Machine Learning Canvases for Project Support and Application of Graph Neural Networks in Invoice Recognition

Motivation: Modern machine learning (ML) projects are complex in many respects. From an organizational perspective, this concerns, for example, the number and type of stakeholders or the collaboration of development teams. From a technical perspective, high requirements arise from the development of algorithms or the size, availability, and processing of data. These various types of complexity make it difficult for small organizations to implement ML projects purposefully. For this reason, conceptual and technical solutions must be sought to support users, managers, and developers in these contexts. From a conceptual point of view, it is advisable to create clear process descriptions and overviews in order to gain clarity about the project at hand. From a technical perspective, concrete ML use cases should be implemented to overcome technical obstacles. In the specific use case of this thesis, this refers to the improvement of Graph Attention Networks (GAT) for invoice recognition (IR). The research question - which methods and concrete models support Small and medium-sized enterprises in implementing ML - as well as the application context are derived from the collaboration with a partner company. Research Method: In this dissertation, a holistic approach is applied to advance the Data Science domain. Design Science Research and Case Studies are used, to make conceptual as well as technical progress for machine learning projects. A comprehensive literature review builds the foundation for the conceptual section, upon which artifacts are built. Model development in program code as well as evaluation form the central part of the technical section. From the big picture to the detailed view, general collaboration and prioritization aspects in ML projects are first explored and then the ML use case of invoice recognition with Graph Neural Networks is examined in detail. This mixed methods approach allows for a comprehensive view of the field of application. Contribution: Five research papers result in two major contributions for the practice and research community - a catalog and a model. The catalog lists ML-canvases and their corresponding fields and questions. It is derived from a comprehensive review of the related literature. This catalog supports the initiation and consistent implementation of machine learning projects. In particular, specific questions help to understand the problem that needs to be solved, the value that is supposed to be generated through the ML project, and the steps to reach a feasible solution. The second contribution is a specific model for invoice recognition. In this use case, special challenges and information types of invoice documents - visually rich documents - are analyzed and categorized. Subsequently, a pre-trained multi-modal Graph Attention Network is enhanced and tested on different datasets. Integrating an LLM in a Graph Attention Network to provide semantic embeddings is a new research approach in this thesis. Applying the model to English and German invoice documents, a significant improvement in token classification can be achieved. These contributions support machine learning projects on both a conceptual and a technical level.