存储预算下稀有物品召回：量化、蒸馏与教师保真度的局限
原标题：Preserving rare-item recall under a storage budget: quantization, distillation, and the limits of teacher fidelity
AI 摘要
该论文研究在固定存储预算下，推荐系统应如何分配字节以保留稀有物品召回率。作者在七个数据集-骨干组合上比较了七种候选方案，发现将字节用于更高秩而非更高精度（如秩300、4比特量化）能保留教师模型95-103%的尾部召回率，而同等字节下秩40模型尾部召回最低。研究还指出，以教师命中保留率衡量压缩模型会误判尾部召回表现，建议用总尾部召回而非教师保真度来评估压缩推荐器。
正文节选
Preserving rare-item recall under a storage budget: quantization, distillation, and the limits of teacher fidelity Austin, Texas, United States 2Unaffiliated, Fairfax, Virginia, United States Abstract Recommenders often have to fit their user and item embeddings into a fixed storage budget, for example the memory of an online feature store. The bytes can go to a small full-precision model, a student distilled from a larger teacher, or a higher-rank model stored at lower precision. We asked which