EXCODER: EXplainable Classification Of DiscretE time series Representations

Discrete representations improve explainability: The paper demonstrates that converting raw time series data into simplified, discrete formats using techniques like VQ-VAE and DVAE makes AI model decisions easier to understand by removing noise and focusing on the most important patterns, rather than trying to explain decisions based on messy raw data.

Explainability doesn't require sacrificing accuracy: A key finding is that these more interpretable discrete representations maintain the same classification performance as traditional deep learning approaches, meaning organizations don't have to choose between having accurate AI systems and being able to explain how they work.

New metric validates explanation quality: The paper introduces "Similar Subsequence Accuracy" (SSA), a measurable way to check whether the parts of time series data that an AI system highlights as important actually match what the system learned during training—addressing a critical gap in verifying that explanations are genuinely faithful to how the model operates.

Practical benefits for real-world deployment: By producing "compact, interpretable, and computationally efficient explanations," this approach could make it easier for organizations to deploy explainable AI systems in practice, particularly for time-based data like medical monitoring, financial trends, or sensor readings where understanding why a decision was made is crucial for trust and accountability.