Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training

_reachsumit · x · 2026-08-12

Sampled softmax is commonly used in large-scale neural recommenders to reduce memory costs. However, under a fixed memory budget, it remains unclear whether to prioritize a larger batch size or more negative items.

Under standard smoothness and variance assumptions, this paper provides theoretical and empirical evidence that the fastest convergence arises from including as many objects as possible (i.e., maximizing batch size $n$). The findings were validated across four real-world sequential recommendation benchmarks, including MovieLens-20M.

Original post →

More from Research

Research channel →