Time Aligned Sliding Graph Embeddings for Dynamic Time Series Analysis

Long, multichannel time series—such as Electroencephalogram (EEG), climate records, or financial data—contain rich temporal dynamics and evolving relationships between segments. Traditional sliding window approaches capture short-term features but ignore how neighboring windows relate. Graph Neural Networks (GNNs) can model these relationships, yet most existing methods treat graphs as static or rely only on final classification outputs. This paper introduces a unified approach that combines sliding-graph construction with pre-final embeddings from a GNN Graph Classification model to create a dense vector timeline. Each embedding captures the structure of a single sliding graph and is linked to its center time, creating a time-aligned sequence of vectors that allows dynamic signals to be queried, compared, and analyzed over time. The framework also captures spatial structure: by generating parallel timelines for multiple channels and comparing them pairwise, it reveals evolving relationships between signal sources. EEG recordings during rest and sleep are used as an illustrative case study, showing how the method uncovers state-dependent connectivity patterns that static or label-only approaches overlook. While demonstrated on EEG, the framework generalizes to other domains, offering new opportunities for structure-aware time series analysis.