While knowledge graphs have emerged as a powerful paradigm for representing and querying interconnected data, distributing them across multiple untrusted nodes introduces significant challenges in guaranteeing query result integrity. We propose a verifiable query processing framework for distributed property graphs that constructs a Merkle Patricia Trie (MPT) over distributed data chunks, enabling efficient cryptographic proof generation and client-side verification without requiring full data retrieval. Experimental evaluation confirms that the verification overhead remains practical, and the framework scales effectively to realistic graph workloads.
Evaluating machine learning models that generate time-dependent predictions in clinical settings remains challenging, as conventional metrics such as time-dependent AUC rely on observed event information and do not account for undocumented intermediate clinical events. In this work, we propose a novel evaluation framework based on class–cluster alignment. The method uses over-time clustering to identify structural patterns in such predictions and introduces alignment-based scores that quantify temporal separation between outcome groups. We evaluate the proposed framework on real-world clinical data from mortality prediction after allogeneic hematopoietic stem cell transplantation. Results show that the proposed scores capture temporally consistent separation patterns that complement conventional evaluation metrics. Overall, the proposed framework provides a novel perspective for evaluating time-dependent risk estimates in longitudinal clinical settings.
Join order (JO) is a fundamental task of query optimizers, as it significantly influences query response time. Determining the optimal JO is an NP-hard problem. Several cost-based algorithms have been proposed. The rise of AI has led to learned-based JO techniques, but their limitations with dynamic workloads have prompted the development of hybrid approaches that combine the strengths of both cost-based and learned methods. This evolution has prompted researchers to investigate the effectiveness of learned solutions compared to cost-based ones, from only a response time perspective. Meanwhile, the database research community is increasingly focusing on energy efficiency (EE) as a critical aspect of system design and optimization. In light of this trend, we argue that a comparative analysis of cost-based JO (CBJO), learned-based JO (LBJO), and hybrid JO from an EE perspective is essential. In this paper, we first present the ecosystem of JO selection, highlighting its major components. Next, we provide an end-to-end experimental study of AlphaJoin (a representative LJO), HybridQO, and PostgreSQL, using the STACK and JOB datasets to evaluate both query response time and energy consumption. We conclude by emphasizing the main findings: LJO methods often trade fast query response for higher energy use and exhibit less maturity and stability than CBJO approaches. HybridQO demonstrates adaptability by switching to CBJO when needed, thus balancing efficiency and robustness across varying workloads.
This paper presents a novel approach to automatic classification of tattoo descriptions extracted from historical prison documents and investigates their utility in record linkage. We utilize transformer-based models to generate both binary and multi-label classification systems for these unique textual descriptions. The generated tattoo labels are then integrated into a comprehensive record linkage workflow, serving as an additional feature alongside existing attributes. We evaluate this integration within both a classical machine learning framework (random forests) and a deep learning record linkage system (Ditto), and find that tattoo labels can contribute positively to record linkage performance in certain configurations. In addition, tattoo labels show potential as a supplement to traditional blocking keys, offering a path toward more efficient record linkage without sacrificing pair quality and pair completeness. This research demonstrates how automatically classified textual data can be incorporated into historical record linkage pipelines, and highlights the broader potential of leveraging unconventional descriptive attributes in this domain.
Entity Resolution (ER) is the task of identifying records that refer to the same real-world entity. A fundamental challenge in data lakes, where heterogeneous datasets lack shared identifiers, producing duplicates and inconsistencies that undermine analytics and data governance. Existing ER approaches treat attribute contributions implicitly, offering no transparent or auditable reasoning over attribute-level evidence. We revisit ER through the lens of Kolmogorov-Arnold Networks (KAN) and propose KAN-ER, a framework operating on attribute-level similarity vectors that learns nonlinear matching decisions. Experiments on twelve datasets show that KAN-ER outperforms DeepMatcher on structured data and achieves competitive F1-scores with DITTO. KAN’s spline functions provide interpretable per-attribute insights critical for data governance. To the best of our knowledge, this is the first application of KAN to the ER task.
We propose a GPU-accelerated conflict analysis method that extracts dependencies from concurrency control information, constructs the dependency graph on GPU memory, and detects cycles via strongly connected component (SCC) analysis. This prevents read-only anomalies in multi-version databases, where a read-only transaction observes an inconsistent snapshot due to dependency cycles. Existing approaches such as Serializable Snapshot Isolation (SSI) avoid exact cycle detection by conservatively aborting transactions upon detecting dangerous structures, which can unnecessarily reduce throughput. In contrast, Read-Safe Snapshot (RSS) guarantees serializability for read-only transactions without aborts by identifying safe snapshot boundaries from exact dependency analysis. However, exact cycle detection is impractical at scale as graph construction and analysis are expensive on CPUs. Our method offloads the entire pipeline—dependency extraction, graph construction, and SCC-based cycle detection—to the GPU. Results show that our method achieves an order-of-magnitude speedup compared to CPU implementations, reducing processing time to milliseconds and enabling near-real-time dependency analysis.
Mapping structured data to large ontologies such as CIDOC CRM remains a challenging task for Large Language Models (LLMs). Previous studies have shown that providing the LLM with a small, domain-relevant subset of the ontology rather than the full specification significantly improves annotation quality. However, constructing this subset has so far required manual expert work, limiting the Automation of semantic integration approaches. In this paper, we present ASOS-CRM (Automated Semantic Ontological Scoping for CRM), a two-stage approach that automatically selects the relevant CIDOC CRM classes and properties for a given dataset. ASOS-CRM combines multilingual sentence embeddings with formal OWL constraints checking to select logically valid ontological classes and properties. Evaluated on four real-world archaeological datasets, our approach yields annotation results that match and marginally surpass those of a manually curated subset, while requiring significantly less preparation time.
Reinforcement learning has become a popular approach for automating physical database design decisions such as which structures to create, retain, or remove in response to dynamic workloads. However, every existing method learns a single expected return for each candidate action and acts on it as if it were a reliable summary of the future. This is a fragile assumption in hybrid transactional/analytical processing (HTAP) environments, where the same action can yield very different long-horizon returns depending on the workload phase, and where bad decisions on write-heavy tables impose persistent maintenance penalties. We propose a different approach based on distributional reinforcement learning. Instead of learning one number per action, our agent learns the full distribution of possible long-horizon returns. This richer representation reveals which actions carry heavy downside tails. Experiments on the CH-Benchmark show that this method is strongest against the matched baselines and performs favorably against prior advisors under the HTAP objective.
Temporal graph learning has emerged as a key paradigm for modeling dynamic interactions in evolving networks. While recent methods have achieved remarkable predictive performance, their computational and energy costs have received comparatively little attention, despite their growing impact on the scalability and sustainability of graph analytics. This paper introduces a multi-objective evaluation framework for assessing temporal graph learning algorithms beyond predictive accuracy. The proposed framework jointly analyzes predictive performance, execution time, and energy consumption, enabling a principled characterization of efficiency-performance trade-offs across diverse graph topologies, temporal dynamics, and hardware settings. Through an extensive empirical study, we show that substantial reductions in computational cost can be achieved without compromising analytical effectiveness. Our results indicate that selecting algorithms under physical-efficiency constraints can reduce execution time by up to 98
Clustering is an important task in data analysis, revealing implicit, previously unknown insights from data. HDBSCAN is a widely used hierarchical clustering algorithm due to its ability to handle diverse data distributions, but like all clustering algorithms its performance degrades in high dimensional spaces. To address this limitation, we propose in this paper a novel algorithm called, ANNOYing HDBSCAN. It is a theoretically and practically more efficient iteration on HDBSCAN that uses approximation, with minimal loss in accuracy.
Results of aggregate queries can be explained by the predicates (attribute-value pairs) that most influence the outcome. Coalition-based game-theoretic power indices (e.g., Shapley value) provide principled attribution but ignore causal structure among attributes and require exponential computation over subsets, which is impractical at scale. Causal methods address structure but assume a complete causal graph, often unavailable in practice. We propose Causally-Constrained Power Indices (CCPI), a framework that integrates partial causal knowledge into coalition-based power-index attribution. Our core idea is to restrict attribution to causally valid subsets consistent with a partial directed acyclic graph that captures the partial knowledge available. We evaluate CCPI, instantiated using three different power indices on synthetic and three real-world datasets. Results show that we achieve speedups from 4× to over 1000× depending on DAG density over unconstrained baselines, supporting practical use at scale. Furthermore, CCPI yields more causally faithful explanations than non-causal baselines.
Deep learning models have achieved strong performance in sensor-based human activity recognition (HAR), but their high computational cost limits deployment on resource-constrained edge devices. To address this issue, we propose ConvTransAKD, a hybrid CNN and Transformer knowledge distillation framework with multi-level Wasserstein Distance alignment. The proposed framework transfers knowledge from a high-capacity teacher to a lightweight student through feature- and logit-level alignment, enabling effective transfer of both intermediate representations and class-wise structure. Experiments on the UCI-HAR and PAMAP2 datasets show that ConvTransAKD consistently outperforms state-of-the-art methods, achieving 98.29
In the continuous process industry, Digital Twins (DTs) are expected to perform diverse predictive tasks, which commonly results in the deployment of isolated, task-specific models. The Time-Series Foundation Model (TSFM) paradigm offers a unified alternative; however, current architectures are not well-suited for the process industry as they fail to utilize the physical topologies governing these systems. Channel-Independent (CI) models isolate variates, while Channel-Dependent (CD) models capture cross-variate dynamics but assume a fully-connected topology. Spatio-Temporal Graph Neural Networks (STGNNs) enforce structural constraints but primarily target supervised training on fixed topologies. To bridge this gap, we introduce a pre-training corpus with directed topologies and propose the Probabilistic Causal Foundation Model (PCFM). PCFM leverages explicit structural priors for topologically constrained predictions, scaling linearly with edge count. Pre-trained via masked sequence reconstruction, PCFM supports zero-shot forecasting, imputation, and structural what-if analysis. Evaluated on the 97-task GIFT-Eval benchmark, PCFM achieves competitive predictive performance against state-of-the-art foundation models.
Efficient data ingestion remains a critical requirement for enabling real-time analytics in modern data ecosystems. While existing tools such as Apache Kafka, Apache Spark, Apache NiFi, and Delta Lake have made significant strides in batch and streaming data processing, they typically operate in isolation and require extensive manual configuration to support heterogeneous, schema-evolving data sources. In this paper, we introduce SmartIngest, a unified and adaptive data ingestion framework purpose-built for lakehouse architectures. Unlike conventional ingestion pipelines, SmartIngest integrates schema adaptation, metadata-driven transformation, and format-aware processing into a cohesive and automated system. SmartIngest abstracts away the complexity of managing disparate data formats and structural changes, enabling ingestion with minimal human intervention. Our framework incorporates real-time monitoring and optimization strategies to sustain low latency and high throughput across diverse workloads. Experimental evaluation across multiple data types and ingestion scenarios shows that SmartIngest achieves substantial improvements in ingestion speed, data quality, and scalability.
AI technology is criticized due to harmful incidents caused by its computational ad-hoc and obscure nature, which largely neglects human principles. In this work, we view agents as fairness guardians and set the proper guardrails in AI pipelines, by a fairness-by-design approach. Multi-role agents are proposed to an end-to-end (human to AI) synergetic framework, so called FAIRTOPIA, governed by an effective algorithm adaptive to each AI decision making scenario contexts. FAIRTOPIA is modeled by a three-layered architecture, to encapsulate the AI pipeline inside an agentic guardian and a self-refining scheme. We enact fairness watch in all of the AI pipeline stages, under robust multi-agent FAIRTOPIA workflows which safeguard human-centric, systematic, and socio-technical principles.
Detecting anomalies in multivariate time series (MTS) remains a challenge for large-scale systems. Due to annotated data scarcity, models typically rely on unsupervised learning to capture nominal dynamics. Recently, Transformer approaches have demonstrated their ability to model complex spatio-temporal correlations. However, their quadratic computational cost necessitates data approximation, resulting in the loss of fine details for accurately detecting anomalies. Furthermore, over-parameterised Transformer autoencoders often boil down to a linear mapping, memorising inputs rather than compressing representations, leading to blind reconstruction of anomalies. To overcome these limitations, we propose a new transformer-based method called the Coordinate-Driven Autoencoder (CODA). CODA bypasses the attention bottleneck through ‘temporal collapse’, an asymmetric cross-attention mechanism that transforms temporal patches into physical variable representations, reducing overhead. To mitigate identity mapping, we recreate signals from a ‘blank canvas’ of spatio-temporal coordinates, creating a bottleneck that restricts trivial copying. Evaluating CODA on 200 datasets from the TSBAD benchmark shows it improves upon state-of-the-art approaches in VUS-PR while achieving statistical equivalence on F1 score and AUC-ROC. Furthermore, unlike standard attention methods which incur prohibitive costs, CODA requires minimal computational resources, enabling highly scalable deployment.
Large Language Models (LLMs) are increasingly adopted to support user interaction with complex data management systems. While recent research has explored their use in tasks such as multidimensional design and translation of analytical queries from natural language into MDX, little attention has been devoted to assisting users in querying data warehouse (DW) metadata. In this paper, we investigate the use of LLMs to answer natural language queries on a DW schema. To this end, we rely on a knowledge graph (KG) that represents DW metadata according to the Dimensional Fact Model. We define a benchmark of queries organized into four categories and evaluate the ability of GPT to answer them, comparing them to the baseline capabilities offered by a CASE (Computer-Aided Software Engineering) tool for DW design. We also assess the benefits of enriching the KG with natural language definitions and descriptions. The results show that LLMs can significantly improve the accessibility of DW metadata, while the accuracy varies depending on the query type and the level of semantic enrichment provided.
Data lakes rely on metadata to remain usable, yet this metadata is often limited or weakly informative for column relationship discovery, especially in ERP-derived datasets with coded or abbreviated schema labels. We propose ColRel, a two-stage method that builds column embeddings from metadata and data available at ingestion time. In difficult cases, such as coded schemata, business dictionaries help interpret column names and support the generation of short natural-language descriptions used in the second stage. Experiments on public benchmarks and an industrial ERP dataset show that ColRel is particularly effective in semantically related, weak-signal settings.
Data lakes stores raw data in their original format. These data can be structured (e.g., tables), semi-structured (e.g., JSON, XML), or unstructured (e.g., plain text). Yet, without proper metadata management, data lakes can quickly become inaccessible and unusable. Table annotation is therefore essential to make the semantics of tabular data explicit. In this paper, we introduce a hybrid annotation approach that jointly leverages language models (LMs) and knowledge bases (KBs) to automatically capture table semantics (column types and column relationships). Unlike existing works, our approach combines both paradigms while fully exploiting their individual strengths. It reconciles heterogeneous annotations and exploits type hierarchies to prioritize fine-grained semantics to produce richer and more reliable annotations. These enriched annotations are represented as annotation graphs that uniformly encode both column-level types and inter-column relations. Our experiments show that our approach achieves competitive performance on standard benchmarks and generates more expressive annotations.
Knowledge Tracing (KT) uses students’ interaction history on the platform to assess students’ mastery of knowledge at the current moment, but often requires experts to label skills. Therefore, we propose a Q-matrix-based Graph Attention Knowledge Tracing to achieve self-labelling of skills. The model uses the graph attention neural network to aggregate the information of past interactions, and realizes the accurate modeling of students’ knowledge state. In addition, it can learn the Q-matrix without expert labeling, enabling self-labelling of skills. Experiments on ASSIST2009 show our model improves AUC by nearly 3.6
