Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training
Under the same memory budget, the paper argues for spending capacity on batch size, not more sampled negatives.
The authors analyze sampled-softmax recommender training where memory scales with batch size times negative candidates. Their rule of thumb is to use as many training objects as possible, with very few negatives, when memory is fixed. They report support from synthetic tests and four sequential recommendation benchmarks, including MovieLens-20M, where that setup converged faster and improved final recommendation quality versus other allocations. ArXiv · AI/CL/LG's note
The authors analyze sampled-softmax recommender training where memory scales with batch size times negative candidates. Their rule of thumb is to use as many training objects as possible, with very few negatives, when memory is fixed. They report support from synthetic tests and four sequential recommendation benchmarks, including MovieLens-20M, where that setup converged faster and improved final recommendation quality versus other allocations. ArXiv · AI/CL/LG's note
score 4