Dynamic Periodic Event Graphs for multivariate time series pattern prediction.

Understanding and predicting outcomes in complex real-world systems necessitates robust multivariate time series pattern analysis. Advanced techniques, such as dynamic graph neural networks, have shown significant efficacy for these tasks. However, existing approaches often overlook the inherent periodicity in data, leading to reduced pattern or event prediction accuracy, especially in periodic time series. We introduce a new method, called dynamic Periodic Event Graphs (PEGs), to tackle this challenge. The proposed method involves time series decomposition to extract seasonal components that capture periodically recurring patterns within the data. It also uses frequency analysis to extract representative periods from each seasonal component. Additionally, motif patterns, which are recurring sub-sequences in the time series data, are extracted. These motifs are used to define event nodes using the representative periods extracted from the seasonal components. By constructing periodic motif pattern-based dynamic bipartite event graphs, we specifically aim to enhance the performance of link prediction tasks, leveraging periodic characteristics in multivariate time series data. Our method has been rigorously tested on multiple periodic multivariate time series datasets, demonstrating over a 5% improvement in link prediction performance for both transductive and inductive scenarios. This demonstrates a substantial enhancement in predictive accuracy and generalization, providing confidence in the technique's effectiveness. Reproducibility is ensured through publicly available source code, enabling future research and applications.