This paper presents the scientific foundation of the ADBIS 2025 tutorial on graph analytics for estimating reciprocity in intellectual communities using property graphs. By integrating Natural Language Processing (NLP), graph analytics, and embeddings, we analyse culturally rich datasets—such as Wikipedia entries and literary corpora from Latin America and Eastern Europe—to uncover epistemic violence, highlight marginalized intellectual contributions, and advance decolonial approaches to information systems.
This paper summarises the fundamental background of the Doctoral Consortium titled “Ethical and Equitable Data Science”. It introduces Freda, a methodology for designing ethical, frugal, and equitable data and algorithm-driven science. It bridges technical innovation with social justice by integrating data sovereignty, fairness-aware analytics, and community-in-the-loop infrastructure. Rooted in decolonial and feminist perspectives, Freda addresses transparency, accountability, and epistemic diversity through policy-aware Spark pipelines, federated learning, and negotiated resource dispatching. A case study illustrates how sovereignty-aware pipelines enable community control, minimize extractivism, and embed plural, justice-centered values into AI systems.
Interactive time series visualization is essential in domains like IoT monitoring but is often constrained by latency and scalability challenges. Various methods have been proposed to address these issues, each with different trade-offs between efficiency, interactivity, and visualization accuracy, making systematic evaluation crucial. Traditional benchmarking approaches, however, fail to capture user-perceived responsiveness and accuracy in real-world exploration scenarios. To bridge this gap, we introduce TimeVizBench, an interactive evaluation platform for scalable time series visualization methods across performance and accuracy dimensions. TimeVizBench enables users to configure different methods, explore visual outputs interactively, and dynamically assess performance and accuracy. It also provides a standardized interface for integrating and comparing additional methods.
Retrieval-augmented generation (RAG) improves the performance of LLM-based applications by incorporating information from external knowledge bases. However, the introduction of search engines and knowledge sources can also introduce new biases and stereotypes into the system. Previous studies have shown that adjusting the bias of retrievers through fine-tuning can influence the overall bias of the RAG system, mitigating bias in RAG. In this work, we propose a re-ranking-based method, termed ReFaRAG, as an alternative to fine-tuning for controlling the bias in retrieval results. We further investigate how biased retrieval output affects different LLMs within the RAG framework.
This paper presents an approach to accelerate the Meta-blocking phase in entity resolution (ER) by leveraging GPU computational power. We enhance the performance of conventional meta-blocking algorithms by utilizing sparse matrix representations of block collections. Our proposed solution remains orthogonal to existing blocking and matching techniques, ensuring that their effectiveness is not compromised. By converting a standard block collection to a one-hot encoded sparse matrix and implementing block purging, block filtering, and edge pruning on GPUs, we achieve up to 40 × speedups compared to CPU-based implementations.
Vector databases are a critical component in modern system infrastructures. In this study, we discuss the principles behind vector database management systems, with a focus on their features, the concept of vector embeddings, and similarity search mechanisms. Furthermore, we examine the synergies between vector databases and language models, which rely on vector embeddings for semantic search and retrieval-augmented generation. We also discuss the challenges arising from the integration of language models with vector databases. Through this discussion, we aim to provide early-stage researchers with an overview of the integration of vector databases and language models.
Deep Neural Network models require significant energy resources, increasing the need to balance performance with sustainability, a field known as Green AI. This paper investigates the intersection of accuracy, privacy, and energy efficiency within Deep Learning-based, Privacy-Preserving Record Linkage. Through a series of experiments, the effects of key parameters on matching efficiency and energy consumption are explored, outlining the impact of encoding, noise addition, and deep learning model configurations. The findings indicate clear trade-offs between energy consumption, privacy, and matching performance.
In the context of teaching computer science, many domainspecific languages (DSLs) used for data manipulation and transformation follow imperative paradigms, yet their semantics remain informal or tool-dependent. This paper proposes a pedagogical framework based on executable formal semantics to improve conceptual understanding and practical competence in such DSLs. Using a minimal imperative DSL developed as a teaching tool to illustrate arithmetic and data transformations, we define its syntax and semantics using denotational semantics and develop an executable interpreter directly derived from the formal rules. The framework enables students to explore and visualize the effects of each language construct, reason about program behavior, and verify correctness properties. We present a case study in which we focus on the gradual use of cross-curricular relationships and gradually build a comprehensive package for students that draws on knowledge from several courses focused on formal methods in software engineering. The paper concludes with a discussion of the potential of this methodology to bridge the gap between formal methods and practical education in the field of computer science.
Data spaces require the federation of open data from diverse providers. To support this, data catalogs play a critical role in managing and exposing metadata that enables dataset discovery from open data portals. However, most existing catalogs rely heavily on high-level metadata at the dataset level (such as title, license, and keywords) often aligned with standards like DCAT (Data Catalog Vocabulary). While useful, this coarse-grained metadata often falls short in supporting discoverability. To address this limitation, this paper proposes a novel extension to the DCAT standard specifically designed to discover relevant open data to be federated in data spaces. The extension enriches dataset descriptions with fine-grained, content-level metadata, including field level details, descriptions, and representative samples of values. These additional metadata elements provide critical context for understanding the structure and semantics of datasets, enhancing their discoverability.
Time series data is essential in many applications of high societal relevance and the problem of efficient and effective prediction is at the heart of many recent research efforts tackling domains like energy demand prediction, weather forecasting, traffic density prediction, market predictions (to name but a few). In the recent years, multiple deep learning approaches have been proposed, each presenting different architectures/models. One universally common theme is “the more, the merrier” in terms of the (training) data. In this paper, we focus on domains in which the time series data items are bound to locations – specifically, energy demand forecasting and weather forecasting, where each data source (e.g., a power transformer units or a weather station) pertains to a particular location. While many recent works have considered the incorporation of spatio-temporal dependencies in the model to improve the effectiveness – as it turns out, the inclusion of multiple sources may have adverse impacts. We present the results of two heuristics that we developed for the purpose of determining when is it that “more is less” in such settings.
Effective data discovery is crucial for collaboration and innovation in data sharing contexts. Data spaces combine the secure sharing features of data ecosystems with economic aspects of data markets in a federated environment. Traditional data catalogs, which primarily rely on high-level metadata (e.g., dataset name, license, keywords), often fail to adequately convey dataset utility to potential consumers. Our solution proposes content-based catalogs to enhance data discovery within data spaces through three key components: (i) high-quality descriptive metadata, (ii) representative data samples, and (iii) advanced discovery services. These components enable consumers to effectively find datasets that align with their requirements and evaluate their relevance prior to access. In this paper, we demonstrate through extensive experimentation across multiple contexts and data quality levels that our sampling technique significantly enhance dataset discoverability while preserving data provider sovereignty.
Managing data across diverse platforms poses significant challenges, including data duplication, vendor lock-in, and inconsistent governance. Lack of a unified table format often leads to complex pipelines, increased storage costs, and hindered interoperability. Apache Iceberg, with its platform-agnostic design, presents a solution by providing a consistent table format for large-scale analytical workloads while addressing cross-platform data accessibility. In this paper, we study the use of Apache Iceberg as a unified table format to enable interoperability between Snowflake and Databricks, with data stored on Amazon S3. Experimental setups include accessing Snowflake-managed Iceberg tables in Databricks and vice versa. Key focus areas include examining query performance, metadata synchronization, and the challenges of managing consistent data across platforms. Optimization strategies, specifically data reordering, were applied to test improvements in query performance for various workloads. The results show that Iceberg reduces the complexity of data management by automating metadata handling and synchronization, ensuring real-time data consistency. Query performance showed improvement in medium-complexity queries with optimized Iceberg tables, while highlighting potential areas for further optimization in full-table scans. These findings underscore Iceberg's potential as a scalable, efficient solution for modern data lake architectures.
Data lakes offer the flexibility to store large volumes of heterogeneous data with minimal curation. However, this flexibility comes at a cost: traditional keyword-based dataset discovery methods require reliable metadata such as table names or column headers, and become ineffective when this metadata is either missing or incomplete. This issue is especially pronounced in open or poorly maintained data lakes, where the quality of metadata cannot be guaranteed. In this paper, we present CoDD, a system for constraint-based dataset discovery in open data lakes. Instead of querying metadata (query-by-metadata), CoDD profiles datasets by extracting structured facts directly from the data using modular, user-definable components. Users can perform query-by-constraint searches by specifying constraints over the profiled facts in an interactive, question-driven interface. Early results from our user study show that CoDD enables users to find relevant datasets when traditional keywordbased approaches fail due to insufficient or misleading metadata. Furthermore, CoDD performs comparably well even when accurate metadata is available, demonstrating that query-by-constraint is a robust and scalable alternative for dataset discovery in open data lake environments.
Navigation in unfamiliar cities is conveniently done using an appropriate mobile app. Such apps must account for different methods of transportation, especially when the use of a car is not a viable option, and public transportation in combination with bicycling or a shared e-scooter is used instead. Existing multimodal route planners developed for this task treat response times as first-class citizens. As a consequence, they often fall short in terms of result quality by potentially missing optimal routes due to restrictive assumptions or insufficient exploration of the numerous nearby public transport access points. This paper presents RideLink, a database-centric routing system that deliberately reverses the usual priorities of these optimization goals such that route quality, i.e. the compound travel time, becomes the primary objective. It evaluates all stops reachable within an empirically chosen cycling/scooter radius (5 km) before pruning alternative routes, thereby discovering near-optimal multimodal journeys, which are often multiple minutes faster than those of other widely used approaches, while still responding within a few seconds on commodity hardware. A fully deployed system demonstrates real-world viability and underpins the contribution to smart urban tourism.
This paper aims to address the complex challenges associated with managing knowledge in the context of financial market data analytics. It presents a novel software architecture that makes use of a knowledge graph for building and executing analytics processes. It describes how the different architectural components interact with each other to fulfil the needs of the user(s). The proposed solution is validated by developing a software prototype and testing it on a scenario involving the computing of a financial market measure. This approach underscores the potential of integrating knowledge management practices into analytical environments, especially in handling the complexity and variability of financial market data analytics processes.
Sharing of data between researchers can benefit data owners, data users, and overall the scientific community. However, sharing presents challenges such as data owners losing control of their data. In this paper we explore a blockchain-based tool for managed data sharing. We present a concrete demonstration where data owners can control access to their data. We also discuss the possibilities and challenges of such an approach. Possibilities include flexible data storage and data access tracing without relying on a trusted third party. Challenges include identity verification, ease of use and blockchain transaction costs.
This research implements the application of Large Language Models (LLMs) in the summarization and distinguish analysis of multiple entities within Resource Description Framework (RDF) graphs. As the volume of structured data on the web is growing exponentially, the need for efficient and effective methods to interpret and summarize this data becomes increasingly important. This study focuses on utilizing LLMs to generate human-readable summaries from RDF graphs and particularly emphasizing on distinguishing between multiple entities. The study apply SPARQL queries to extract relevant data from DBpedia, subsequently a thorough process of frequency analysis and property unification to refine the dataset. Three LLMs including ChatGPT, DeepSeek, and Mistral have been evaluated for their ability to generate coherent and informative summaries. The evaluation process combines human-based assessments with automated metrics for the thorough analysis of generated texts. Key outcomes include the effectiveness of LLMs in generating summaries that are both informative and contextually relevant. The research also reflects the importance of data preprocessing techniques, such as frequency analysis and property unification in enhancing the quality of the generated summaries. Moreover, the study provides insights into the strengths and limitations of different LLMs in summarizing RDF data that offers a foundation for future research in this area. A framework for evaluating the performance of LLMs in summarization tasks has been designed in this research opens the way for future explorations in the application of advanced AI technologies in data interpretation and knowledge representation.
Overtourism presents complex and often hidden challenges for urban environments, impacting residents, infrastructure, and visitor satisfaction. This study proposes a novel, data-driven methodology to detect and analyze latent overtourism-the early, subtle warning signs of excessive tourism-before visible breakdowns occur. By leveraging user generated content from Tripadvisor, a temporal circulation multidigraph is modeled to capture tourist mobility. Using frequent subgraph mining algorithms, the approach identifies recurring tourist movement patterns across different urban scales. These patterns are then analyzed in both spatial and temporal dimensions to detect hotspots and evaluate dynamic attractiveness through a Huff-based probabilistic model. The approach is applied to three cities of varying sizes revealing consistent tourist flows and areas under increasing pressure, suggesting early overtourism.
Modern database management systems (DBMS) face significant challenges when executing analytical tasks on exponentially growing datasets, often relying on search operation for key lookup. Traditional optimization methods focus on minimizing execution times in host-based systems. In contrast, smart storage devices enable offloading of query plan execution on-device, presenting opportunities for new optimizations. However, these devices operate under strict computational constraints and necessitate efficient resource management. Prevailing DBMS implementations predominantly employ binary search, because of its performance and robustness. In contrast, interpolation search algorithms yield considerable computational savings in smart storage settings, however they are not always robust. In this paper, we propose a novel adaptive search algorithm iSearch, which combines a configurable number of interpolation search iterations with a fallback to binary search. This hybrid approach ensures robust and predictable runtime performance, regardless of the underlying data distribution. We further demonstrate that commodity consumer devices benefit more from adaptive search approaches than traditional host systems, highlighting the potential for improved performance and efficiency in both contexts.
