WEBIST 2026 Abstracts

Area 1 - HCI in Mobile Systems and Web Interfaces

Full Papers

Paper Nr:	30
Title:	Browser-Constrained Location Logging as a Behavioral Filter for Pedestrian Mobility Analysis
Authors:	Mizuaki Izumi, Iori Sasaki, Min Lu, Tomihiro Utsumi, Ryo Sato and Masatoshi Arikawa
Abstract:	Mobile Web map applications offer zero-installation access to users of regional services but are subject to browser execution constraints that suspend geolocation updates when the application is not in the foreground. The resulting intermittent logs can be regarded as incomplete observations when compared with continuous native-app reference data. This study takes a different position: each observation is conditioned on an information-seeking action by the user, and the data can therefore be reinterpreted as foreground-conditioned data, anchored to moments of active map consultation. To investigate whether such intermittent trajectory records remain useful for pedestrian mobility analysis, a field experiment with 21 participants using a stamp-rally Web map interface was conducted. Reconstructed trajectories achieved a mean Jaccard Index of 0.89 and a mean Root Mean Square Error (RMSE) of 10.3 m, indicating that route similarity was largely preserved across reduced sampling rates and contrasting screen-viewing styles. These findings suggest that the browser execution model can function not only as a technical constraint but also as a behavioral filter for Web-based location services, selectively recording moments of user engagement.
Download

Paper Nr:	61
Title:	Influence of Mobile Interaction between Light and Dark Modes
Authors:	Paula Diaz-Alvarez, Sara Vecino, Martin Gonzalez-Rodriguez, Javier De Andrés and Daniel Fernandez-Lanvin
Abstract:	People spend increasing amounts of time with different screens, especially mobile phones, making UI design especially important, as it affects users’ experience, efficiency, and purchase decision. A vital part of UIs is colour, since it influences users’ purchase intention, perceptions, feelings, trust, visual performance, and visual discomfort. A popular theme since 2018 is dark mode, as it has become attractive due to its supposed advantages such as power saving, visual comfort, and a decrease in visual fatigue. These advantages have been studied since the 1980s to verify if they are true, but despite the considerable number of studies, no consensus has been reached. This study aims to shed light on this matter by using an eye-tracking device in the area of mobile phones and e-commerce, as the combination of these has not been studied yet.
Download

Short Papers

Paper Nr:	42
Title:	Co-Design against Production: A Dimensional Cross-Benchmark of Environmental Decision-Support Data-Driven Dashboards
Authors:	Carlo Martino, Alessio Caccamo and Mattia Francesco Ucchiello
Abstract:	Environmental data-driven dashboards mediate complex ecological data for conservation decisions, yet the design knowledge that shapes them is split between two production cultures that are rarely examined side by side: market-deployed platforms and prototypes produced through participatory co-design. This paper compares the two along a single information-design and usability framework. A corpus of twelve publicly deployed environmental dashboards operated by agencies, research institutions and firms (Corpus A) and a corpus of nine high-fidelity prototypes co-designed in a university living lab (Corpus B) are read through one shared instrument that combines the System Usability Scale, Cairo's Visualization Wheel, representational-efficiency measures, interaction principles and accessibility criteria. The comparison is deliberately dimensional rather than statistical, because the two corpora were measured in different settings. Co-designed prototypes reach a higher usability ceiling and a higher mean (76.39 against 71.08) but show far greater variance, and they diverge systematically from production along the abstraction, originality and non-human-representation axes. From this divergence we derive a set of design guidelines for environmental decision-support data-driven dashboards, distinguishing confirmations of established principles, guidelines strongly supported by the observed data, and more speculative design implications.
Download

Paper Nr:	72
Title:	Does Perceived Bad Usability Degrade Performance? A Mouse Tracking Study on User Irritation in E-Commerce
Authors:	Mujahid Mahmood, Sara Vecino, Daniel Fernandez-Lanvin, Javier de Andres and Martin Gonzalez-Rodriguez
Abstract:	Previous studies have primarily examined the impact of poor web design on immediate experiences, yet the relationship between design-induced user irritation, perceived usability, and subsequent task execution remains less clear. This study investigates whether exposure to poor web design degrades objective performance on a standardized subsequent task, or if performance is maintained despite subjective dissatisfaction. We conducted an experiment with 186 participants in a between-subject design shoe-purchasing task on an e-commerce website. The participants were shown one of two homepages: a well-designed homepage (Version A) and a poorly designed homepage (Version B) with color contrast problems, tiny lettering, pop-ups, and layout problems. The two groups were then led to an identical well-designed Cart page where 16 mouse tracking metrics were measured. System Usability Scale (SUS) was used to assess perceived usability. The results showed that the perceived bad usability did not affect objective performance: subjects exposed to the bad design condition rated the overall usability 11.52 points lower than subjects in the good design condition (59.72 vs. 71.24; t(176.38) = 4.62, p < 0.001, Cohen’s d = 0.68). However, no significant differences were observed on the Cart page for objective mouse tracking metrics. The results indicate that while initial poor design quality induces user irritation and lowers perceived usability, users maintain motor execution and task efficiency on subsequent standardized interfaces. The findings reflect the complementary nature of subjective and behavioural metrics for usability evaluation.
Download

Area 2 - Internet Technology

Full Papers

Paper Nr:	26
Title:	Broken Contracts? Towards Automated Conformance Testing for HTTP Semantics
Authors:	Matthias Keckl, Lorenz Gruber, Peter Müller, Andreas Tennert, Markus Ende, Peter Braun and Nikolas Herbst
Abstract:	The Hypertext Transfer Protocol (HTTP) serves as the implicit contract governing the semantic interactions of modern web services. Despite the prevalence of explicit interface descriptions like OpenAPI, adherence to the underlying HTTP specifications remains inconsistent. This divergence creates significant barriers for API usage, while simultaneously introducing security vulnerabilities and performance degradation. Existing conformance tools often lack comprehensive coverage, focusing narrowly on specific subsets of the protocol or omitting specific architectural roles. To address this gap, we present a novel framework designed to translate HTTP specifications into executable, architectural roles specific rules. By extracting 390 rules from the HTTP Semantics specification, we demonstrate that entire protocol specifications can be translated into a unified, executable model. By applying it to a microservice reference implementation, we show that semantic HTTP conformance testing can be effectively automated, identifying 225 runtime violations. Furthermore, we utilized the framework to analyze 3.916 publicly available OpenAPI documents, uncovering over 2.8 million violations. Notably, only approximately 1% of the validated documents were completely free of rule violations. These findings reveal a systemic gap between protocol specifications and real-world implementations, underscoring the necessity for automated conformance testing and specialized tooling throughout the API life-cycle.
Download

Paper Nr:	43
Title:	Secure IoT Event Processing at the Edge: A Stress-Test Evaluation of Cryptographic Overhead in CEP-Based Monitoring Infrastructures
Authors:	Antonello Calabrò, Eda Marchetti and Tauheed Waheed
Abstract:	Secure IoT and IIoT monitoring infrastructures increasingly process encrypted event streams at the edge, where limited computing resources, bursty traffic, and persistence operations can turn per-message security costs into system-level performance bottlenecks. This paper evaluates the performance envelope of a CEP-based secure monitoring pipeline deployed on a Raspberry Pi 5 edge node. The pipeline integrates the CONCERN runtime monitoring infrastructure with the SECURER cybersecurity testing framework and a Drools-based rule engine for Complex Event Processing. We focus on the cost of processing AES-protected IoT events inside the rule execution path, rather than at the communication boundary. The evaluation uses four rule configurations designed to isolate different overhead amplification mechanisms: single decryption, repeated decryption within one rule firing, decryption combined with multiple database writes, and decryption followed by event fanout. We measure latency, throughput, backlog growth, CPU utilization, working-memory pressure, and stabilization time under both periodic arrivals and short-burst traffic generated by an O-RAN-based IIoT testbed. Results show that encryption is not only an additive per-event cost. When embedded in CEP rules, cryptographic processing can be amplified by rule structure, persistence, and derived-event generation, leading to non-linear degradation under bursty traffic. The study provides practical design guidance for secure IoT monitoring infrastructures: avoid redundant decryption inside rules, keep persistence outside the critical processing path, and control event fanout on constrained edge nodes.
Download

Paper Nr:	56
Title:	Architectural Smells and Anti-Patterns in Modular Monolith Applications: A Multivocal Literature Review
Authors:	Maurício Carvalho, Juliana de Melo Bezerra, Karla Donato Fook and Waldemar Neto
Abstract:	Modular monoliths are increasingly adopted as an alternative to microservices for long-lived web back ends. When designed with disciplined modularity, a single deployable unit of cohesive, loosely coupled modules can offer maintainability and progressive scalability without the operational overhead of distributed systems. In practice, these module boundaries erode silently as the system evolves, and the knowledge of how this happens is scattered across grey literature. The community has consolidated peer-reviewed catalogs of architectural smells and anti-patterns for microservices, but no equivalent exists for modular monoliths. This paper addresses that gap. We conduct a multivocal literature review across six digital libraries and the grey literature. From 32 included sources, 19 primary studies and 13 grey, we propose a catalog of seventeen architectural smells in four families. Each smell is classified against the microservices catalogs as inherited, transformed when the boundary becomes internal to a single process, or novel. In the toolchains that dominate practice, an in-process boundary violation still compiles, so smells stay latent until a team tries to extract a module into a microservice. We name this the Enforcement Gap and pair every smell with a build-time architectural fitness function, giving practitioners executable checks to assess and evolve their systems.
Download

Paper Nr:	69
Title:	MTDGW: A Shared-Storage Gateway That Decouples Entry-Point Discovery from Back-End Data Reachability
Authors:	Mitsunobu Ogawa and Hiroshi Koide
Abstract:	Shared storage for web applications may expose its entry point to an untrusted process on an authenticated host. MTDGW separates discovery of the entry point from back-end disclosure by presenting a synthetic candidate namespace at the known entry and binding the real back-end namespace only after a valid path is traversed. We implement a proof-of-concept reachability-control data plane with FUSE and SMB and analyze 860 retained trials spanning static and rotating paths, multiple candidate-space sizes, depth-first search, and direct complete-path probing. A deterministic check confirmed candidate-only entry listings, non-disclosure at wrong terminals, and a minimal regular-file lifecycle through an accepted path. Larger candidate spaces sharply reduced attacker-observed back-end reach rates, broadly matching the analytical models. Within the evaluated budgets, one-second rotation provided no evidence of a lower rate. Lookup-derived strict history showed no evidence of further reducing the reach rate under depth-first search; internal counters indicated that low-level file-system events did not reliably mark logical traversal boundaries. Direct complete-path probing retained nonzero reachability consistent with the model. Under this threat model, candidate-space design mattered more than temporal rotation; the strict-history result motivates an explicit traversal identifier, which remains to be implemented and evaluated.
Download

Paper Nr:	71
Title:	Request Cancellation in Web Application Frameworks
Authors:	Tudor Alexandru Matei and Luís Ferreira Pires
Abstract:	Modern web applications must account for client requests terminating earlier than expected due to tab closure, navigation away, or connection loss. Improper termination leaves request computation redundant, since the user is no longer present to receive the response. As a result, the server sustains unnecessary resource usage or produces undesirable behaviour. This paper investigates how request cancellation in web application frameworks is influenced by their underlying architectures. This paper classifies web frameworks by cancellation signal source, which determines the route a disconnect takes to running application code, yielding three mutually exclusive architecture classes, namely passive, cooperative, and reactive. We then compare these classes using identical measurement infrastructure and idiomatic implementations across three candidate frameworks, empirically comparing the life-cycle of detection, propagation, and termination of the cancellation signal. Spring MVC does not detect the signal by specification; the Spring WebFlux is load-invariant; ASP.NET Core detection degrades under thread-pool saturation with a predictable cliff. Abort protocols determine the verifiability of work stoppage, as the reactive framework is decoupled while cooperative cancellation is itself the termination event. Finally, incomplete cancellation across all architectures follows Little’s law for resource occupancy, while state consistency depends entirely on the transaction boundary. Our results can be used by application designers to select web application frameworks according to their request cancellation requirements.
Download

Short Papers

Paper Nr:	15
Title:	An AI-Enhanced Web Platform for Dynamic Density-Plot Optimization and Interactive Analysis
Authors:	Roopal Dixit, Haiquan Chen, Nicholas Agudelo and Bang Tran
Abstract:	Density plots are an important tool for exploratory data analysis, however their utility often depends on careful tuning of parameters such as kernel type and bandwidth. In practice, these parameters are frequently selected using heuristics or manual trial-and-error, which can lead to over-smoothed distributions, spurious modes, and inconsistent interpretation across datasets. This paper presents an AI-Enhanced Dynamic Density Plot Optimization Platform that automates density plot configuration and enables interactive, quality-guided refinement. The system integrates (i) classical kernel density estimation (KDE) baselines, (ii) attention-based neural density estimation models, and (iii) reinforcement learning (RL) agents that adaptively tune bandwidth to maximize a composite visualization quality score. The platform also incorporates dynamic color adaptation to improve contrast and readability while preserving statistical fidelity. A Streamlit-based dashboard provides real-time parameter controls, multi-algorithm comparison, and exportable reports. Using MNIST as a benchmark, the proposed AI-assisted optimization improves density plot quality and reduces the burden of manual tuning, achieving higher composite quality scores than manual and library-based KDE while providing an interpretable workflow for users.
Download

Paper Nr:	41
Title:	Setting the Horizon for Front-End Performance: The Impact of Apple Silicon Generation on the Power Consumption of Web Clients
Authors:	Janne Kalliola, Vesa Hirvisalo and Juho Vepsäläinen
Abstract:	It is well acknowledged that the ICT industry will consume more energy in the future due to increasing usage of related services. A significant share of this increasing energy demand is contributed by web technologies that are widely used as they already reach the majority of people on the globe. Although it is well understood how and where this energy is consumed, there are still open questions related to specific use cases, such as web browsing, and the impact of hardware. We investigated the theme in detail as we benchmarked Apple Silicon (M1, M3, M4) against a variety of tasks to find out if there are notable differences between the generations of this relatively new processor family. We found out that newer versions of the processor yield higher performance, although this does not seem to translate to clear gains in energy consumption for web browsing. Therefore, it can be stated that while the performance of these processors improved, energy efficiency under low loads, such as web browsing, has not improved at all, posing a problem since devices using these processors are used mainly with low loads in regular usage. There are still open questions related to different types of websites and processor architectures to see if these observations generalize.
Download

Paper Nr:	44
Title:	MapUp: A Declarative Interface for the Leaflet Library
Authors:	Tizian Berger, Tobias Münch, Anis Koubaa and Andreas Schmidt
Abstract:	Interactive web maps are used in many domains, including logistics, tourism, and data journalism. The Leaflet library provides a flexible technical foundation for building such applications, but their use often requires JavaScript knowledge and an understanding of the underlying object model. Existing web mapping platforms reduce this complexity, but typically within a specific application environment. This raises the question of to what extent Leaflet-based web mapping functionality can be provided via a declarative interface without sacrificing it’s flexibility. To address this question we present MapUp, a work-in-progress approach that combines Leaflet with Web Components. MapUp provides reusable custom HTML elements for maps, tiles, markers, layer groups, and GeoJSON-based data. These components can be configured through attributes, declaratively nested, and connected via DOM events and public methods. The prototype shows how web mapping tasks can be expressed with less JavaScript while still allowing advanced users to access the Leaflet objects. We illustrate the approach through representative scenarios. The paper discusses the benefits and limitations and outlines future work on larger datasets, backend integration, and evaluation.
Download

Paper Nr:	48
Title:	Toward User-Centric IoT Cybersecurity Compliance: Challenges, Gaps, and a Methodological Framework
Authors:	Tauheed Waheed, Eda Marchetti and Antonello Calabrò
Abstract:	Cybersecurity compliance is a critical challenge for IoT manufacturers because threats, standards, and regulatory obligations evolve quickly. This paper reports a PRISMA-informed rapid review and proposes a Cybersecurity Compliance Methodology (CCM) that connects product and user information with standards recommendation, regulatory mapping, compliance procedures, and test specifications. The CCM is user-centric, traceable, and incrementally implementable; a smart-home keyless-entry example illustrates its preliminary operationalization.
Download

Paper Nr:	50
Title:	Towards Trustworthy Educational Data Ecosystems: From Empirical Evidence to Governance-Oriented Architecture
Authors:	Thiago Medeiros, André Araújo, José Silva and Alenilton Silva
Abstract:	Educational institutions increasingly rely on data to support decision-making, planning, and policy development. However, challenges related to data quality, interoperability, accountability, and traceability continue to limit effective educational data management. Although the literature has widely discussed Data Governance, empirical evidence of its implementation in educational environments remains limited. This study examines Data Governance in educational ecosystems through a mixed-methods approach combining a literature review, a survey, and semi-structured interviews with education professionals. The findings reveal recurring challenges, including fragmented information systems, limited interoperability, insufficient governance structures, data quality issues, and weak accountability mechanisms. Based on evidence from both the literature and empirical investigation, this work proposes a governance-oriented conceptual architecture to address these challenges through data integration, quality management, security, and Blockchain-supported traceability services. The study provides an updated synthesis of research on educational Data Governance, empirical evidence from educational practice, and an evidence-based architectural proposal to support trustworthy, transparent, and accountable educational data ecosystems.
Download

Paper Nr:	68
Title:	A User Assistant Based on a Local Native-Language Model and RAG Architecture for University Web Information Systems Controlled at the Level of Access and Context
Authors:	Mateusz Hoffman, Tomasz Wojdyński and María José Escalona Cuaresma
Abstract:	Universities increasingly rely on web information systems to support prospective, active, removed and postgraduate students, lecturers and administrative staff. Even when information is formally available, it remains distributed across regulations, announcements, schedules, grading systems and user portals, so students and other users still rely on administrative support to find the correct answer. This position paper argues that user assistants based on Retrieval-Augmented Generation (RAG) should not be designed as general chatbots connected to external language models, but as web information systems controlled at the level of access and context. By control we mean architectural management of which data are retrieved, which documents are available to a given user role, what context is passed to the model, and when the system should refuse to answer. The position is illustrated by AIKIDO, a prototype local user assistant that combines the Bielik 11B native-language model, a single ChromaDB vector database with metadata-based access filtering, a backend orchestration layer, a demonstration API simulating a user information system and a web-based chat interface. Role-dependent access is implemented not by separate RAG databases, but by filtering retrieval results before context construction.
Download

Paper Nr:	39
Title:	Early-Phase Cloud Infrastructure Cost Estimation Using Model-Driven Development
Authors:	Mario Matías Urbieta, Martin Krahl Vitale and Sergio Firmenich
Abstract:	The selection of an appropriate software architecture is a critical decision in the development of any software system. With the advent of cloud computing, architects now have a plethora of architectural styles to choose from, each with its own set of trade-offs. In this work, we present a development approach based on MDD to the early evaluation of software architectures from a cost point of view. To characterize the application usage demand, we conducted two surveys: one to determine the adoption of serverless computing, and another to identify typical usage patterns. Based on the survey results, we defined several usage scenarios and generated application prototypes using JHipster with both serverless and monolithic architectures. These prototypes can be used to assess the cost of app’s infrastructure according to the usage scenario. We present a use case showing the benefits of the approach. Our findings provide insights into the cost-effectiveness of each architecture under various conditions, aiding software architects in making informed decisions.
Download

Paper Nr:	78
Title:	Implementation of Hadoop Clusters Using Containers: A Systematic Review
Authors:	João Fernandes Santos Filho, José Aprigio Carneiro Neto and Edward David Moreno
Abstract:	Technological evolution has driven the development of computational architectures with high processing capacity to enable the execution of more complex applications. One of the solutions that offers a high level of performance is the use of clusters, which are formed by a set of computers that execute tasks in a distributed manner. Parallelism libraries are used to manage the system and coordinate communication, allowing tasks to be executed in parallel. Apache Hadoop is one of the main libraries for processing and storing large volumes of data, but it has load balancing issues when processing small datasets. A strategy that can contribute to optimizing balanced task distribution and process management is the use of containers. In this context, this work presents a systematic review of studies related to the implementation of container-based Hadoop clusters. Searches were conducted in the Scopus, ACM, Web of Science, IEEE Xplore, and ScienceDirect databases, resulting in the selection of 12 works after applying the established criteria. The analysis of the studies allowed for the identification of relevant information on the topic, such as containerization technologies, types of applications executed, monitoring tools, and different application scenarios, contributing to the development of new research in this area.
Download

Paper Nr:	79
Title:	Dynamic Pair-Wise Confusion-Aware Loss for Imbalanced Intrusion Detection
Authors:	Xiaoying Li, Xiaojuan Cai and Hiroshi Koide
Abstract:	Persistent pair-wise confusion caused by feature overlap and behavioural similarity remains a major challenge for minority attack detection in imbalanced intrusion detection systems. Existing loss functions mainly allocate optimization effort according to class frequency or sample difficulty, while dynamically evolving pair-wise confusion relationships are rarely incorporated into the optimization objective. To investigate this limitation, we propose a new method named Dynamic Pair-wise Confusion-aware Loss (DPCL). DPCL dynamically estimates pair-wise confusion from the confusion matrix during training and constructs adaptive pair-wise weights that are integrated into the loss function to emphasize persistently confusing class pairs. Experiments on the CICIoT2023 dataset compare DPCL with Cross-Entropy, Focal Loss, and Class-Balanced Loss under identical settings. The results show that dynamically modelling pair-wise confusion provides a new optimization strategy for reducing persistent inter-class confusion while maintaining competitive overall classification performance.
Download

Area 3 - Social Network Analytics

Short Papers

Paper Nr:	53
Title:	Beyond Third-Party Cookies: Measuring Meta Tracking Technologies on German Websites
Authors:	Steven Brumm and Konstantin Knorr
Abstract:	Tracking technologies remain a key component of the online advertising ecosystem despite increasing regulatory pressure and browser-level privacy protections. Meta’s tracking infrastructure combines client-side scripts, first-party identifiers, and URL-based mechanisms to enable cross-site user profiling. While previous studies have investigated Facebook tracking on a global scale, empirical evidence on the German web remains limited. In this paper, we present a large-scale measurement study of Meta-related tracking technologies across German websites and university domains. We developed MetaTrackInspector, an automated framework based on Scrapy and Playwright that detects Meta Pixel deployments, consent interactions, first-party tracking identifiers, and URL-based identifier propagation. Using a dataset comprising approximately 20,000 German domains and a comprehensive collection of German university websites, we analyzed tracking behavior before and after user consent. Our results show that Meta-related tracking is widespread, affecting up to one third of the analyzed websites in average. We further identify pre-consent tracking on up to 6% of websites, indicating that tracking is initiated before explicit user consent is obtained. Moreover, tracking prevalence differs substantially among universities depending on their sponsorship, with private institutions exhibiting significantly higher adoption rates than public and church-affiliated universities. Finally, Meta Click Identifiers are automatically converted into persistent first-party cookies on more than 87% of websites that employ Meta tracking. These findings demonstrate that Meta’s tracking infrastructure continues to operate at scale despite regulatory requirements and highlight the increasing importance of first-party identifier-based tracking in the modern web ecosystem.
Download

Paper Nr:	52
Title:	Towards a Workflow for Author Impact Analysis via Co-Citation Networks and Early Insights from Extracted Author Data
Authors:	Vincenza Carchiolo and Michele Malgeri
Abstract:	Co-Citation analysis is a foundational technique in bibliometrics, yet its application to author-level scientific influence requires reproducible and scalable methodological frameworks. This paper presents a workflow for acquiring, preprocessing, and temporally aligning bibliographic data from OpenAlex and comparable sources. To demonstrate its applicability, the workflow is used to construct longitudinal timelines describing the evolution of scholarly output and citation impact during the first five years of a researcher’s career. This study constitutes the first step of a broader research programme on author-level co-citation behaviour from temporal and, in future work, structural perspectives. The resulting timelines provide an empirical baseline for analysing citation and referencing dynamics. While future developments will incorporate co-citation networks, structural indicators, and higher-order relational analyses, the present contribution focuses exclusively on the temporal dimension.
Download

Area 4 - Web Intelligence and Semantic Web

Full Papers

Paper Nr:	23
Title:	Risk Vectors in Large Language Models (LLMs)
Authors:	Renato Fernandes Pereira, Admilson de Ribamar Lima Ribeiro and Gilton José Ferreira da Silva
Abstract:	The increasing integration of Large Language Models (LLMs) into various applications has raised critical security concerns. This study aims to identify and categorize the primary risk vectors associated with LLMs reported in the literature, consolidating current knowledge to guide future research and the development of more secure models. From a researcher’s perspective, the study characterizes existing technical vulnerabilities, explicitly excluding from its scope topics such as data privacy and legislation, social engineering, threat intelligence, or generic risk management. To achieve this objective, a Systematic Literature Mapping (SLM) was conducted based on the guidelines by Kitchenham and Charters (2007). This methodological approach provides a comprehensive overview of the state of the art, allowing for the mapping of the volume of available evidence and topics of interest in LLM security without the need for an exhaustive, in-depth analysis. The research corpus consisted of 25 papers extracted from recognized scientific databases, including the ACM Digital Library, Compendex, IEEE Xplore, ScienceDirect, Scopus, and SpringerLink. As a result, the study identifies the most common risk vectors in LLMs.
Download

Paper Nr:	27
Title:	An Ontology for Animal Biosecurity Risk Assessment: An Extension of the Common Ontology of Value and Risk
Authors:	Jéssica de Almeida Berlezi, Alencar Machado, José Palazzo M. de Oliveira and Vinicius Maran
Abstract:	Recent outbreaks of highly pathogenic avian influenza (HPAI) in Brazil and the United States have revealed limitations in animal health governance, highlighting the need for improved computational and decision-support tools. This work presents an initial effort to address this gap through the development of a domain ontology grounded in sanitary regulations and complemented by international frameworks, including the World Organization for Animal Health (WOAH) and Food and Agriculture Organization (FAO) guidelines. The ontology was developed using the Unified Process for ONtology building (UPON), specified in OntoUML, and grounded in the Unified Foundational Ontology (UFO). Beyond describing domain concepts, semantic validation and inference were used through automated reasoning with the HermiT reasoner and SWRL rules in Protégé. For validation purposes, synthetic datasets of simulated epidemiological scenarios were used, such as the impact of pathogen exposure on a batch of animals, enabling the ontology to demonstrate its inferential capacity in realistic risk assessment contexts. This provides a domain foundation for the development of surveillance tools to be integrated into existing animal health defense and disease control information systems.
Download

Paper Nr:	32
Title:	ECHO-G: Evidence-Centered Hybrid Ontologization Generic with Semantic Idempotence Evaluation Using Local Mixture-of-Experts Language Models
Authors:	Claire Ponciano, Markus Schaffert and Jean-Jacques Ponciano
Abstract:	Large language models have made the automatic construction of knowledge graphs from unstructured documents practical, yet the resulting graphs are hard to audit: extracted statements rarely carry traceable evidence, and no standard criterion exists to quantify whether the graph faithfully represents the source. We present ECHO-G, an evidence-centered ontology-construction framework that addresses both gaps. ECHOG uplifts documents into a TBox-constrained RDF graph in which every assertion is materialised as a named echo:ExtractedFact node carrying source span, confidence, certainty, polarity, validation status, and PROVO-aligned provenance. The pipeline runs entirely on local hardware through a Mixture-of-Experts openweights model. To evaluate faithfulness we introduce semantic idempotence: the document, after the round trip PDF → MD → TTL → MD′ , should preserve every fact and invent none. We operationalise the criterion through LLM-based atomic fact extraction with embedding-based fuzzy matching, yielding recall, precision, and hallucination-rate metrics on two reconstruction channels: the text the graph retains, which bounds the noise of the measurement itself, and the reified assertions alone, which measure extraction fidelity. On a multilingual benchmark of three Wikipedia articles on Orthohantavirus (English, German, French), ECHO-G retains 0.951 of the source facts at a hallucination rate of 0.050, and restates 0.481 of them from its assertions alone. A qualitative analysis shows the dominant failure mode is not fabrication but language drift inside the evaluation instrument. We release the framework, ontology, benchmark, and audit trails under an open-source licence.
Download

Paper Nr:	37
Title:	Automated Assessment of Children’s Writing Competence Using Large Language Models
Authors:	Alejandro Alvarez-Varela, Haithem Afli, Marisol Cueli, Daniel Fernandez-Lanvin, Martin Gonzalez-Martinez and Javier De Andres
Abstract:	This study examines the validity of DeepSeek-R1 for the automated assessment of children's writing competence in Spanish. We analyzed compare-contrast texts produced by 126 4th-grade students from three primary schools in Spain across three analytic dimensions: Structure, Coherence, and Quality. The evaluation spanned five distinct measurement phases. Automated scores generated by the Large Language Model were compared against human expert consensus using the Intraclass Correlation Coefficient model (3,1), Spearman rho, Mean Absolute Error percentage, and Wilcoxon signed-rank tests with Holm-Bonferroni correction. Intraclass Correlation Coefficient values were uniformly poor across all 15 phase-by-dimension combinations (range: 0.128–0.443). This performance was driven by a systematic negative bias that emerged from the first test onwards and peaked at the second test, with an underscoring of -17 to -25 points per dimension. Conversely, Spearman rho values improved substantially over time, rising from 0.23–0.31 at Pretest to 0.45– 0.63 at Posttest and Follow-Up. This dissociation between rank ordering and absolute calibration indicates that the current pipeline cannot follow students' score gains during instruction. Therefore, the system is unsuitable for longitudinal growth tracking without phase-specific recalibration. Baseline screening remains the most viable near-term application.
Download

Paper Nr:	47
Title:	On the Future of Software Reuse in the Era of AI Native Software Engineering
Authors:	Antero Taivalsaari, Tommi Mikkonen and Cesare Pautasso
Abstract:	Software development is currently under a paradigm shift in which artificial intelligence and generative software reuse are taking the center stage in software creation. Earlier opportunistic software reuse practices and organic software development methods are rapidly being replaced by ”AI Native” approaches in which developers place their trust on code that has been generated by artificial intelligence. This is leading to a new form of software reuse that is conceptually not all that different from cargo cult development. In this paper we discuss the implications of AI-assisted generative software reuse, bring forth relevant questions, and define a research agenda for tackling the central issues associated with this emerging approach, followed by some predictions on the future of open source and software reuse more broadly.
Download

Short Papers

Paper Nr:	28
Title:	Contextualized Obfuscation: A Privacy-Preserving Architecture for Safe LLM Integration in Judicial Text Processing
Authors:	Rafael Marconi Ramos, Ézio Paes Pacheco, João Victor Rodrigues Viana Omati, Oliver Henrique Ferreira de Jesus, Pedro Carvalho Brom and Li Weigang
Abstract:	The adoption of Large Language Models (LLMs) in judicial environments is constrained by strict privacy requirements and the risk of exposing sensitive information under regulations such as the Brazilian General Data Protection Law (LGPD). This work presents a privacy-preserving adaptation pipeline for judicial document processing based on contextualized obfuscation, in which sensitive entities are replaced by semantically consistent placeholders that preserve referential integrity and contextual coherence. The proposed architecture enforces a fully local preprocessing layer, ensuring that no original sensitive data is transmitted to external LLM services. The pipeline combines LLM-based entity extraction, rule-based enhancement, and post-processing techniques to improve entity normalization and replacement consistency. Experimental evaluation using real-world judicial documents and nine different LLMs, including both local and commercial models, demonstrated that contextualized obfuscation consistently outperformed traditional anonymization approaches in downstream interpretation tasks. Across all evaluated commercial LLM judges, the proposed approach achieved substantial gains in contextual understanding, with average improvements of around 3× and gains reaching up to 17× in some scenarios, while consistently preserving actor traceability and semantic relationships more effectively than traditional masking strategies. Results also indicate that the proposed pipeline remains computationally feasible in resource-constrained environments, supporting practical and scalable privacy-preserving integration of LLMs into judicial workflows.
Download

Paper Nr:	31
Title:	Enhancing Product Descriptions Using LLMs for a Better e-Commerce User Experience
Authors:	Sara Vecino, Haithem Afli, Javier de Andres, Martin Gonzalez-Rodriguez and Daniel Fernandez-Lanvin
Abstract:	With the increase in online shopping, we aim to improve the users’ experience on product details pages by using Large Language Models. Our goal is to use LLMs to generate improved product descriptions by using the already existing ones and the information of a newer feature, the questions and answers. We generated product descriptions using two different LLMs: a smaller one run in a modest environment, Llama2, and a commercial, bigger one, ChatGPT4. We found statistically significant differences in the quality of the descriptions based on the users’ ratings. ChatGPT4 descriptions are considered to be of better quality than both Llama2 and the original descriptions. We also found that participants reading the ChatGPT4 descriptions make significantly fewer first-pass regressions, that is, less rereading, compared to the Llama2 and the original descriptions. We conclude that the use of a commercial and bigger LLM can be used to enhance product descriptions, thus improving user experience on e-commerce websites.
Download

Paper Nr:	34
Title:	SmartTutor: A Course-Grounded and Policy-Aware RAG Tutoring Framework for Moodle
Authors:	Freddy Chedid, Rima Kilany Chamoun, Wadad Wazen, Mario Gharib and Anthony Bassil
Abstract:	Learning Management Systems centralize educational resources, but students still face difficulty locating reliable answers across files, slides, pages, announcements, and forums. Large Language Models offer conversational access to such material, yet unconstrained assistants introduce hallucination, privacy, and academic-integrity risks. This paper presents SmartTutor, a Moodle-integrated tutoring framework that combines closed-domain Retrieval-Augmented Generation, deterministic workflow orchestration, evidence approval, anonymization, and privacy-preserving personalization. SmartTutor binds every request to a course context, identifies course concepts, associates them with student mastery signals, retrieves only instructor-approved resources, and follows a cite-or-refuse strategy that prevents generation when evidence is weak or policy constraints fail. Before model interaction, sensitive information is redacted to reduce exposure of personal data. The system also supports multiple pedagogical modes, including Socratic, hint-based, practice, exam, revision, and teacher-defined modes, without allowing modes to bypass grounding or privacy constraints. We evaluate OCR-based ingestion, local LLM latency and usability, and governance controls within the same grounded pipeline. Results indicate that course-bounded retrieval, mastery-aware adaptation, and deterministic gating can provide explainable tutoring while maintaining institutional control over learning data. The design is also applicable beyond Moodle, to other controlled, privacy aware RAG domains.
Download

Paper Nr:	38
Title:	Multi-Horizon Temperature Forecasting with LSTM and Transformer Models
Authors:	Gerasimos Vonitsanos, Andreas Kanavos and Phivos Mylonas
Abstract:	Accurate temperature forecasting supports decision-making in agriculture, transportation, energy management, and environmental monitoring. This study compares Long Short-Term Memory (LSTM) and Transformer architectures for multi-horizon average-temperature forecasting using daily observations from a selected grid cell of the JRC MARS Meteorological Database. Meteorological and temporal predictors are incorporated through 30-day input sequences, including temperature-related variables, precipitation, vapor pressure, wind speed, seasonal encodings, lag variables, and rolling averages. Persistence, Random Forest, and Gradient Boosting models are used as reference baselines. Separate models are trained for 1-day, 3-day, and 7-day forecasting horizons and evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The Transformer achieves the lowest errors at every horizon, reducing MAE relative to Persistence by 13.4%, 14.0%, and 21.3%, respectively. Random Forest and Gradient Boosting remain competitive in several settings but do not consistently match the Transformer. SHapley Additive exPlanations (SHAP) at the 1-day horizon identify temperature-related and atmospheric variables as the most influential predictors. These findings demonstrate the effectiveness of attention-based architectures for multi-horizon temperature forecasting while providing feature-level insights into model behavior.
Download

Paper Nr:	45
Title:	CRACH-OS: A Retrieval-Augmented and Human-in-the-Loop Web Intelligence Framework for Adaptive Incident Response
Authors:	Antonello Calabrò, Eda Marchetti and Sanaz Nikghadam-Hojjati
Abstract:	This paper presents CRACH-OS, a reference architecture and operational specification for an application-oriented LLM engineering pipeline for incident response. CRACH-OS extends CEP/SIEM infrastructures with a reasoning layer that retrieves contextual evidence, constructs causal hypotheses, generates countermeasures, and exposes structured recommendations to a human analyst. The framework separates generative reasoning from deterministic decision logic: LLMs support evidence synthesis and option generation, while risk-cost models rank actions under operational, business, and policy constraints. Analysts validate assumptions, refine impacts, inspect explanations, and may override recommendations. Observed outcomes support controlled updates to detection rules, causal templates, cost models, and remediation knowledge. The framework is illustrated through a multi-stage IT/OT incident scenario involving ambiguous evidence and human-guided optimization.
Download

Paper Nr:	51
Title:	Hypermedia as a Substrate for the Agentic Web: From Documents to Affordances
Authors:	Juho Vepsäläinen and Fernando Miguel Carvalho
Abstract:	Websites and applications are increasingly used by software agents operating in behalf of their human users. The problem is that often it can be challenging for agents to understand available actions and meaning exposed by contemporary web applications. We argue that web applications intended for agentic usage should expose their valid, state-dependent actions through hypermedia controls. The aim is to allow agents to discover and execute transitions based on represented application state that is served both for human and machine users through a single interface. We connect this position to the Semantic Web, REST/HATEOAS, Hypermedia-Driven Applications, and contemporary agent interfaces to show the value of this viewpoint. The resulting synthesis establishes a research direction for evaluating agent interaction while considering maintenance related tradeoffs in hypermedia-oriented web applications.
Download

Paper Nr:	64
Title:	Platform-Aware Misogyny Detection on Social Media: A Comparative Analysis of Class Imbalance, Text Normalisation and Model Selection
Authors:	Fatima Chiroma and Alaa Mohasseb
Abstract:	Online misogyny remains a major problem across social media, yet most detection systems are developed and evaluated on a single platform, raising questions about whether successful design choices generalise across platforms. This study addresses that gap through a multi-platform analysis of misogyny detection on X (Twitter), Reddit, and a combined dataset, covering three imbalance ratios (IR = 1.07, 8.38, and 3.33). Results show that design choices are highly platform dependent and can have opposite effects. Class balancing provides minimal benefit on the near-balanced X (Twitter) dataset but improves misogyny F1 by over 18% on the highly imbalanced Reddit dataset. In contrast, synonym-based augmentation consistently reduces performance, suggesting lexical substitutions distort context-sensitive misogynistic cues across platforms. Model rankings also vary by platform. BERT performs best on individual homogeneous platforms (85% F1 on Reddit; 78.6% on X (Twitter)), while HateBERT proves more robust on the heterogeneous combined dataset (82.6–83.8% F1). Transformer models generalise better than FastText to unseen misogynistic expressions, but all models struggle with implicit, humor-driven, and stereotype-based misogyny, highlighting implicit detection as an ongoing challenge.
Download

Paper Nr:	70
Title:	Aascore: A Core Ontology for Semantic AAS Embeddings
Authors:	Daniel Spieldenner, Hacane Hechehouche and André Antakli
Abstract:	The Asset Administration Shell (AAS) is an established standard to describe assets in predominantly Industrie 4.0 contexts. Gathering information in these environments for complex tasks, for example assembly line or work plan optimization, requires information retrieval from a highly heterogeneous system, including multi domain data, such as processes, worker capabilities or industrial machines. However, query capabilities of AAS are limited, making finding and retrieving relevant data difficult. This is especially true for aggregating cross domain data from distributed data sources. While systems like Eclipse BaSyx allow for more complex search queries, including information from non AAS sources or making full use of semantic identifiers within AAS models to infer relations and properties via semantic classes is not possible. With the AAS meta model ontology available, an important step towards a semantically exploitable representation is done but lacks sufficient semantic depth and expressibility. In this paper we present a higher level AAS core ontology, allowing for linking concepts from arbitrary domains to an AAS based system, treating the AAS meta model ontology as a AAS domain ontology and allowing for extensive querying and reasoning over multiple AAS.
Download

Paper Nr:	75
Title:	A Multi-Perspective Graph-Based Framework for Heterogeneity Analysis and Variant Discovery in Object-Centric Event Logs
Authors:	Donia Sidia, Asma Mejri and Sonia Ayachi Ghannouchi
Abstract:	Process Mining (PM) has long relied on case-centric event logs, a paradigm that imposes artificial constraints on processes involving multiple interacting objects. The emergence of Object-Centric Event Logs (OCELs) overcomes these limitations by allowing a single event to relate to objects of several types simultaneously, without any lossy projection onto a single case identifier. While Object-Centric Process Mining (OCPM) has produced techniques for discovery, conformance checking, and performance analysis, the systematic comparison of process variants and the quantification of process heterogeneity remain underexplored. This paper proposes MOGHA (Multi-perspective Object-centric Graph Heterogeneity Analysis) framework, a graph-based framework that transforms OCELs into event and object graphs, identifies structural and behavioral process variants through graded similarity measures, and quantifies process heterogeneity from both perspectives through dedicated indicators.
Download

Paper Nr:	76
Title:	Ontologies as the Semantic Backbone of LLM-Driven Software Development: A Position on Grounding the AI-Assisted Lifecycle
Authors:	José Palazzo M. de Oliveira and Vinícius Maran
Abstract:	Large Language Models (LLMs) are increasingly adopted throughout the software development lifecycle. As this adoption deepens, the field is reproducing, at the scale of the whole engineering process, a common issue: the silent divergence, across successive artefacts, between what stakeholders meant, what the specification says, and what the generated code does. One of the factors that are directly related to is the semantic drift and two structural properties of unmediated generation: The absence of any guarantee of referential identity across artefacts, and the absence of a derivational record. This work proposes that a domain ontology should be the persistent semantic backbone grounding LLM activity across the software development lifecycle aid by LLMs, rather than an after-the-fact validator attached to a single phase. We make the position actionable through a taxonomy of five ontology roles with their integration points, a reference architecture, and a case study on a real animal-disease-spread ontology contrasted against a structured but ungrounded baseline. The work argue that grounding does not eliminate drift but relocates it to a single observable, human-auditable boundary, and close with a scoped research agenda.
Download

Paper Nr:	80
Title:	Context–Capacity Ontologies for Selecting Process Chunks in Loosely Coupled Cloud Workflows
Authors:	Yosra Lassoued
Abstract:	Web-based cloud workflows increasingly combine business-process fragments supplied by organizations that share neither a stable process model nor a pre-negotiated partner set. Selecting a replacement fragment here is not the same as selecting an executable web service: the consumer knows only the intended business outcome and the operating conditions of the missing fragment, while the provider keeps its internal realization private. We formulate the selection problem for process chunks (PCs) in loosely coupled inter-organizational cloud workflows. We separate context-when a fragment may be used-from capacity-the business outcome it delivers-and represent both in two linked OWL ontologies with explicit matching relations (exact, plugin, subsume). On this semantic layer we define two selection strategies: MAC, which keeps the single best-aligned chunk, and MIC, which assembles complementary chunks under governance rules, preceded by a five-stage semantic prefilter. A drug-development case study validated by pharmacology experts, and a benchmark on the public QWS dataset, evaluate the approach. Across a campaign of 2,550 runs, the prefilter cuts the candidate set by 88–97% and lowers genetic-search time while preserving admissibility. We state explicitly which claims the evidence supports and which it does not.
Download

Paper Nr:	55
Title:	A Multi-Axis Heterogeneous Deep Ensemble for Stock Price Forecasting
Authors:	Athanasios Bakalis and Christos Makris
Abstract:	Stock price forecasting is a widely studied problem in financial AI, complicated by the non-stationary and noisy nature of financial time series. Ensemble learning addresses the bias-variance tradeoff by aggregating base learners, whose error diversity is the primary driver of the resulting gain. We construct an ensemble heterogeneous along three axes (input modality, architecture family, and prediction regime) using an LSTM network, a Sequence-to-Sequence model, a Conditional Wasserstein GAN, and a Diffusion Differential Graph Transformer. We then test whether this heterogeneity yields the required error diversity. It does not. The two generative models lead on every metric, and a regression-versus-diffusion ablation attributes the graph transformer’s advantage to its diffusion head. The MLP meta learner partly suppresses the weakest model’s errors but fails to outperform the strongest single model, a deficit that error-correlation and ambiguity analyses trace to the combination stage.
Download