Improving Uncertainty Quantification of Variance Networks by Tree-Structured Learning