Time Series Data Classification using the Knowledge-Augmented Transformer Model

Time Series Classification predicts outcomes by analyzing sequential data that changes over time. As the complexity of time series data increases, particularly in scenarios involving dynamic relationships, existing models face limitations. This paper proposes a classification model using an encoder-based transformer with knowledge augmentation to address these challenges. The model preprocesses time-series data into a graph structure, enabling effective learning of intricate relationships. Relational attention and temporal attention enhance focus on critical relationships and time points, respectively. Additionally, symbolic reasoning through knowledge augmentation allows the model to capture complex, implicit information. Experimental results show that the proposed model outperforms existing SOTA model on complex datasets like SC2EGSet.