Time series classification requires effective feature representations that capture temporal dependencies and structural patterns. Symbolic Aggregate approXimation (SAX) can reduce dimensionality while preserving essential time series patterns, but SAX-based approaches often lack contextual information beyond local symbol occurrences. Here, we introduce a novel method that combines SAX with Graph Neural Networks (GNNs), a robust deep learning framework that generates meaningful embeddings for time series classification. SAX words are extracted and structured as a directed graph, where consecutive words form weighted edges. GNN s are then applied to learn embeddings, which are classified using a Random Forest (RF) classifier. The method is evaluated on a diverse set of benchmark datasets from the UCR Time Series Archive, which provides insight into its performance across different types of time series data.