18
1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

Embed Size (px)

Citation preview

Page 1: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

1

Learning User Clicks in Web Search

Ding Zhou et al.The Pennsylvania State

UniversityIJCAI 2007

Page 2: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

2

What is Click Prediction?

microsoft xbox 360 kinect game1. http://www.xbox.com/zh-TW/kinect2. http://www.xbox.com/en-US/kinect3. http://en.wikipedia.org/wiki/Kinect

Page 3: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

3

What is Click Prediction?

microsoft xbox 360 kinect game11 http://www.xbox.com/zh-TW/kinect11 http://www.xbox.com/en-US/kinect11 http://en.wikipedia.org/wiki/Kinect

Page 4: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

4

Click Prediction Model

P(http://www.xbox.com/zh-TW/kinect | microsoft xbox 360 kinect game)

P(http://www.xbox.com/en-US/kinect | microsoft xbox 360 kinect game)

P(http://en.wikipedia.org/wiki/Kinect | microsoft xbox 360 kinect game)

Page 5: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

5

Two approaches

1. Full Model: PF(http://www.xbox.com/zh-TW/kinect | “microsoft xbox 360 kinect game”)

2. Independent Model: PI(http://www.xbox.com/zh-TW/kinect | microsoft, xbox, 360, kinect, game)

Page 6: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

6

Two Approaches: Pros and Cons

Full Model High prediction accuracy Low coverage

Independent Model Low prediction accuracy High coverage

Page 7: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

7

New Approach: Conditional Probability Hierarchy• P(url | microsoft xbox 360 kinect game) =

PF(url | “microsoft xbox 360 kinect game”) + (1-)PI(url | microsoft, xbox, 360, kinect, game)

Page 8: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

8

New Approach: Conditional Probability Hierarchy• P(url | microsoft xbox 360 kinect game) =

PF(url | “microsoft xbox 360 kinect game”) + (1-)PI(url | microsoft, xbox, 360, kinect, game)

• PI(url | microsoft, xbox, 360, kinect, game) =P(url|microsoft)P(url|xbox)P(url|360)P(url|kinect) P(url|game)

Page 9: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

9

New Approach: Conditional Probability Hierarchy• P(url | microsoft xbox 360 kinect game) =

PF(url | “microsoft xbox 360 kinect game”) + (1-)PI(url | microsoft, xbox, 360, kinect, game)

• PI(url | microsoft, xbox, 360, kinect, game) =P(url|microsoft)P(url|xbox)P(url|360)P(url|kinect) P(url|game)

Page 10: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

10

New Approach: Conditional Probability Hierarchy

microsoft xbox 360 kinect game

P(url|microsoft xbox)P(url|360 kinect)

P(url|360 kinect game)

P(url|microsoft xbox 360 kinect game)

Page 11: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

11

New Approach: Conditional Probability Hierarchy• P(url | microsoft xbox 360 kinect game) =

PF(url | “microsoft xbox 360 kinect game”) + (1-)PI(url | microsoft, xbox, 360, kinect, game)

• PI(url | microsoft, xbox, 360, kinect, game) =f(PF(url | “microsoft xbox”),

(1-)PI(url | microsoft, xbox), PF(url | “360 kinect game”),

(1-)PI(url | 360, kinect, game))

Page 12: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

12

New Approach: Conditional Probability Hierarchy• P(url | microsoft xbox 360 kinect game) =

PF(url | “microsoft xbox 360 kinect game”) + (1-)PI(url | microsoft, xbox, 360, kinect, game)

is directly proportional to the occurrence frequency of “microsoft xbox 360 kinect game”

Page 13: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

13

How to predict? What’s the threshold probability for considering

P(url | microsoft xbox 360 kinect game) as a click???

Page 14: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

14

How to predict? What’s the threshold probability for considering

P(url | microsoft xbox 360 kinect game) as a click??? Not said in paper…

Page 15: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

15

Experimental Corpus

CiteSeer 1,826,817 query-click pairs

Page 16: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

16

Prediction Accuracy

Page 17: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

17

Prediction Coverage

Page 18: 1 Learning User Clicks in Web Search Ding Zhou et al. The Pennsylvania State University IJCAI 2007

18

Accuracy-Coverage Tradeoff