Try it out from the Priority Inbox settings tab. Doug Aberdeen, Ond ř ej Pacovsk ý , Andrew Slater
PASSIVE-AGGRESSIVE LOGISTIC REGRESSION • Crammer, Dekel, Keshet, Shalev-Shwartz, Singer: “ Online Passive-Aggressive Algorithms ”, 2006 � 1 − p max( e − ǫ , 0) if important; w i ← w i + f i e = � f � 2 + 1 otherwise . − p 2 λ • A message is important if it’s read/replied/starred/marked within a time limit. • λ is a regularisation parameter that controls “aggressiveness’’. • ε is the “passiveness”, related to the hinge loss.
SIMPLE TRANSFER LEARNING • Glut of data globally , dearth of data per user. f 1 g 1 + Global model: f n g n + s f 1 w 1 + User model: f n w n w n + k w n +1 f n +1 f n + k User only features
SCALING Row key prefix is user ID. But fast Bigtable reads are not in row order! Task 1: Task 0: Row prefix 8-f Row prefix 0-7 1 ~100k users - 0 x Profile/ fi 1/8th of users e per shard. r last action P pass 3 BT fetch - 2 In task sharding x 20 -- 30k f/sec/core. fi e e r P m i t Model 5 - update 4 x pass Why not fi e r P map-reduce? 7 Writeback - 6 x fi e r P
FEATURES ~200 global features + personal Social features Content features Thread features Label features Spam features
Recommend
More recommend