Hyperkantprediktion i temporala kunskapsgrafer : Undersökning av entitet-till-kategori-kantprediktion med ett kategorimedvetet regelbaserat ramverk

Temporal Knowledge Graphs have emerged as a powerful way of not only modeling static relationships between entities but also the dynamics of how relations evolve over time. As these informational structures can be used to store information from a real world setting, such as a news flow, predicting future graph components to a certain extents equate predicting real world events. Most of the research in this field focuses on embedding-based methods, often leveraging convolutional neural net architectures. These solutions act as black boxes, limiting insight. In this thesis, we explore an extension to an established rule-based framework, TLogic, that yields a high accuracy in combination with explainable predictions. This offers transparency and allow the end-user to critically evaluate the reason for predictions at the end of the prediction stage. The new rule format named C-TLogic, incorporates entity category as a key component with the purpose of limiting rule application only to relevant entities. In addition, we implement another aggregation method, Max+, as an alternative to the original Noisy-OR method and adapt these to aggregate results both for the purpose of entity and category prediction. To evaluate the frameworks, they are applied to various datasets with varying categorization schemes. In addition to categories assigned in production, clustering-based categorization based on text embeddings and taxonomoy-based categorization is applied to the datasets ICEWS18 and FinDKG. The results suggests the new more complex rule format detracts more from accuracy than the added requirements of matching categories between rule mining and application benefits from leveraging category specific patterns. Strengthening this point, strong performance of the new combination of TLogic with Max+ aggregation on the ICEWS18 dataset indicates the confidence estimation process of TLogic more accurately reflecting the actual accuracy of rules on the test set. This is true in both entity and category predictions, yet the best aggregation method differs across datasets and models. Applying the original TLogic model to the FinDKG dataset yields a very high performance, indicating TLogic’s value in real-world applications on a state-of-the-art dataset in a financial news context.