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.
More from Research
- Reconstructing Articulated Objects from a Single Rest-State Image — snu · 2026-08-12
- DistilVDR: A 524M Compact Visual Document Retriever Distilled from 8B — nanovdr · 2026-08-12
- AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss — Kwai-Kolors · 2026-08-12
- AI Agents Still Can't Conduct Open-Ended Research Due to Cognitive Failures — mikeflache · 2026-08-12
- A Simple Equation to Solve Trainer-Inference Mismatch in LLMs — willccbb · 2026-08-12
- EigenLabs Builds Multi-Agent Environments for Autonomous Scientific Research — gajesh · 2026-08-12