Try it out from the Priority Inbox settings tab. Doug Aberdeen, Ond - - PowerPoint PPT Presentation

Doug Aberdeen, Ondřej Pacovský, Andrew Slater

Try it out from the Priority Inbox settings tab.

PASSIVE-AGGRESSIVE LOGISTIC REGRESSION

• Crammer, Dekel, Keshet, Shalev-Shwartz, Singer: “Online Passive-Aggressive Algorithms”, 2006
• 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.

wi ← wi + fi max(e − ǫ, 0) f2 +

1 2λ

e = 1 − p if important; −p

• therwise.

SIMPLE TRANSFER LEARNING

• Glut of data globally, dearth of data per user.

+

f1 fn

w1

wn

+

f1 fn

+

s

Global model: User model:

fn+1 fn+k

User only features

g1

gn

wn+1

wn+k

~100k users per shard. 20 -- 30k f/sec/core.

SCALING

Profile/ last action pass Model update pass BT fetch Writeback t i m e

Row key prefix is user ID. But fast Bigtable reads are not in row order!

P r e fi x

• 1

P r e fi x 2

• 3

P r e fi x 4

• 5

P r e fi x 6

• 7

Task 0: Row prefix 0-7

1/8th of users

Task 1: Row prefix 8-f

In task sharding

Why not map-reduce?

FEATURES

Social features Content features Thread features Label features Spam features

~200 global features + personal