One-Step Graph-Structured Neural Flows for Irregular Multivariate Time Series Classification
Researchers have introduced One-Step Graph-Structured Neural Flows (GSNF), a novel machine learning model designed to improve the classification of irregular multivariate time series. While existing Neural Flow methods efficiently learn Ordinary Differential Equation (ODE) solution trajectories, they often overlook inter-variable interactions due to their one-step mapping nature. GSNF addresses this limitation by incorporating two auxiliary-trajectory self-supervision strategies: interaction-aware trajectory generation via re-initialization and reverse-time trajectory generation. These techniques induce trajectory divergence to expose graph-induced interactions and enforce forward-backward consistency to regularize graph learning. Theoretical bounds on divergence are derived to support the approach. Experimental evaluations across five real-world datasets demonstrate that GSNF achieves state-of-the-art classification performance. Furthermore, the model maintains highly competitive training times and memory usage compared to existing methods, offering a robust solution for complex time-series data analysis in artificial intelligence applications.
Wire timeline
One-Step Graph-Structured Neural Flows for Irregular Multivariate Time Series Classification
Researchers have introduced One-Step Graph-Structured Neural Flows (GSNF), a novel machine learning model designed to improve the classification of irregular multivariate time series. While existing Neural Flow methods efficiently learn Ordinary Differential Equation (ODE) solution trajectories, they often overlook inter-variable interactions due to their one-step mapping nature. GSNF addresses this limitation by incorporating two auxiliary-trajectory self-supervision strategies: interaction-aware trajectory generation via re-initialization and reverse-time trajectory generation. These techniques induce trajectory divergence to expose graph-induced interactions and enforce forward-backward consistency to regularize graph learning. Theoretical bounds on divergence are derived to support the approach. Experimental evaluations across five real-world datasets demonstrate that GSNF achieves state-of-the-art classification performance. Furthermore, the model maintains highly competitive training times and memory usage compared to existing methods, offering a robust solution for complex time-series data analysis in artificial intelligence applications.
cs.AI updates on arXiv.org