+ All Categories
Home > Documents > Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John...

Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John...

Date post: 29-Mar-2015
Category:
Upload: hadley-fender
View: 213 times
Download: 1 times
Share this document with a friend
Popular Tags:
33
Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research
Transcript
Page 1: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

Tagommenders:Connecting Users to Items

Through Tags

Shilad SenMacalester College

Jesse Vig, John RiedlGroupLens Research

Page 2: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 3: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 4: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 5: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 6: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 7: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

Tagommenders

1. Analyze user interactions to infer liking (preferences) for tag concepts.

2. Recommend items related to tag concepts liked by users.

Page 8: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 9: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

9 of 32

Tagommender Goals

• Recommend items using just tags. (Delicious)

• Improve item recommendations with ratings by by using tags. (LibraryThing / Amazon)• accuracy• flexibility• explainability (Vig, IUI 2009).

Page 10: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

10 of 32

Tagommender Flow Chart

WALL-E

animation robots pixar

tag preference inference

tag-based recommendation

Page 11: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

11 of 32

MovieLens Tagging

• Tagging introduced in 2006• 15,000 distinct tags• 127,000 tag applications:

<user, tag, movie>• 4000 users applied >= 1

tag• 7700 movies with >= 1 tag

app

Page 12: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

12 of 32

Outline

• Tag preference inference• Item recommendation• Auto-tagging and wrap-up

Page 13: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

13 of 32

Outline

• Tag preference inference• Item recommendation• Auto-tagging and wrap-up

Page 14: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

14 of 32

Step 1: Tag Preference Inference

animationrobotspixar?

Infer a user’s interest in tags from:• tags user applied• tags user searched for• user’s clicks on movie hyperlinks• user’s movie ratings

Page 15: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

15 of 32

118,017 ratings

by 995 users

Page 16: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

16 of 32

Preferences for Tags Searched / Applied

average pref applied searched for0

1

2

3

4

5

Av

era

ge

Ta

g P

refe

ren

ce

Page 17: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

17 of 32

Movie-rating algorithm

cars

Page 18: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

18 of 32

Movie-Rating Algorithm

cars

4 of 12 1 of 369 of 380.8 0.10.9

Page 19: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

19 of 32

Bayes-Rating Algorithm

Generative Model:• Expressive probabilistic processes.• Model movie ratings.• Separate model for every user, tag.

Page 20: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

20 of 32

Jill’s Ratings for animated Movies

0.0 1.0 2.0 3.0 4.0 5.00

0.2

0.4

0.6

0.8

1

animation

Star Rating for Movies With Tag

Fre

qu

en

cy

N(μ=3.8,σ=0.7)

Bayes-Rating Algorithm

Page 21: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

21 of 32

all possible normal dists for ratings for animated movies

WALL-E

not tt = animation

p(t | WALL-E) 1.0 - p(t | WALL-E)

N(μu,t,σu,t) N(μu,σu)

0 1 2 3 4 50

0.5

1

N(μ=2.0,σ=1.0)N(μ=4.0,σ=0.5)

Bayes-Rating Algorithm

Page 22: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

22 of 32

All movies m rated by Jill tagged with

animation

not tt = animation

Toy Story

WALL-E

Shrek

0 1 2 3 4 50

0.5

1

all possible normal dists for ratings for animated movies

Bayes-Rating Algorithm

Page 23: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 24: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

24 of 32

Outline

• Tag preference inference• Item recommendation• Auto-tagging and wrap-up

Page 25: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

25 of 32

Tagommender Flow Chart

WALL-E

animation robots pixar

tag preference inference

tag-based recommendation

Page 26: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

26 of 32

Step #2: Tag-Based Recommendation

• Standard machine learning problem• With / without ratings• Six standard recommender baselines• Evaluate predictive performance

Page 27: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.
Page 28: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

28 of 32

Outline

• Tag preference inference• Item recommendation• Auto-tagging and wrap-up

Page 29: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

29 of 32

Inferred pref for girlie movie:

Rating for “Runaway Bride”

Alice

Bob

Mike

(other users) …. …

cosine similarity = 0.45

Using Tag Preferences for Tag Inference

Page 30: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

30 of 32

Top 10 Inferred Tags Not Already Applied

movie tag cosine sim

Pearl Harbor (2001) disaster 0.47

Runaway Bride (1999) girlie movie 0.45

Beauty and the Beast (1991) talking animals 0.42

Armageddon (1998) will smith 0.41

Cinderella (1950) cartoon 0.40

Inconvenient Truth (2006) documentary 0.40

The Little Mermaid (1989) musical 0.40

Gone in 60 Seconds (2000) exciting 0.39

My Best Friend’s Wedding (1997) chick flick 0.39

Billy Madison (1995) very funny 0.39

Page 31: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

31 of 32

Summary of Tagommenders

• Tag preference inference:• Systems can infer user preferences for tags.• Item ratings help tag pref inference.• Tag prefs can be used for auto-tagging.

• Tagommenders outperform traditional recommenders:• Without ratings: moderate edge (10%).• With ratings: slight edge (2%).

Page 32: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

32 of 32

Future Work

1. Alternative modalities for tags.

2. Quality vs. preference.

Thank You!

3. GroupLens.

4. MovieLens users.

5. NSF grants IS 03-24851 and IIS 05-34420.

6. Macalester College.

Page 33: Tagommenders: Connecting Users to Items Through Tags Shilad Sen Macalester College Jesse Vig, John Riedl GroupLens Research.

Shilad Sen

[email protected]

(photo by flickr user SantiMB)


Recommended