Optimizing the joint item- and token-level hyperparameters in SToICaL

Identify the performance-optimal combination of item-level reweighting and token-level prefix-tree marginalization hyperparameters in the combined SToICaL loss to balance nDCG and recall-at-k metrics.

Background

The authors present supplemental experiments combining fractional item-level reweighting (parameterized by α) with token-level target distributions from trie-based marginalization (parameterized by β). While some settings improve nDCG, the optimal joint configuration is not determined, and performance does not consistently surpass item-level reweighting alone. They explicitly leave finding the ‘sweet spot’ combination to future work.