Keywords: conceptual modeling, encoding, machine learning
Astract: The integration of Conceptual Modeling (CM) and Machine Learning (ML) has given rise to a growing research field known as Machine Learning for Conceptual Modeling (ML4CM), where ML techniques are applied to support modeling tasks such as classifica-tion, completion, or repair. A crucial factor in these applications is the transformation of conceptual models into ML-compatible representations, called encodings. A wide variety of encoding strategies exist that draw on different information sources within conceptual models, depending on the specific use case. However, existing ML4CM studies tend to treat encodings as fixed and focus predominantly on tuning ML algorithms or hyperparameters. Consequently, encoding strategies and their internal configuration options receive limited scrutiny during evaluation, making it difficult for researchers and practitioners to select and adapt optimal encodings for specific tasks.This thesis addresses this gap by developing and evaluating a set of configurable semantic encodings for conceptual models. Specifically, it investigates how semantic information (e.g. names, types, contextualrelationships) within models can be systematically extracted and transformed into ML-compatible representations. The work adopts the Design Science Research methodology and extends the CM2ML framework with an ArchiMate parser and four semantic encoders: Bag-of-Words (BoW), Term Frequency (TF), Embeddings,and Triples. Each encoder captures distinct semantic aspects and supports extensive configurability to enable experimentation and task-specific adaptation. Furthermore, all encodings can be interactively visualized within the framework, offering real-time insight into parameter effects and traceability to link encoded features back to their source model elements.To evaluate the proposed encodings, the thesis combines a qualitative comparison based on defined criteria with a quantitative assessment through two representative ML tasks.The first task, dummy classification, employs TF encodings to distinguish dummy views from valid ones and explores the impact of common NLP parameters and weighting schemes. The second task, node classification, aims to predict element types based on local context, using triple encodings enriched with word embeddings for element names and one-hot vectors for types. The results demonstrate the suitability of the encodings for specific ML4CM tasks and that certain encoding configurations can have a substantial influence on model performance.

Keywords: Process Patterns, Teaching Modelling

Keywords: Conceptual modeling, Data modeling, ER, Genetic algorithms, Graph neural networks, Model Modularization
Astract: Due to the crucial role conceptual models play in explicitly representing a subject domain, it is imperative that they are comprehensible and maintainable by humans. Modularization, i.e., decomposing an overarching, monolith model into smaller modules, is an established technique to make the model comprehensible and maintainable. Genetic Algorithms (GA) have been applied to modularize conceptual models by formulating desired structural model characteristics as multiple objectives. Recently, Graph Neural Networks (GNN)-based methods have shown promising performance in graph processing tasks, including graph clustering but outside the conceptual modeling domain. In this paper, we present a novel approach for GNN-based conceptual model modularization and comparatively analyze our approach against an existing multi-objective GA-based one. Furthermore, we provide a comparative analysis of our novel GNN model against two existing GNN-based graph clustering approaches. We investigate the dependence of the quality of the modularized solutions on the model size. We discuss the comparative results of our novel GNN-based approach and the existing GA-based approach to derive future research lines. Furthermore, our results show, that our proposed GNN-based modularization outperforms the existing GNN-based graph clustering approaches and provides a suitable alternative compared to the GA-based modularization.

Keywords: API classification, BERT, GPT, OpenAPI Specification
Astract: Application Programming Interfaces (APIs) enable interaction, integration, and interoperability among applications and services, contributing to their adoption and proliferation. However, discovering APIs has relied on manual, time-consuming, costly processes that jeopardize their reuse potential and accentuate the need for effective API retrieval mechanisms. Leveraging the OpenAPI Specification as a basis, this paper presents an exploratory study that combines BERT and GPT machine learning models to propose a novel API classifier. Our investigation explored the zero-shot learning capabilities of GPT-4 and GPT-3.5 using relevant terms extracted from API descriptions using BERT. The evaluation of our approach on two datasets comprising 940 API descriptions sourced from public repositories yielded an F1-score of 100% in the small dataset (17 APIs) and 39.1% in the large dataset (923 APIs). These results surpass state-of-the-art on the small dataset with an impressive 29-point improvement. The large dataset showed GPT can suggest labels not in the provided list. Manual analysis revealed that GPT’s suggested labels fit the API intent better in 18 out of 20 cases, highlighting its potential for unknown classes and mismatch detection. This emphasizes the need to improve dataset quality and availability for API research. Our findings show the potential of automated API retrieval and open avenues for future research.

Keywords: Conceptual Modeling, Sustainability

Keywords: Conceptual modeling, Domain modeling, Generative AI, LLM, UML
Astract: The information systems engineering community is increasingly exploring the use of Large Language Models (LLMs) for a variety of tasks, including domain modeling, business process modeling, software modeling, and systems modeling. However, most existing research remains exploratory and lacks a systematic approach to analyzing the impact of prompt content on model quality. This paper seeks to fill this gap by investigating how different levels of description granularity (whole text vs. paragraph-by-paragraph) and modeling strategies (model-based vs. list-based) affect the quality of LLM-generated domain models. Specifically, we conducted an experiment with two state-of-the-art LLMs (GPT-4o and Llama-3.1-70b-versatile) on tasks involving use case and class modeling. Our results reveal challenges that extend beyond the chosen granularity, strategy, and LLM, emphasizing the importance of human modelers not only in crafting effective prompts but also in identifying and addressing critical aspects of LLM-generated models that require refinement and correction.

Keywords: Agility, Business Process Management, Conversational Agents, LLM, Social media
Astract: The traditional approaches in dynamic and collaborative environments that use Business Process Management (BPM) methodologies usually lack the ability to adapt to real-time changes in case of heavy human involvement in repetitive processes. The agility of social BPM is, however, still limited because of a lack of context-sensitive tool support. This paper proposes a mapping framework that leverages conversational AI agents on a social media platform to enhance BPM agility. AI-driven conversational agents are mapped to the respective phases of the BPM lifecycle to provide real-time guidance, recommendations, and context-sensitive feedback. The agents’ collaborative features enable inclusive co-construction, interactive task execution, and continuous monitoring of the processes. That allows dynamic adaptation of the processes in case of changes so that tasks remain aligned with the users’ needs and contextual demands. This framework is developed through an exploratory approach that integrates literature review, deductive design, and use case-based evaluation. This framework could bridge gaps in the current BPM practices by integrating BPM, AI, and social media, thereby offering a new model for agile and collaborative business process management.

Keywords: Artificial Intelligence, Event Log Completion, Event Log Creation, Event Log Quality Improvement, Multimodal data
Astract: Process mining uses data from event logs to understand which activities were undertaken, their timing, and the involved entities, providing a data trail for process analysis and improvement. However, a significant challenge involves ensuring that these logs accurately reflect the actual processes. Some processes leave few digital traces, and their event logs often lack details about manual and physical work that does not involve computers or simple sensors. We introduce the Business-knowledge Integration Cycles (BICycle) method and mm_proc_miner tool to convert raw and unstructured data from various modalities, such as video, audio, and sensor data, into a structured and unified event log, while keeping human-in-the-loop. Our method analyzes the semantic distance between visible, audible, and textual evidence within a self-hosted joint embedding space. Our approach is designed to consider (1) preserving the privacy of evidence data, (2) achieving real-time performance and scalability, and (3) preventing AI hallucinations. We also publish a dataset consisting of over 2K processes with 16K steps to facilitate domain inference-related tasks. For the evaluation, we created a novel test dataset in the domain of DNA home kit testing, for which we can guarantee that it was not encountered during the training of the employed AI foundational models. We show positive insights in both event log enrichment with multimodal evidence and human-in-the-loop contribution.

Keywords: Design Process, Design Science Research, Soft Systems Methodology
Astract: Several efforts have been undertaken to define generic guidelines that address specific gaps in the ‘build’ activity of Design Science Research (DSR) artifacts, i.e., constructs, models, methods and frameworks, and instantiations. Unfortunately, explicit guidance is still lacking on how to coherently operationalize such guidelines when building a DSR artifact, particularly a framework. In addition, there is no an elaborate procedure or logical thinking pattern that can be followed when building a DSR artifact, particularly a framework for solving an unstructured problem. Consequently, structural compositions of some artifacts insufficiently subscribe to several general design guidelines, which often hinders the artifacts from fulfilling their intended purposes. To address this gap, Soft Systems Methodology can be leveraged during the design cycle of a DSR initiative, to elaborate the ‘build’ activity and simultaneously support the coherent operationalization of existing general design guidelines. This is demonstrated herein by presenting a Technique of Building Frameworks for guiding Interventions against unstructured problems (TBUFI). From 2011 to 2023, TBUFI has undergone 11 evaluation iterations, which involved: (a) using it to support the building of frameworks for guiding digital interventions in ten research studies; and (b) engaging information systems specialists in a group walkthrough meeting to deliberate its structural composition. Evaluation iterations since 2011 (including feedback from information systems specialists) confirm TBUFI’s ability to successfully guide the design of frameworks that can direct interventions against complex and unstructured problems, by making their ‘build’ activity more elaborate, coherent, and aligned with existing general design guidelines. Thus, TBUFI can be perceived as a supplement for the ‘build’ activity in DSR.

Keywords: Conceptual fidelity, Domain modeling, Return on modeling effort, Views
Astract: Humanity has long since used models, in different shapes and forms, to understand, redesign, communicate about, and shape, the world around us; including many different social, economic, biological, chemical, physical, and digital aspects. This has resulted in a wide range of modeling practices. When the models as used in such modeling practices have a key role to play in the activities in which these practices are ‘embedded’, the need emerges to consider the effectiveness and efficiency of such processes, and speak about modeling capabilities. In the latter situation, it also becomes relevant to develop a thorough understanding of the artifacts involved in modeling practices/capabilities. One context in which models play (an increasingly) important role is model-driven systems development, including software engineering, information systems engineering, business process engineering, enterprise engineering, and enterprise architecture management. In such a context, we come across a rich variety of modeling related artifacts, such as views, diagrams, programs, animations, specifications, etc. In this paper, which is actually part of an ongoing ‘journey’ in which we aim to gain deeper insights into the foundations of modeling, we take a fundamental look at the variety of modeling related artifacts as used in the context of model-driven (systems) development, while also presenting an associated framework for understanding, synthesizing the insights we obtained during the ‘journey’ so-far. In doing so, we will also argue that the aforementioned artifacts are actually specific kinds of models, albeit for fundamentally different purposes. The provided framework for understanding involves definitions of domain model, the Return on Modeling Effort (RoME), the conceptual fidelity of domain models, as well as views as a mechanism to manage the complexity of domain models.