D-optimal Coreset
D-optimal design is useful because fitting information depends mainly on feature dimension rather than raw candidate count. D-optimal design selects a small informative coreset from the candidate set using non-negative candidate weights. C…
1 sources - 5 claims
D-optimal design is useful because fitting information depends mainly on feature dimension rather than raw candidate count. D-optimal design selects a small informative coreset from the candidate set using non-negative candidate weights. Candidates enter the coreset when their D-optimal weights exceed a threshold. Weighted least squares fits the surrogate by minimizing weighted squared error over coreset candidates. Lower D-optimal thresholds usually enlarge the coreset and improve coverage of true top-10 candidates, but do not guarantee better prediction ranking.