Machine learning enables the extraction of useful information from large, diverse datasets. However, despite many successful applications, machine learning continues to suffer from performance and transparency issues. These challenges can be partially attributed to the limited use of domain knowledge by machine learning models. This research proposes using the domain knowledge represented in conceptual models to improve the preparation of the data used to train machine learning models. We develop and demonstrate a method, called the Conceptual Modeling for Machine Learning (CMML), which is comprised of guidelines for data preparation in machine learning and based on conceptual modeling constructs and principles. To assess the impact of CMML on machine learning outcomes, we first applied it to two real-world problems to evaluate its impact on model performance. We then solicited an assessment by data scientists on the applicability of the method. These results demonstrate the value of CMML for improving machine learning outcomes.

Data centers play a central role in digital transformation, especially in the field of artificial intelligence (AI). However, their energy consumption is enormous, e.g., 16 billion kWh in Germany in 2020. At the same time, energy costs are rising and climate neutrality requirements are increasing. These factors pose major challenges for the sustainable and cost-effective operation of data centers. This paper introduces the ESCADE project (05/2023 - 04/2026), an ongoing research initiative funded by the German Federal Ministry of Economics and Climate Action, aiming to optimize the energy-efficiency of AI in data centers. AI compression techniques such as knowledge distillation, quantization and neural architecture search result in smaller, more energy-efficient AI models that deliver comparable performance. When combined with neuromorphic hardware, these models can achieve energy savings of up to 80The ESCADE consortium, a multidisciplinary collaboration of seven industry and academic partners, explores energy- efficient AI in two use cases: visual computing for scrap sorting in steel industry and natural language processing for software development. This paper provides a comprehensive overview of the ESCADE project, outlining its objectives, work packages, and anticipated outcomes. A central contribution is the introduction of first results in terms of the information system EAVE: Energy Analytics for Cost-effective and Sustainable Operations. By using AI-based analyses, EAVE optimizes the relationship between AI performance and operating costs of AI applications in data centers. The system measures and predicts the energy consumption, CO emissions and operating costs of different AI model configurations, including hardware options. At the same time, it analyzes which factors significantly influence these values. This enables decision-makers to manage the operation of data centers in a data-based and efficient manner while meeting environmental targets.

Adaptive and personalized AI systems in healthcare rely on user-specific and contextual information to provide support. However, incomplete, unreliable, and outdated data prevents both patients experiencing illness, pain, or cognitive impairment, as well as therapists, in making proper and informed decisions. Patients specifically may not have the knowledge to comprehend complex medical information, or effectively communicate symptoms. AI- driven mental models and user models can bridge these cognitive gaps, ensuring personalized and effective patient care. The FedWell research project (09/2023–08/2026), funded by the Federal Ministry of Education and Research (BMBF), explores the integration of artificial mental models (AMMs) and user models from various sources into adaptive AI systems to assist patients in decision-making. The project focuses on two key applications: rehabilitation support after knee/hip surgery and treatment decision assistance for patients with cognitive impairments (e.g., multiple sclerosis, dementia). FedWell employs a combination of structured surveys, contextual data collection, and AI techniques to model patient behavior, attitudes, and intentions. A decision support system MENTALYTICS is developed from fine-tuned large language models (LLaMA-2, LLaMA-3, Mistral, Phi-3), that employs AMMs. By the end of the project, FedWell aims to deliver robust AMMs capable of representing patient beliefs and decision-making processes, ultimately guiding them toward treatment options that best fit their individual needs.

Artificial intelligence (AI) can support patient-centered care in prevention and rehabilitation. In Germany, almost 1.9 million patients were treated in rehabilitation hospitals in 2023, mostly due to musculoskeletal disorders. The success of rehabilitation depends on cooperation between patient, doctor, and therapist as well as active participation. However, cognitive limitations, language barriers, and psychological factors tackle decision-making and communication abilities of patients. This leads to incomplete or distorted data and impairs individualized therapy. A potential solution approach is to apply artificial mental models (AMMs) that anticipate patients’ unknown mental models. These concepts are based on cognitive science theories and world models from AI. AMMs can optimize treatment decisions, correct misjudgments, and thus increase the success of rehabilitation. Particularly in knee rehabilitation, an AI agent can determine how patients perceive their recovery and enable individual adjustments. The BMFTR project FedWELL investigates the use of AMM in rehabilitation. A non-discriminatory base model was developed using data from online forums, user studies, and machine learning models. Initial results show that AI-supported models can predict individual assumptions and expectations of patients within the rehabilitation process and enable personalized therapies. This article presents the research design of the project and reports the first results of the initial survey phase.

Large Language Models (LLMs) are being rapidly adopted for many activities in organizations, business, and education. Included in their applications are capabilities to generate text, code, and models. This leads to questions about their potential role in the conceptual modeling part of information systems development. This paper reports on a panel presented at the 43rd International Conference on Conceptual Modeling where researchers discussed the current and potential role of LLMs in conceptual modeling. The panelists discussed applications and interest levels and expressed both optimism and caution in the adoption of LLMs. Suggested is a need for much continued research by the conceptual modeling community on LLM development and their role in research and teaching.

Window segmentation and vectorization remains a significant challenge, particularly in the absence of clean facade images. To extract complete window segments from building façade images with occlusions, this article proposes an occlusion-aware window segmentation (OA-WinSeg) network with conditional adversarial training guided by prior structural information. This architecture combines the power of image segmentation and generative capabilities to handle occlusions. First, OA-WinSeg automatically detects occlusions and generates a rectangular boundary guidance from a coarse window segmentation, which incorporates structural information about the building layout into the process. Subsequently, the network refines the coarse segmentation and generates window segments in the missing regions by attending to contextual information of the nonoccluded parts of the façade. Finally, our approach generates accurate vector representations, information needed for building modeling systems. Experimental results demonstrate the effectiveness of our model with simulated and occluded real-world datasets. In addition, we evaluate our model on various ablation studies to explore the contribution of the different modules. Finally, we have analyzed the potential applications of the proposed segmentation network and the completed window segments, including building façade inpainting.

Reliable routing in multimodal transport networks requires more than minimizing travel time: it demands accounting for the risk of delays and disruptions. This paper presents ROMY (Risk-Optimized Mobility), a decision support system that integrates heterogeneous data sources, advanced feature engineering, and a graph neural network (GNN)–based predictive core to deliver personalized, risk-aware route recommendations. ROMY models the transport network as a directed, attributed graph, combining spatial, temporal, modal, and semantic attributes to capture complex dependencies between network segments. The predictive core employs edgeconditioned message passing to incorporate contextual information such as travel mode, time-of-day, and navigation instructions into segment-level risk estimation. A pilot implementation demonstrates the system’s ability to identify high-risk segments, outperform baseline models, and offer alternative routes that reduce risk with minimal impact on travel time. The results highlight ROMY’s potential to enhance the reliability of multimodal mobility services and support data-driven transport planning.

Artificial intelligence (AI) is rapidly transforming software development. At the same time, approaches that align with the intellectual tradition of agile, including code-first methods, rapid prototyping, and vibe coding, have become more prominent, due to their speed and flexibility. The role of conceptual modeling has become less clear in this emerging environment. However, danger exists in discarding modeling, as it has traditionally shielded software development from risks and vulnerabilities of developing a solution without careful and systematic deliberation. These risks are now dramatically amplified due to the speed and scale of AI. This manifesto finds a new place for conceptual modeling in software development with AI, and as a result, proposes a new way to conduct software development more broadly. The manifesto comprises core principles, termed SAFE-AI, and the process of implementing the principles, MADE with AI. Together, they pave the way for a synergistic collaboration between AI and human developers with the aim of producing effective and responsible software.

Large language models (LLMs) like GPT-4 and LLaMA-3 offer transformative potential across industries, e.g., enhancing customer service, revolutionizing medical diagnostics, or identifying crises in news articles. However, deploying LLMs faces challenges such as limited training data, high computational costs, and issues with transparency and explainability. Our research focuses on distilling compact, parameter-efficient tailored language models (TLMs) from LLMs for domain-specific tasks with comparable performance. Current approaches like knowledge distillation, fine-tuning, and model parallelism address computational efficiency but lack hybrid strategies to balance efficiency, adaptability, and accuracy. We present ANON - an adaptive knowledge distillation framework integrating knowledge distillation with adapters to generate computationally efficient TLMs without relying on labeled datasets. ANON uses cross-entropy loss to transfer knowledge from the teacher's outputs and internal representations while employing adaptive prompt engineering and a progressive distillation strategy for phased knowledge transfer. We evaluated ANON's performance in the crisis domain, where accuracy is critical and labeled data is scarce. Experiments showed that ANON outperforms recent approaches of knowledge distillation, both in terms of the resulting TLM performance and in reducing the computational costs for training and maintaining accuracy compared to LLMs for domain-specific applications.

Energy storage operation strategies depend on forecasts of the expected prices in the dispatch period. The quality of the forecast influences the economic success of the dispatch strategy. Nevertheless, current literature only evaluates the fore-cast with performance metrics on the time series, not with realized revenue. In addition, dispatch strategies are mostly evaluated with perfect foresight, which is not available in real life applications, or synthetically generated price forecasts, that do not precisely reflect shortcomings of real forecasting models. To examine the influence of forecasting performance on the realized revenue and the robustness against imprecise forecasts, we examine the use case of energy arbitrage in the day-ahead energy market with a battery storage. Different forecasting models are implemented to predict the resulting clearing price of the day-ahead auction before its closure based on the historic realized price of the past days. The price forecast is then used in a mixed integer linear programming-based dispatch model for the battery storage to derive the optimal schedule for market participation. To benchmark the realized revenue, we use the actual realized price to compute the optimal dispatch decision, resulting in the theoretical maximal revenue under perfect foresight. The impact of different forecasting models is evaluated by comparing the resulting revenues. Furthermore, we study the impact of including the ageing costs of the battery in the dispatch decision-making, which affects the forecast-based dispatch. The results show significantly higher missed revenue of 84.84 % compared to the theoretical maximum revenue when ageing costs are included and only 28.27 % missed revenue when ageing costs are neglected.

In the evolving domain of prevention and rehabilitation, adaptive and personalized Artificial Intelligence (AI) systems are pivotal for delivering patient-centric care. This study investigates the application of Artificial Mental Models (AMM) within healthcare AI systems, specifically in knee rehabilitation, to address cognitive challenges that result in incomplete or biased data, impacting patient decision-making and communication. Leveraging Large Language Models (LLMs), we develop and fine-tune AMMs to accurately capture individual patients' mental models and enhance support during rehabilitation. Our research adopts a Design Science Research (DSR) methodology encompassing two phases: elicitation and individualization. In the elicitation phase, a domain-specific AMM is generated through a quantitative study involving 150 participants and indirect patient observations, ensuring a discrimination- and bias-free model. The individualization phase utilizes curated and non-curated patient data to fine-tune the AMM for individual patients. The effectiveness of these patient-specific AMMs is evaluated in real-world rehabilitation settings through A/B testing, comparing patient and AMM predictions of pain with actual pain assessments. The study's outcomes highlight the potential of AI systems to provide accurate, personalized, and bias-free patient care, significantly improving rehabilitation outcomes. The methodology and findings suggest broader applicability across various healthcare domains, enhancing the integration of AI in routine patient care and advancing the effectiveness of therapeutic interventions.

The following discussion paper summarizes the results of a panel discussion conducted on July 10, 2023 at the International Conference on Software Technologies (ICSOFT) in Rome, Italy. The panelists included Jordi Cabot from Luxembourg Institute of Science and Technology, Luxembourg, Wolfgang Maass from University of Saarland, Germany, and Marten van Sinderen from University of Twente, Netherlands. The panel was moderated by Hans-Georg Fill from University of Fribourg, Switzerland.

This paper explores the integration of graph knowledge from linguistic ontologies into multilingual Large Language Models (LLMs) using adapters to improve performance for low-resource languages (LRLs) in sentiment analysis (SA) and named entity recognition (NER). Building upon successful parameter-efficient fine-tuning techniques, such as K-ADAPTER and MAD-X, we propose a similar approach for incorporating knowledge from multilingual graphs, connecting concepts in various languages with each other through linguistic relationships, into multilingual LLMs for LRLs. Specifically, we focus on eight LRLs --- Maltese, Bulgarian, Indonesian, Nepali, Javanese, Uyghur, Tibetan, and Sinhala --- and employ language-specific adapters fine-tuned on data extracted from the language-specific section of ConceptNet, aiming to enable knowledge transfer across the languages covered by the knowledge graph. We compare various fine-tuning objectives, including standard Masked Language Modeling (MLM), MLM with full-word masking, and MLM with targeted masking, to analyze their effectiveness in learning and integrating the extracted graph data. Through empirical evaluation on language-specific tasks, we assess how structured graph knowledge affects the performance of multilingual LLMs for LRLs in SA and NER, providing insights into the potential benefits of adapting language models for low-resource scenarios.

In the field of prevention and rehabilitation, adaptive and personalized Artificial Intelligence (AI) systems play a vital role in enhancing patient-centric care. These AI systems analyze extensive data on user behavior and situational context to provide tailored support that meets unique needs of individuals, such as patients recovering from knee surgeries. Such patients often face cognitive challenges that interfere with their ability to process complex medical information, make informed decisions, and communicate effectively about their symptoms. These challenges may lead to the generation of incomplete, inaccurate, or biased data, significantly impacting the effectiveness of their treatment. This research introduces the concept of artificial mental models (AMM) integrated into healthcare AI systems. AMMs are cognitive frameworks that encapsulate patient's perceptions and expectations about their therapy and recovery journey. They are beneficial in scenarios requiring nuanced understanding and adaptation to patient's changing conditions, e.g., to assist an amateur soccer player recovering from knee surgery. Here, the AMM acts as a liaison between patient and therapist, helping to devise a personalized exercise regimen that adapts to patient's pain and progress. This work explores the generation of AMMs using Large Language Models instrumental in both eliciting and fine-tuning AMMs to individual patient needs. The research encompasses a prospective study with two phases: elicitation and individualization. The elicitation phase involves creating a bias-free, domain-specific basis AMM through extensive data collection from both quantitative studies and indirect observations, including data on personality traits and expected pain during specific exercises. This basis AMM is evaluated in a technical experiment to ensure it is free from bias and discrimination. In the individualization phase, the basis AMM is refined using direct observations of specific patients, incorporating curated data such as medication details, rehabilitation plans, and patient-reported outcomes, as well as non-curated data like movement patterns and fitness status. The final AMM tailored for individual patients is assessed through action research, involving real-world application to evaluate its impact on rehabilitation outcomes. The research questions focus on whether the predictions made by the AMM about pain are consistent with the patients' experiences and the assessments made by therapists. Results validate the effectiveness of AMMs in real-time clinical settings, demonstrating their potential to significantly enhance the personalization and effectiveness of patient care in knee rehabilitation. Broader implications suggest that AI-driven approaches could enhance patient care across various areas of healthcare, supporting the ability of systems to predict patient needs and improve overall outcomes by more targeted and effective interventions.

Presents an adaptive knowledge distillation framework to create efficient domain-specific language models.