Weight Vector Tuning and Asymptotic Analysis of Binary Linear Classifiers

Unlike its intercept, a linear classifier's weight vector cannot be tuned by simple grid search. Hence, this paper proposes tuning of generic binary classifier through the parameterization decomposition discriminant scalar which controls trade-off between conflicting informative and noisy terms. By varying parameter, original is modified in meaningful way. Applying method to number classifiers under variety data dimensionality sample size settings reveals that classification performance loss due non-optimal native hyperparameters can compensated for tuning. This yields computational savings as proposed reduces compared hyperparameter, may involve repeated generation along with burden optimization, reduction, etc., depending on classifier. It also found significantly improves Linear Discriminant Analysis (LDA) high estimation noise. Proceeding from second finding, an asymptotic study misclassification probability parameterized LDA growth regime where are comparable conducted. Using random matrix theory, shown converge quantity function true statistics data. Additionally, estimator derived. Finally, computationally efficient parameter using demonstrated real