Structured data is widely used in domains such as industry, finance, and healthcare. However, real-world datasets often suffer from issues such as data sparsity, distribution imbalance, and missing features, which significantly hinder the generalization performance of machine learning models. To address these challenges, this paper proposes a structured data augmentation method based on high-entropy sample masking and large language model-assisted completion. The method first employs a lightweight RandomForest classifier to estimate the prediction entropy of each sample, identifying high-uncertainty instances. For these samples, only key features—selected based on feature importance—are masked, and the masked inputs are then completed using DeepSeek-R1. To ensure the consistency of generated samples with the original data distribution, Kullback-Leibler (KL) divergence is introduced as a filtering criterion to retain only the most distribution-aligned samples. Furthermore, an iterative enhancement framework is established, which terminates automatically when the average KL divergence between generated datasets across two consecutive iterations falls below a defined threshold (0.01). Experimental results on multiple classification tasks demonstrate that the proposed method achieves superior performance in terms of accuracy, stability, and generation efficiency compared to conventional augmentation methods and full-field LLM generation. This work offers an efficient, controllable, and trustworthy solution for structured data augmentation.
The sharing of electronic medical records (EMRs) can significantly improve the quality of medical services and is crucial for disease research and the diagnostic process of doctors. However, due to insufficient trust and concerns over data security, data holders often refuse to share EMRs data with data users. On the premise of ensuring data security, how to motivate data holders to conduct trusted sharing is a major challenge currently faced by the academic community. To address this challenge, this paper proposes a distributed sharing model based on hybrid consortium chains and Stackelberg games. This model integrates the decentralized feature of the consortium chain with the game strategy optimization method to construct a trusted and secure shared framework. By analyzing the decision-making conflicts of the participants and designing the incentive mechanism to ensure reasonable returns, the ADMM (Alternating Direction Method of Multipliers) is utilized to achieve the dual maximization of the profits of data providers and the utility of consumers. The AHC (Agglomerative hierarchical clustering) algorithm was adopted to cluster consensus nodes to improve the traditional PBFT (Practical Byzantine Fault Tolerance) consensus algorithm, enhancing the efficiency of EMRs data sharing. And it has reduced the phenomenon of data silos. The simulation experiments show that the medical consortium chain system constructed in this paper can provide secure EMRs management and transactions, reduce the communication overhead and delay of medical nodes when sharing EMRs data, increase the throughput, and promote the interoperability among medical institutions and relevant departments.
Multivariate probabilistic time series forecasting is crucial for decision-making tasks that rely on future values, uncertainty quantification, and inter-variable dependencies. Recently, attentional copulas have gained popularity for successfully combining the expressiveness of deep learning with the modular structure of copula models, which separate the modeling of marginals from that of the joint distribution. Despite their promise, existing methods such as TACTiS face challenges related to convergence stability and the accurate calibration of dependencies. We propose T-PAKC, a Transformer-based Proportional Attentional Kernelized Copula model to address these challenges through: (1) replacing discrete copula binning with a Gaussian kernelized estimator for smooth gradients and expressive density modeling, and (2) adding a Proportional Absolute Error (PAE) regularizer to improve joint calibration without harming marginals. Experiments on five real-world datasets show T-PAKC consistently outperforming strong baselines in CRPS and Energy Scores, with more stable convergence. This work offers a robust, domain-agnostic framework that combines statistical rigor with deep learning flexibility, advancing scalable, uncertainty-aware probablistic multivariate forecasting.
The analysis of Environmental, Social, and Governance (ESG) data presents significant challenges for traditional question answering (QA) systems. These systems, typically designed for small and well-structured tables, struggle to handle the high-cardinality, ambiguity, and constant updates that characterize real-world ESG datasets. To address this, we present ESGTabQA, a table-based question-answer system capable of handling online, open-domain, natural language queries over high-cardinality ESG tables. The system integrates three key components: an MPNet-based retriever to identify relevant data segments, the TAPAS model to extract answers from tables, and GPT-4 to generate clear responses in natural language. A central innovation of our system is an entity-specific query decomposition that breaks down complex multi-entity queries into parallel, independent sub-queries that align with the data’s structure. This is particularly effective for comparative questions, allowing the system to locate and reason over data spread across multiple sections of a large dataset. We evaluate ESGTabQA on 32 real-world ESG queries, covering simple, ambiguous, and comparative scenarios. The proposed system achieves strong results, including a semantic similarity score of 0.876, 79.5
Graphs are widely used to model complex relationships among entities in domains such as recommendation systems and social networks. A fundamental problem in these applications is quantifying the connectivity between two entities, often formulated as estimating the number of source-to-target (s–t) paths under specific constraints. In this paper, we study the problem of hop-constrained s–t path counting, which aims to estimate the number of paths from a source vertex s to a target vertex t within a given hop bound k. We propose kPathAssess, an efficient path counting algorithm inspired by the recent PathAssess algorithm. kPathAssess balances estimation accuracy and computational complexity by leveraging locally computable lower bounds on path counts. Extensive experiments on real-world graphs demonstrate that kPathAssess achieves high estimation efficiency with significantly reduced computation time, particularly as the hop constraint increases.
The occluded person re-identification (ReID) remains a challenge due to feature loss in occluded regions, which not only introduces noise interference but also poses significant difficulties in feature alignment. To systematically address this issue, we propose an innovative Pose-guided Feature Restoration Transformer (PRT) framework, designed to adaptively restore features within occluded areas. Our approach commences with an occlusion sample augmentation strategy. By integrating real-world images with noisy images, we simulate a wide spectrum of complex occlusion scenarios, enhancing the model’s generalization ability across diverse occlusion conditions. Subsequently, based on visibility scores, we develop a method that fuses common and unique features to recover the characteristics of occluded keypoints. To further optimize feature transfer, an adaptive directional graph convolution method is introduced to construct higher-order feature representations, effectively alleviating feature degradation caused by occlusion. Finally, leveraging the attention mechanism of the Transformer, we integrate both global and local contextual information to enhance the overall feature representation comprehensively. Comprehensive experiments conducted on both occluded and holistic datasets demonstrate the superior effectiveness of our PRT framework.
Given a start point s, a destination t, and the set of keyword types that must be satisfied, the goal of Top-k Keywords-Aware Optimal Route (k-KAOR) query is to return top-k shortest routes that covers all query keyword types. Existing solutions can be divided into two classes: candidate based algorithms and path expansion based algorithms. However, they all face a trade-off between query efficiency and route quality. To solve this problem, in this paper, we further study the k-KAOR problem and introduce a Splice-Based Expansion (SBE) algorithm. The SBE algorithm efficiently integrates the ideas of candidate set and path expansion. First, to avoid a vast search space, we introduce a Point of Interest (POI) filtering refinement strategy based on the spatial index IG-Tree. Then, to cover more combinations of POIs while avoiding brute-force enumeration, we pre-construct a keyword distance matrix and splice POIs using the graph partitioning of the IG-Tree. Finally, we perform path expansion by concatenating different POI combinations until the top-k paths are discovered. Extensive experiments on real-world road networks show that our method can reduces response time by an order of magnitude and improves result quality by 25
In this paper we apply the aspect-oriented programming paradigm to GraphQL, a query language for web services. Aspect-oriented programming is a technique to quickly and easily add support for cross-cutting concerns to programs. Data also has cross-cutting concerns such as provenance, versioning, and testing. These cross-cutting data concerns can be represented as advice that annotates data. This paper shows how to add support for aspect-oriented advice to GraphQL schemas and queries to create Aspect-oriented GraphQL. We describe a software architecture that leverages GraphQL itself to implement Aspect-oriented GraphQL.
In recent years, many studies have focused on improving knowledge graph completion (KGC) using pre-trained language models (PLMs). Most PLM-based methods focus on encoding textual triples but fail to account for relation directionality and structural context, which are crucial for capturing complex relational semantics. To address this limitation, we propose BiRKGC, a Bidirectional Relation and Graph-Aware model for KGC. The framework consists of three collaborative components: a Bidirectional Relation Encoder (BiR) for explicit direction modeling, a Graph-Aware Sequence Encoder (GaS) for integrating subgraph-level structure, and a Masked Token Modeling (MTM) task for semantic pre-training. Together, these components enhance the model’s ability to fuse textual semantics with structural information for more effective reasoning. In addition, BiRKGC introduces a dynamic subgraph sampling mechanism to improve structural adaptability. Experiments on WN18RR and FB15k-237 show that BiRKGC consistently improves MRR, Hits@1, and other metrics over strong baselines, while also narrowing the performance gap between forward and backward reasoning. Ablation studies validate the effectiveness of each module, demonstrating BiRKGC’s potential and competitiveness in KGC tasks.
The majority of current approaches to fake news detection seek to improve accuracy by leveraging social context based on pair-wise graph structures. However, these methods encounter two limitations. Firstly, simple graph structures are inherently limited in their ability to capture the complex relational dynamics that exist beyond dyadic associations within social networks. Secondly, social context-based fake news detection research, predominantly focused on user engagement, overlooks the crucial relationship between news content and its author. The author-topic connection is relevant to fake news detection, as author credibility is topic-dependent. To address these, we utilize our model to construct a hypergraph and introduce a hypergraph neural network that reflects both the engagement of spreaders with news, the relationships between news, spreaders, authors, and topics, and, indirectly, the author-topic connections via hyperedge interactions. Experiments on two real-world fake news datasets show that our model outperforms other GNN-based models.
Web-based dashboards are fundamental to digitalization, serving as central components in various organizational processes. However, their development is typically restricted to software engineering teams who must translate domain-specific requirements from stakeholders into functional interfaces. Despite this process often being repetitive and structured in nature, it remains complex and largely inaccessible to non-experts. While template-based tools such as Power BI increase accessibility, they frequently lack the flexibility needed for highly customized and adaptive solutions. To overcome these limitations, we present DIVE (Data Insights and Visualization Engine), a novel stakeholder-driven, hybrid neuro-symbolic framework for generating highly adaptable dashboards. By leveraging natural language interaction and flexible Knowledge Graph (KG) structures, DIVE empowers users to define, query, and visualize data without programming knowledge. At its core, DIVE integrates an interaction of symbolic and neural techniques together with a Retrieval-Augmented Generation (RAG) architecture that integrates coding-relevant context to generate precise queries and synthesize dynamic User Interface (UI) components in real time. To assess DIVE’s effectiveness, we conducted a mixed-method evaluation using the NASA Task Load Index (NASA-TLX) and the System Usability Scale (SUS), supplemented by participant and expert interviews for qualitative insights. The results demonstrate that DIVE supports completing complex analytical tasks with high usability and low cognitive load. Study participants and experts emphasized the system’s innovative approach, practical applicability, and potential for substantial time and cost savings, thus confirming its capacity to produce customizable, future-proof dashboards.
Modern recommender systems usually rely on inner product search to find relevant items. The inner product inherently considers both the norm and direction of vectors, which naturally aligns well with purchasing behavior. Unlike the metric space, the inner product space does not follow the triangle inequality and the identity of indiscernibles. This makes search in the inner product space more challenging as the neighbor transferring is not as stable as metric space, which directly affects the connectivity and efficiency of graph-based methods. Existing representative methods, including ip-nsw and its successors, are insert-based graph construction algorithms. These insert-based algorithms build the edges between non-candidate points which makes them unsuitable for maximum inner product search (MIPS). In this paper, we analyze neighbor transferring in the inner product space and find that normalized points can improve the efficiency of neighbor transferring. Based on this idea, we propose a new inner product search method that has the following advantages: (i) It constructs a highly accurate proximity graph that can help the inner product search and reduce the index size to a minimum of 1
The main objective of the analysis is to assess the effectiveness of current neuro-symbolic approaches, with a focus on their capabilities in knowledge representation and reasoning in latent space for web applications. Today, neuro-symbolic approaches are increasingly influential in bridging neural representation learning and symbolic reasoning, particularly for knowledge graph completion and semantic interoperability in web engineering. A central challenge in this field is the lack of methods that simultaneously link word embeddings capturing semantic information from unstructured text and concept embeddings encoding formal knowledge from ontologies such as OWL and RDF(S), while enabling unified approximate deductive inference across both modalities. This disconnection limits the potential to perform integrated reasoning over textual data and structured knowledge bases. In this paper, we analyze whether state-of-the-art neuro-symbolic inference methods can represent symbolic entities in a shared vectorial space while preserving the formal semantics dictated by ontologies. Our analysis examines how well these models align the geometry of learned embeddings with ontological constraints, and whether this alignment supports effective inference over knowledge graphs. Finally, we conclude with a discussion of open research directions, including the need for joint embedding models that combine neural and symbolic reasoning for heterogeneous web data, as well as diversification strategies to enhance inference robustness and retrieval effectiveness.
Multivariate time series imputation is vital for domains such as meteorology, finance, and transportation. However, most existing approaches employ a separated imputation-then-forecasting pipeline, which struggles to model complex spatio-temporal dependencies effectively and tends to incur significant error accumulation, especially in scenarios with high missing rates. To address these challenges, we propose Mamba-based Time Series Completer (MambaTSC), an end-to-end framework that integrates time series imputation and forecasting within a unified state space modeling architecture. Specifically, we first devise a Multi-Scale Temporal Enhancement Mamba (MS-Mamba) module, which augments state space modeling with multi-scale temporal feature extraction, enabling the capture of both short-term fluctuations and long-term trends. Additionally, we develop a Score-Gated Graph Convolutional Network (SGCN) that dynamically adjusts information propagation based on observation reliability, thereby improving spatio-temporal representation learning. Extensive experimental results on three real-world datasets show that our proposed MambaTSC achieves state-of-the-art performance, demonstrating superior effectiveness and robustness for time series with substantial missing data.
Temporal Knowledge Graph (TKG) reasoning aims to infer missing facts based on historical information. Pre-trained Language Models (PLMs) have excelled in knowledge-intensive tasks, motivating numerous studies to explore their semantic modeling capability for TKG reasoning. However, existing PLM-based methods primarily focus on temporal semantic modeling while neglecting structural dependencies and global historical patterns, which limits their effectiveness. To address these challenges, we propose PSG, a PLM-based framework that takes into account Structure and Global information for TKG reasoning. Specifically, we design specific contextualized time-aware soft prompts to combine and balance semantic and structural information, which are then processed by both the frozen PLMs and the graph model. Besides, based on our analysis, we capture timely and relevant global information by utilizing global frequency, which emphasizes interactions involving recent historical facts. We evaluate PSG on three datasets and conduct extensive analysis, validating the effectiveness of PSG in TKG reasoning.
Knowledge Tracing (KT) is a key task in online education that aims to model the dynamic knowledge states of students. To achieve high-quality KT, question difficulty estimation plays an important role during the information extraction phase. However, most existing methods based on statistical models face two major challenges: (1) unstable estimates across different datasets, and (2) the cold-start problem when estimating the difficulty of new questions. To address these issues, we propose a Retrieval-Augmented Generation-based Knowledge Tracing framework (RAG-KT). This framework introduces a novel RAG-based difficulty estimation method that generates stable and accurate difficulty scores, and can estimate the difficulty of unseen questions by referencing semantically similar ones. The LLM-generated scores, along with other relevant information from students’ learning histories, are then fed into a KT model for the prediction task. Specifically, to improve the quality of difficulty estimation, we apply contrastive learning to fine-tune a domain-specific embedding model for RAG in educational contexts. We also carefully design task-specific prompts to guide LLMs for better difficulty estimation performance. Extensive experimental results demonstrate that RAG-KT achieves stable difficulty estimation and effectively addresses the cold-start problem. With more reliable difficulty estimates, the performance of KT models can be significantly improved.
Data augmentation is widely employed in sequential recommendation to mitigate data sparsity issues without requiring additional supervised labels. This process generally consists of two stages: augmented data construction and augmented data filtering. Although existing methods have investigated various effective strategies for both stages, they still encounter two primary limitations: (1) limited semantic diversity during the augmented data construction stage, and (2) inefficient filtering mechanisms during the augmented data filtering stage. To address these limitations and achieve more effective data augmentation, this paper proposes Semantic-guided Data Augmentation and Filtering for Sequential Recommendation (SDARec), which leverages semantic information of items to guide both the construction and filtering of augmented data. In the augmented data construction stage, dataset retrieval augmentation and sequence retrieval augmentation strategies are designed to construct semantically controllable and diverse augmented data. In the augmented data filtering stage, a large language model is employed to evaluate and filter the augmented data through a chain-of-thought prompting approach, eliminating the need for expert knowledge or complex heuristic designs. SDARec performs augmentation at the dataset level without requiring model training, making it highly generalizable and adaptable to various conventional recommendation models. Extensive experiments on three public datasets demonstrate the effectiveness of SDARec. Source code is available at https://github.com/zzzzzdev/SDARec .
A good recommender system can enhance sales both offline and on the Web, while lowering maintenance and labour costs. Recent research shows that hybrid recommender systems, built on memory-based and knowledge-based techniques, deliver good results. Hybrid recommender systems leverage strong predicting power both for existing users as well as for new users, making these methods useful for a wide range of applications. In this paper, we use a Natural Language Processing (NLP) embedding technique word2vec in the context of sequential sales data, incorporating meta-data from multiple sources. We extend the Meta-Prod2vec method with jointly trained embeddings for users and related meta-data. The embeddings are made on a unique dataset representing sequential sales and website behaviour data in multiple European countries. The embeddings and features are concatenated and fed to a Neural Network (NN) to model consumers’ behaviour. The proposed Meta-User2vec method, which advantages from both product and user meta-data, outperforms existing methods Meta-Prod2vec and user2vec.
In the realm of urban mobility, the increasing availability of large-scale transportation data has paved the way for advanced and sophisticated routing queries. Among these, the time-dependent constrained path optimization (TD-CPO) problem optimizes route selection based on user-defined preferences such as safety, navigability, road width, or scenic value, while adhering to specific constraints (travel time). This problem has broad applications in urban navigation, helping commuters and travelers find paths that align with their individual needs. However, solving the TD-CPO problem is challenging, as it can be reduced to the NP-hard arc-orienteering problem. Existing state-of-the-art research has predominantly focused on solving the TD-CPO problem for a single departure time. However, this approach is often too limiting for users who prefer to depart within a time range. Extending these methods to handle a range of departure times typically results in redundant computations and inefficiencies. This paper proposes an extension of the TD-CPO query, called interval TD-CPO, that accounts for a range of departure times. This extension aims to identify the most suitable path and a corresponding departure time that maximizes the user’s preference metric within the given constraints. Experimental results on real-world graphs show that our algorithm delivers comparable solution quality in just 5 s, a major improvement over the 120 s needed by state-of-the-art methods adapted for interval TD-CPO.
Recently, domain-specific large language models (LLMs) have been rapidly developed and prosperous in many specialized fields, yet their performance in traditional Chinese medicine (TCM) remains unclear because of the lack of a reliable and user-friendly benchmarking framework. In this paper, we propose TCM-Eval, an automated, scalable, and open-source system for evaluating LLMs on TCM scenarios. In TCM-Eval, we design six core dimensional tasks modules plus 1 integrated module, and standardize a total of 6,813 questions covering 29 specialized data sources. Furthermore, we integrate 16 LLMs, including the popular commercial, open-source, and state-of-the-art proprietary LLMs, into TCM-Eval. To provide a fine-grained and systematic assessment, we also design 4 multi-dimensional metrics across 7 types of questions. Our framework shapes a new standard for TCM LLM evaluation, offering a robust foundation for the relevant studies and clinical applications.
