On the Regularization Landscape for Linear Recommendation Models

Many linear recommendation models inspired by deep architectures show surprisingly identical performance on standard benchmarks. This paper reveals that beneath varying heuristics such as dropout or autoencoders, these methods essentially reduce to either nuclear-norm or Frobenius-norm regularizations. While nuclear-norm variants offer closed-form low-rank solutions at the cost of expressiveness, Frobenius-norm counterparts remain expressive but rely on full-rank representations or intricate numerical solvers. By generalizing Frobenius regularizers, the authors introduce closed-form, low-rank solutions that synthesize the strengths of both paradigms.