Shuwen Xiao

Taobao's recommender system just got a 1.65% CTR boost by compressing ultra-long user behavior sequences with a hierarchical codebook and sparse attention, proving that personalized interest centers can be learned efficiently.