PIP: Pictorial Interpretable Prototype Learning for Time Series Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time series classification models are often black boxes, hindering deployment in critical applications due to lack of user trust.
• Existing interpretable models for time series data generate explanations too complex for non-engineers.

Purpose of the Study:

• To develop a novel deep learning architecture for time series classification that provides user-tailored explanations.
• To enhance user trust and understanding of time series classification models through interpretable visual prototypes.

Main Methods:

• Introduction of PIP (Pictorial Prototype) deep learning architecture.
• Joint learning of classification models and visual class prototypes.
• Training the model using user-selected class illustrations for personalized explanations.

Main Results:

• PIP generates user-friendly explanations by leveraging end-user defined visual prototypes.
• An end-user experiment demonstrated PIP's effectiveness in communicating learned concepts to non-experts.
• PIP achieved an improved balance of accuracy and interpretability compared to baseline methods.

Conclusions:

• PIP offers a novel approach to interpretable time series classification.
• Visual, user-defined prototypes enhance the understandability and trustworthiness of AI models.
• The architecture shows promise for broader adoption of time series classification in real-world applications.