Surprisal-Based Episode Segmentation and Dynamic Cluster Graph Propagation for Multivariate Time Series Relationship Learning

Key Points

• The research aims to improve time series prediction performance by capturing relationships in sensor data patterns.
• Used surprisal-based boundaries for episode segmentation of input time series.
• Implemented Gumbel-Softmax to cluster episodes into representative vertices.
• Applied dynamic relationship modeling via a graph convolutional network.
• Reduced total vertices from 127 to 64, cutting GCN computation by about 66.6% and inference time by 33.7%.
• Achieved similar or improved mean squared error (MSE) in some conditions relative to existing models.
• Consistent performance improvements observed in the Weather dataset, confirming the generalization of the proposed method.