Shapelet Bottleneck Models for Explainable Time Series Classification

The patent segments the time series data into shapelet patterns and extracts meaningful temporal motifs from the data. By breaking down the complex time series into recognizable patterns (shapelets), the model maintains high classification performance while providing interpretable explanations about which patterns were detected and why they are significant for the classification decision.

The patent introduces shapelet patterns as intermediary representations between the raw time series data and the classification decision. These shapelets act as mediators that capture temporal dependencies and relationships, enabling the model to achieve state-of-the-art performance while maintaining interpretability through the extracted patterns.

The patent merges the shapelet concept bottleneck model with a deep neural network in a hybrid architecture. The SCBM handles the interpretability aspect by extracting and explaining shapelet patterns, while the DNN handles the classification performance. This combination allows the system to achieve both explainability and high performance without excessive training complexity, as each component focuses on specific aspects of the problem.

The patent employs regularization techniques and additional classification loss functions that modify the training parameters and objective function. By adjusting these parameters and adding constraints, the model learns to produce both accurate classifications and interpretable explanations, balancing the trade-off between performance and explainability through parameter optimization.