关于

• 解决的问题
  1. id太多，embedding参数太多，在线打分困难
     • hash trick
     • low-rank factorization：ALBERT: A Lite BERT for Self-supervised Learn- ing of Language Representations
     • quantization：Differentiable Product Quantization for End- to-End Embedding Compression
  2. 不同频率的item的embedding需要不同的容量(长度/状态数)

Differentiable Product Quantization

• KD encode