Interpretable Machine Learning

Structured and transparent models for sparse learning, multiscale approximation, time-series reconstruction, uncertainty, and validation.

Focus

My interpretable machine-learning work is about making modeling decisions inspectable. In many scientific and operational settings, the model must reduce noise, compress data, reconstruct a signal, or produce a prediction while preserving enough structure for review, uncertainty assessment, and downstream decisions.

Core Themes

• Sparse multiscale learning for transparent data reduction.
• Hierarchical regularization and bias-variance-compression tradeoffs.
• Confidence and prediction intervals for compressed models.
• SVD-based augmentation for surrogate modeling under limited data.
• Localized penalized splines for noisy and irregular time series.
• Outlier screening, reconstruction, derivative estimation, and uncertainty bands.

Representative Work

• Forward-backward greedy sparse multiscale learning.
• Hierarchical regularization networks for sparsification-based learning on noisy datasets.
• Hierarchical approximations for data reduction and learning at multiple scales.
• SVD-enabled data augmentation for machine-learning surrogate modeling of nonlinear structures.
• ALPS framework for modeling time series of land ice changes.
• Glacier surge monitoring from temporally dense elevation time series.

Decision-Science Connection

The common pattern is not just prediction. The work asks which information is decision-relevant, which detail is redundant or unstable, how much approximation error can be tolerated, and how uncertainty should be shown before a model output is trusted.