iLPCtd: Interpretable Link Prediction on Continuous Time Durations with Extracted Logical Rules from Temporal Knowledge Graphs
Shijie Yang, Shaochong Du, Yongzhan Lu, Hong Huo, Tao Fang · 2024
Temporal Knowledge Graphs (TKGs) augment traditional knowledge graphs by integrating temporal attributes, transforming them into a more nuanced and dynamic data structure capable of modeling real-world events. Representation learning techniques, particularly embedding-based methods, have been widely adopted for reasoning over TKGs. However, these methods often suffer from a black-box nature, which limits their ability to provide transparent reasoning chains and logically consistent link prediction outcomes. Moreover, many of these approaches overlook the duration of events, despite the fact that most real-world events span over periods of time rather than occurring instantaneously. This paper proposes a novel model, Interpretable Link Prediction on Continuous Time Durations (iLPCtd), which addresses both the interpretability of link predictions and the consideration of continuous temporal dynamics. By leveraging a random walk approach, iLPCtd extracts first-order temporal logical rules from given TKGs and applies them for subsequent link predictions. Importantly, the model incorporates the continuous nature of event timestamps to more accurately reflect real-world scenarios. To evaluate the model's performance, we constructed a new dataset, Satellite Observation to Earth (SOE), based on satellite meteorological observation workflows. Experimental results demonstrate that iLPCtd outperforms existing methods on this dataset. Additionally, experiments conducted on public datasets also show promising results, indicating that iLPCtd is not only capable of generating interpretable logical reasoning chains but also excels in handling complex temporal link prediction tasks, offering both theoretical and practical improvements in TKG reasoning.