Work

Tulkittavuusohjattu regularisointi asemakohtaisessa sääennustamisessa

Year: 2026
Type: dissertation
Abstract: This thesis studies interpretability-driven regularization for station-level weather forecasting. Each data point corresponds to one calendar day at the Finnish Meteorological Institute station Nuuksio; the feature vector contains same-day weather variables and cyclical calendar features, and the label is next-day maximum temperature. The baseline predictor is a multilayer perceptron trained by empirical risk minimization using mean squared error. Two regularization mechanisms are studied: feature-importance regularization using a learnable input gate and a pseudo-label vector, and monotonicity regularization using local comparison pairs and hinge-style penalties. The empirical results show that feature-importance regularization strongly improves alignment between the learned gate and the pseudo-label prior with only small changes in predictive error. However, parameter-level gate alignment does not by itself guarantee corresponding function-level feature use. Monotonicity regularization has a smaller main effect in the selected task because the baseline predictor already exhibits very low monotonicity violation rates on the selected local comparison set, but negative-control experiments show that the regularizer is active and can substantially change learned behavior when incorrect priors are imposed. The thesis contributes a precise applied machine learning formulation in which user-specified interpretability priors are encoded directly in the training objective and evaluated jointly through predictive performance, parameter-level alignment, function-level diagnostics, and negative controls.
Language: English
Open Access status: green
OpenAlex ID: W7165502345
Cites: 30
Cited by:
Related to: 10