies; specifically, we present federated adaptations of equal-width, equal-frequency, entropy-based, and kernel density-based methods. Experiments conducted on ten benchmark datasets demonstrate that the proposed framework achieves performance comparable to centralized discretization approaches, with the kernel density-based and equal-frequency variants yielding the most competitive results.
Knowledge Graph (KG) summarization is the process of producing compact representations to improve KG comprehension, support efficient search, and facilitate downstream applications. Existing summarization techniques have made significant progress in preserving either semantic or structural information. However, they cannot directly control trade-offs between the two dimensions in a lossy setting. Existing lossy summarization approaches evaluate information loss after the summary is generated (post-hoc); but they lack a mechanism to proactively control loss during the summarization process. This study introduced a dual-utility metric that evaluates both structural preservation and semantic cohesiveness. Three algorithms were developed: LIDUS, LDUS, and aLDUS. The first algorithm performs lossless summarization while preserving structural and semantic integrity. The second is a utility-guided lossy summarization method that iteratively merges ranked entity pairs, minimizing utility loss at each iteration. The third is an optimization algorithm that reduces time complexity by incrementally computing semantic loss, while guaranteeing a bounded utility loss. Experiments were conducted on five real-world and synthetic KGs to demonstrate the accuracy of the top-10
A transient community is a short-lived group of nodes in a temporal or dynamic network that forms suddenly, persists briefly, and disappears quickly as the network evolves. Detecting such groups is especially valuable for targeted marketing because they reflect short-term bursts of collective interest where users are most responsive to relevant information. However, most existing community detection methods in dynamic networks integrate all time network timesteps into a single community structure, thereby overlooking short-lived communities that emerge and dissolve around specific events, trends, or transient behaviors. To address these challenges, a novel evolutionary incremental approach called Event Adaptive Incremental Learning Non-Negative Matrix Factorization (EA-iNMF) is proposed. The proposed method integrates event-adaptive learning with an incremental update mechanism to efficiently identify optimal time spans corresponding to peak user engagement. To capture structurally cohesive regions, an event-centric core–periphery extraction strategy is introduced, where core nodes are selected based on the Temporal Core Index (TCI), and maximal biconnected components around high-TCI nodes are identified as highly interactive regions. These core–periphery subgraphs are then incrementally combined with adjacent snapshots to model community evolution. Furthermore, a modularity-based candidate window selection mechanism is designed to identify time intervals with maximum community cohesiveness. Extensive experiments on four real-world networks demonstrate that EA-iNMF achieves an average runtime improvement of 8 to 10
Machine learning increasingly drives decisions in domains such as finance and healthcare, where ethical considerations, such as fairness, are central. In such contexts, ensuring fairness is essential, especially when decisions impact individuals and social groups. Federated learning (FL) provides a decentralized training paradigm, yet client heterogeneity and demographic imbalance can amplify disparities across subpopulations. Existing fairness-aware FL methods remain limited, often focusing on group fairness in binary classification and lacking explicit control over the trade-off between fairness and predictive performance. We introduce FedFairLAB, a FL method that enforces group, intersectional, and multiclass fairness simultaneously at both the local and global levels. A tunable performance budget allows practitioners to control how much predictive performance can be sacrificed to improve fairness. Experiments on six real-world datasets show that FedFairLAB substantially improves fairness while keeping models accurate and usable in realistic FL settings.
Constraint-based Deep Active Clustering (CODAC) integrates actively selected pairwise constraints into deep representation learning to efficiently improve existing cluster structures, even under tight query budgets. CODAC encodes the constraint information into the embedding so that the learned representation can generalize to unconstrained data, leading to a rapid improvement of the clustering quality even on large datasets. CODAC makes minimal assumptions regarding the data and can be combined with a wide variety of deep clustering models. It does not require the number of clusters to be known a priori and is even effective if the initial estimate is badly misspecified. Across diverse image, text, and tabular datasets, CODAC consistently attains higher cluster quality with fewer queries than the previous state-of-the-art, and can substantially improve the clustering quality with just 100–200 queries compared to the deep clustering baselines.
The heterogeneous layers of attributed multiplex networks (AMNs) encode structurally distinct and spectrally diverse relational patterns. However, traditional Graph Neural Network models often overlook this frequency diversity, applying uniform filtering and layer fusion strategies that fail to capture cross-layer complementarity. We propose AMN-ARMA, a frequency-adaptive representation learning framework that models each layer as operating in a distinct spectral regime. It combines adaptive ARMA (AutoRegressive Moving Average) filters tailored to layer-specific frequency responses with multi-band, entropy-regularized attention to ensure balanced and interpretable integration. We further introduce a training objective that encourages frequency specialization across layers while maintaining both intra- and inter-layer consistency. Experiments on three real-world AMN datasets demonstrate that AMN-ARMA achieves up to 1.8
Named Entity Recognition (NER) in specialized domains like biomedicine suffers from acute data scarcity, requiring expensive expert annotations. While data augmentation offers a promising solution, it inevitably introduces noisy and mislabeled samples that can degrade model performance. This problem is amplified in few-shot scenarios where every training example matters. We introduce PALAUNER (Policy-based Active Learning to Augment Named Entity Recognition), a reinforcement learning framework that learns to select high-quality samples from augmented data pools. Using a deep Q-network, our agent evaluates samples based on content features and model predictions, deciding which examples will improve NER performance. Experiments across five BioNER benchmarks demonstrate that PALAUNER consistently enhances diverse augmentation methods, from simple perturbations to GPT-based generation. Average F1 improvements are of 0.5−7.1 points in few-shot settings. PALAUNER’s modular design enables seamless integration with emerging augmentation techniques, providing a generalizable solution for training data quality enhancement. We publicly release our code on GitHub: ( https://github.com/picuslab/palauner ).
Continual test-time adaptation (CTTA) is the task of deployed models adapting to the massive data stream with potential distribution shifts in the test environment to maintain generalizability over time. Current research has primarily focused on two major challenges, catastrophic forgetting and error accumulation, while paying relatively little attention to adaptation efficiency. Moreover, although distribution shifts occur in a wide range of forms, most existing methods often ignore the differences among these shift types and rely on unified optimization strategies that remain insensitive to the magnitude of the shift, making them prone to overfitting for mild shifts and insufficiently adaptive under severe shifts. To address the above problems, a Shift-Level Regulated Continual Test-Time Adaptation Framework (SLRF) is proposed to determine the shift level and subsequently select the most appropriate adaptation strategy from a set of well-designed options. Therefore, SLRF mainly consists of two modules: an adaptive shift-level detector (ASLD) and an adaptation strategies decision module (ASDM). Considering the influence of distribution shifts on both generalization and adaptation, ASLD categorizes shift level into three types: aligned, weak shift, and strong shift. Based on the reliable ASLD results, ASDM selects the adaptation strategy that is constructed according to the detection criteria for efficient adaptation. Specifically, aligned samples are leveraged to maintain long-term adaptability; weak-shift samples are used to further improve generalization; strong-shift samples are utilized for domain alignment to avoid error accumulation and enable rapid adaptation. Furthermore, SLRF provides a stable mechanism for transitioning across different shift scenarios and preventing catastrophic forgetting, while maintaining low computation overhead. Extensive experiments on CIFAR10C, CIFAR100C, and ImageNet-C demonstrate that SLRF achieves superior performance in both generalization and adaptability.
In this study, we introduce MO-MAB-KG, an adaptive Multi-objective multi-armed bandit framework designed to enhance retrieval strategies in Knowledge Graph-augmented, Retrieval-Augmented Generation (KG-RAG) systems operating under non-stationary conditions. The proposed framework addresses two key challenges in real-world deployment: (1) adapting to dynamic environments driven by shifting query distributions and continual updates to knowledge graphs, and (2) simultaneously optimizing multiple, often conflicting objectives such as accuracy, retrieval recall, and latency. MO-MAB-KG offers three principal innovations: (i) a contextual bandit model that dynamically selects among dense, sparse, knowledge graph-based, and hybrid retrieval strategies based on query-specific features; (ii) a reward function grounded in the Generalized Gini Index to support principled multi-objective optimization; and (iii) an online learning mechanism that incrementally updates retrieval strategies based on implicit user feedback. Experiments on real-world QA datasets from a Nexacro-based GenAI chatbot assistant service demonstrate that MO-MAB-KG improves the hit rate by 12.7
Knowledge Graph Completion(KGC) tasks play a crucial role in applications such as intelligent question answering and recommender systems by inferring missing entity relationships from known knowledge graph information. However, existing research faces three key limitations: (1) insufficient ability to capture and represent textual information, (2) inability to effectively handle complex relationships, such as many-to-one and one-to-many, and (3) failure to effectively learn the structural information of the knowledge graph. To address the above challenges, this paper proposes Text Embedding Optimized Knowledge Graph Completion (TEOKGC). First, the model employs a dual text encoder, utilizing the pre-trained Sentence-BERT language model to enhance the semantic representation of text features. Second, to improve the model’s ability to handle complex relationships, such as many-to-one and one-to-many, this paper introduces a deformable adaptive attention module to uncover latent text features more effectively. Finally, to enhance the learning of knowledge graph topology, this paper introduces a path learning optimization mechanism that improves text embedding by computing inter-entity path distances based on triad similarity fusion. The proposed model, TEOKGC, is evaluated on multiple public datasets, WN18RR, FB15k-237, and Wikidata5M. Experimental results demonstrate that TEOKGC achieves significant improvements, verifying the effectiveness and potential of our model. Our code is available at https://github.com/herenohere/TEOKGC .
Graph Neural Networks (GNNs) struggle on heterophilous graphs where edges connect dissimilar nodes. Existing rewiring methods increase homophily via heuristics without theoretical guarantees, fundamentally transforming graph character. We propose LIGR (Label Informativeness-Guided Rewiring), which maximizes an information-theoretic measure quantifying how much neighbors’ labels reveal about node labels. Building on the theoretical framework of LIMO et al. (2025), we extend the binary-class threshold analysis to multi-class settings, establish provable conditions for beneficial edge modifications, and prove convergence guarantees. Experiments on 9 benchmarks across three GNN architectures (GCN, GraphSAGE, GAT) show consistent improvements, with LIGR achieving best average performance and excelling particularly on citation networks (Cora: +6.5 |Δ h| = 0.029 vs. 0.071–0.080) while maintaining competitive accuracy, suitable for applications requiring interpretability (biological networks, social graphs). Code: https://github.com/smlab-niser/ligr .
Causal discovery from observational data alone in the presence of unobserved common causes is crucial yet challenging. We categorize the causal relationship between two random variables X and Y into the following four categories: two direct-case cases ( X → Y or X ← Y ), a case in which X and Y are independent, and a case in which X and Y are confounded by unobserved common causes C , and aim to decide among them from observational data alone. We call this the Reichenbach problem, since this categorization is valid if the Reichenbach’s common cause principle is assumed to be true. Although many methods have been proposed to address causal structure learning in the presence of C , they typically impose assumptions on the nature of C that require knowledge of how unobserved confounders behave, and it is difficult to guarantee in many practical cases compared to assumptions about observed variables. In our previous study (Kobayashi et al. in: 2022 IEEE international conference on big data (Big Data). IEEE, pp 45–54, 2022), we proposed a causal discovery method without such assumptions regarding the nature of unmeasured confounders, named 𝖢𝖫𝖮𝖴𝖣 , for discrete data. Building on the Reichenbach’s common cause principle, if neither the two direct-cause models nor the independence model can explain the data well, we can infer the involvement of unobserved common causes in generating X and Y . To realize this idea in practice, 𝖢𝖫𝖮𝖴𝖣 imposes structural assumptions on the direct-cause models such as additive noise models ( 𝖠𝖭𝖬 s) (Peters et al. in: Proceedings of the thirteenth international conference on artificial intelligence and statistics. JMLR workshop and conference proceedings, pp 597–604, 2010), while allowing the confounded model to represent any joint distribution on ( X , Y ) by placing no assumptions on C . 𝖢𝖫𝖮𝖴𝖣 then employs the normalized maximum likelihood (NML) codelength (Shtar’kov in Problemy Peredachi Informatsii 23(3):3–17, 1987) as an information criterion and compares models of varying capacities. This study extends 𝖢𝖫𝖮𝖴𝖣 to include all data types. We provide NML formulations and theoretical guarantees for consistency in model selection for mixed and continuous cases. Through extensive experiments on both synthetic and real-world data, we demonstrated that 𝖢𝖫𝖮𝖴𝖣 is more effective than existing methods in inferring causal relationships.
Counterfactual explanations (CFEs) have emerged as a key tool in eXplainable Artificial Intelligence (XAI) for interpreting complex machine learning and deep learning models. However, most CFE methods neglect the high computational cost of generating explanations. This limitation can be particularly severe for high-dimensional data such as time series. Moreover, many time-series CFE approaches treat validity, the requirement that the counterfactual actually changes the model prediction, solely as an objective to be optimized rather than as a strict constraint, often leading to invalid explanations and limiting practical applicability. In this work, we propose FastPACE, an efficient method tailored to the generation of CFEs for time-series classification that includes an invalidity penalty to guide the search, while enforcing validity procedurally. FastPACE substantially reduces the runtime of current state-of-the-art methods without compromising the quality of the explanations. Extensive experiments on datasets from the UCR and UEA repositories show that FastPACE matches, and in several cases improves, the explanation quality of existing approaches while being significantly faster.
To characterize global seismic activity accurately, monitoring systems require the onset times and amplitudes of specific seismic phase arrivals in observed waveform data. However, phase amplitudes remain difficult to measure automatically, with current methods achieving only 64% accuracy and leaving 36% of windows to manual correction by experts, creating a major real-time bottleneck in seismic monitoring workflows. To address this limitation, we introduce two knowledge-driven deep learning systems that learn expert decision-making patterns from analyst-corrected seismic data, thus improving automated amplitude measurement. Our approach leverages a U-Net architecture and a novel hybrid loss function that encodes expert preferences for pixel accuracy, shape regularity, and temporal consistency in amplitude window selection. We evaluated 80,648 seismic signal windows to assess the capabilities of our deep learning models. The models achieved substantially higher accuracy in amplitude measurement than traditional rule-based systems. When their outputs diverged from analyst-selected labels, we observed considerable variability among analysts, highlighting the subjective nature of the task. However, independent seismologists showed a strong preference for our model outputs. These findings suggest that our deep learning models successfully capture expert decision-making patterns, demonstrating effective knowledge transfer in an inherently subjective domain.
Features with considerably larger or smaller observations than the rest of the dataset, causing noticeable skewness in the feature distributions, are prevalent in practical applications. Traditional clustering methods often assume symmetric data, leading to poor performance with skewed features. This challenge becomes further complicated in datasets with low separation between true clusters in the feature space. These problems are encountered in a wide range of important practical areas, such as cell grouping, forest fires, maritime search and rescue, urbanization studies, and neuroimaging. Bayesian model-based clustering methods can accurately capture the skewness in the data and centers of poorly separated true clusters. However, they are computationally inefficient due to their Bayesian nature. We propose a Bayesian model-based clustering framework to address these issues by utilizing the generalized multivariate log-gamma distribution with a Dirichlet process mixture. Comparative numerical experiments on 30 benchmark datasets with traditional and Bayesian model-based clustering algorithms demonstrate the superior performance of the proposed method, particularly for skewed datasets with low true cluster separability. The proposed approach, implemented in R, also shows better computational efficiency than its Bayesian alternatives. The computer codes to implement our approach are provided to facilitate practical applications.
The cryptocurrency market represents a significant innovation in the financial ecosystem, built upon cryptographic principles to ensure secure and transparent transactions. Cryptocurrencies experienced a global adoption, driven by their decentralized nature that enables borderless transactions without third-party intermediaries. The price of cryptocurrencies is characterized by a significant volatility, that introduces both opportunities and challenges. In this context, the development of accurate methods for the forecasting of price variation, able to work in real-time on data streams, has become vital for various stakeholders. In this paper, we propose a novel approach, called LEMON, for the online prediction of the price variation of cryptocurrencies, that leverages possible temporal correlations among them. Our approach stems from the empirical evidence that cryptocurrencies tend to form groups characterized by similar trends, a behavior often attributed to shared market dynamics and common external factors. Through the analysis of temporal correlations, LEMON dynamically identifies these groups, that are then exploited to learn multiple multi-target tree-based models, specifically designed for processing continuous data streams. LEMON also introduces a novel adaptive non-parametric weighting scheme, that automatically adjusts the importance of each instance based on the observed data distribution in real-time, improving the forecasting of the price variation. Our experiments, performed on 16 datasets related to 16 cryptocurrencies, demonstrate that LEMON outperforms state-of-the-art approaches in two distinct prediction tasks: forecasting the closing price variation (regression) and predicting the market trend direction (classification), making it an effective tool to support stakeholders requiring accurate real-time predictions.
