This repository contains a Graph Neural Network (GNN) model designed for time series analysis. The model leverages graph-based representations to capture complex dependencies in sequential data.
Traditional time series models often struggle to capture relationships beyond simple temporal dependencies. By representing time series data as graphs, the GNN model can learn both temporal and relational patterns, improving forecasting and anomaly detection tasks.
-
Graph Construction
- Each feature/signal is treated as a node in a graph.
- Edges represent dependencies (e.g., temporal adjacency, similarity, or domain-specific relationships).
-
Node Feature Encoding
- Raw time series values or extracted features are assigned to nodes.
- Optional preprocessing includes normalization or embedding.
-
Graph Neural Network Layers
- Message passing aggregates information from neighboring nodes.
- Multiple GNN layers allow the model to learn higher-order dependencies.
-
Readout & Prediction
- Node or graph-level embeddings are used for downstream tasks (e.g., forecasting future values, classification).
- Output layers map embeddings to predictions.