Category: Machine Learning

Summary: Testing whether late-applied weight decay and class-balanced reweighting interact non-additively on minority-class generalization under compositional shift.

Weight decay and class-imbalance reweighting are both common tools, but they are usually evaluated one at a time. This experiment asks whether applying weight decay late in training changes the effect of class-balanced reweighting on minority-class performance in a way that cannot be predicted by simply adding their separate benefits.

The model trains classification systems under compositional shift while comparing combinations of timing for weight decay and class-imbalance reweighting. The focus is on minority-class generalization, where interaction effects are likely to matter most.

That makes the project a mechanism test rather than a parameter sweep for best accuracy alone. The result targets whether optimization timing and data reweighting jointly shape generalization in a genuinely nonlinear way.

Method: Controlled training sweeps combining late weight decay schedules with class-balanced loss reweighting under compositional shift.

What is measured: Minority-class generalization, interaction effect size, class-wise performance, compositional-shift accuracy, and comparison with additive expectations.