@mastersthesis {pub5907,
	title = {Knowledge Graph Completion on a Situational Knowledge Graph Describing Household Actions},
	author = {Mariam Arustashvili},
	year = {1970},
	month = {January},
	abstract = {The primary objective of this thesis is to explore the potential of various KGC methods to improve the accuracy and adaptability of situational knowledge graphs in a robotics context. Specifically, my goal is to assess how well traditional graph-based machine learning and language models can predict parent actions {\textemdash} high-level tasks {\textemdash} and anticipate next sub-actions {\textemdash} granular steps {\textemdash} based on knowledge of the environment and user behavior. This exploration aims to identify the strengths and limitations of these methods as a step towards improving prediction accuracy and enhancing the adaptability of robotic systems. To achieve this, I focus on three main objectives. The first objective is to establish baseline models for predicting both parent actions (high-level tasks such as {\textquotedblright}making coffee{\textquotedblright}) and sub-actions (granular steps like {\textquotedblright}grabbing a cup{\textquotedblright} or {\textquotedblright}pouring water{\textquotedblright}). Developing these baselines provides a foundational comparison point for more advanced techniques, ensuring that any improvements can be measured against a simple, rule-based approach. The second objective is to explore and evaluate various graph-based machine learning models for link prediction within situational knowledge graphs. This involves using the PyKEEN library [1] to implement state-of-the-art models like TransE [5], RotatE [41], and DistMult [51], among others. These methods are tested on their ability to predict the next sub-action and the parent action from a sequence of sub-actions, assessing how well they capture complex, multi-step
relationships in the action sequences. The third objective is to integrate and evaluate the performance of LLMs as a complementary tool for action prediction. Given the cost-efficiency and potential for real-time application, GPT-4o-mini will be used to compare its predictive capabilities against the graph-based methods [37]. I assess the strengths and limitations of LLMs in this context, particularly their ability to adapt to dynamic environments and their computational efficiency compared to graph-based models. These methods are evaluated in terms of prediction accuracy, computational efficiency, explainability, and applicability to real-time robotics systems. by improving the ability of situational knowledge graphs to predict actions, I aim to contribute to the development of more intelligent, user-adaptive robotic systems that can operate effectively in dynamic, real-world environments.},
	publisher = {University Mannheim},
	booktitle = {Mannheim Univeristy}
}