hungarian-discrete-vs-soft-ot-metrics

Summary

The Hungarian algorithm locks each item to exactly one partner, so the only fair score is whether that single top pick was correct, while soft optimal-transport methods spread an item's "mass" fractionally across several partners, which means they also get graded on average rank and embedding similarity. This matters because the two families of methods live in different output spaces and use different scoring rubrics, so their numbers are not directly comparable on a single leaderboard.