Shapelet Discovery by Lazy Time Series Classification.

Area of Science:

• Machine Learning
• Data Mining
• Time Series Analysis

Background:

• Time series shapelets are crucial for discriminative feature representation in classification.
• Existing shapelet-based models often suffer from high computational complexity and poor interpretability due to evaluating shapelets on the entire dataset.

Purpose of the Study:

• To enhance the efficiency and interpretability of shapelet discovery in time series classification.
• To address the limitations of traditional methods by incorporating instance-specific and classwise feature frequency information.

Main Methods:

• A lazy strategy fusing global and local similarities is employed to improve shapelet discovery efficiency.
• A novel approach learns instance-specific evaluation datasets during prediction to reduce classification uncertainty.
• A shapelet coverage score is introduced to quantify the discriminability of time stamps across different classes.

Main Results:

• The proposed method demonstrates competitive performance against benchmark time series classification techniques.
• The approach provides enhanced insights into discriminative features within individual time series and across different classes.
• Improved computational efficiency and interpretability were achieved compared to traditional shapelet-based models.

Conclusions:

• The developed method offers a more efficient and interpretable solution for time series shapelet discovery.
• The instance-specific evaluation and shapelet coverage score contribute to better classification performance and feature understanding.
• This research advances the field of time series analysis by providing a novel and effective shapelet-based classification strategy.