Rinko - Towards a Rubust Modeling of Temporal Interest Change for Behavioral Targeting

Towards a Robust Modeling of Temporal Interest Change Patterns for Behavioral Targeting

Mohamed Aly(Seeloz Inc.), Sandeep Pandey(Twitter),

Vanja Josifovski(Google), Kunal Punera(Relate IQ)

WWW2013

Presented by Annaka 2013-11-07

Summary

2

SVMWEB

(i)(ii)

conversion/email/

conversion()4%

Summary

3

SVMWEB

(i)(ii)

conversion/email/

conversion()4%

Outline

1. / WEB

2. 3 + 4

3.

4.

4

focus

Outline

1. / WEB

2. 3 + 6

3.

4.

5

6

() [19th ~ ]

SVMWEB[2000~]

7

WEB

8

SVMWEB[2000]

frequencyintensityrecency

absolute-long-termabsolute-short-termrelative-long-term-positiverelative-long-term-negativerelative-short-term-positiverelative-short-term-negative

6

Outline

1. / WEB

2. 3 + 6

3.

4.

9

3 + 4

10

P1 P2 P3

1

2

3

click etc

3 + 4

11

FrequencyX

Decayed-intensityX

RecencyX

3

3 + 4

12

- Fr

equency

2p 4q

d

ecayed intensity

-

Recency

()

()

3 + 4

13

absolute-long-term

absolute-short-term

relative-long-term

relative-short-term

4

3 + 6

14

pi

pi

absolute-long-term6

3 + 6

15

absolute-short-term6

Pi

3 + 6

16

relative-long-term-positive6

negative

i[/day]

i[/day]

(i) /i

3 + 6

17

relative-short-term-positive6

negative

1i[/day]

(i) / (i)

2i[/day]

Outline

1. / WEB

2. 3 + 6

3.

4.

18

19

58282005283.4 66%33%

conversion-weighted average of AUC

20

33 P

Outline

1. / WEB

2. 3 + 6

3.

4.

21

22

33

23

24

25

P=14[day]

P

26

27

28

2 twitter