Increasing Diversity Through Furthest Neighbor-Based Recommendation

Post on 13-Jul-2015

576 views 0 download

transcript

Increasing Diversity Through Furthest

Neighbor-Based Recommendation

Alan Said, Benjamin Kille, Brijnesh J. Jain, Sahin Albayrak

1

Agenda

Problem

Approach: k Furthest Neighbor

Experimental settings

Results

Conclusions

Discussion

27.02.2012 Information Retrieval & Machine

Learning

2

Problem: Missing diversity

27.02.2012 Information Retrieval & Machine

Learning

3

• Accurate recommendations

• However, all items appear

similar

Problem: Missing diversity

27.02.2012 Information Retrieval & Machine

Learning

4

Intersection:

Actors

Intersection:

Plot

Harry Potter and the Chamber of Secrets 49,0% 32,3%

Harry Potter and the Prisoner of Azkaban 52,9% 26,7%

Harry Potter and the Goblet of Fire 39,2% 29,2%

Harry Potter and the Order of the Phoenix 49,0% 16,8%

Harry Potter and the Half-Blood Prince 41,2% 15,5%

Harry Potter and the Deathly Hallows: Part 1 43,1% 16,8%

Harry Potter and the Deathly Hallows: Part 2 49,0% 22,4%

Source: imdb.com

Desired features of Recommendations

Reflect a user‘s preferences

Correct ranking

Novelty

Serendipity

Diversity

27.02.2012 CC IRML Folie 5

Idea: Combining orthogonal

recommendation

k Furthest Neighbor

27.02.2012 Information Retrieval & Machine

Learning

6

dislike

like

k Furthest Neighbor

27.02.2012 CC IRML Folie 7

Experimental settings

Data set: randomly sampled 1 million ratings out of

MovieLens (1M100k)

excluded: 100 most popular movies (rating frequency)

excluded: users with < 40 ratings

44,214 users; 9,432 movies

Evaluation:

precision @N

recall @N

overlap

N ϵ {5; 10; 25; 50; 100; 200}

27.02.2012 Information Retrieval & Machine

Learning

8

Approaches:

– kNN Pearson

– kNN cosine

– kFN Pearson

– kFN cosine

Results I

Precision @ N

27.02.2012 Information Retrieval & Machine

Learning

9

N 5 10 25 50 100 200

Pearson Similarity 0,0007 0,0110 0,0170 0,0280 0,0410 0,0900

Cosine Similarity 0,0050 0,0070 0,0160 0,0270 0,0570 0,0000

Results II

Recall @ N

27.02.2012 Information Retrieval & Machine

Learning

10

N 5 10 25 50 100 200

Pearson Similarity 0,0080 0,0130 0,0210 0,2300 0,0140 0,0100

Cosine Similarity 0,0020 0,0060 0,0070 0,0060 0,0050 0,0040

Results III

Overlap

27.02.2012 Information Retrieval & Machine

Learning

11

Conclusion

„The enemy of my enemy is my friend“ seems to hold in

the context of recommender systems

kFN achieved worse precision

kFN provided higher recall with N > 50

kFN did provide orthogonal recommendations

27.02.2012 Information Retrieval & Machine

Learning

12

27.02.2012 CC IRML Folie 13

Thanks for your attention!!!

http://recsyswiki.com

27.02.2012 CC IRML Folie 14

Contact

Benjamin Kille

Researcher of Competence Center

Information Retrieval

& Machine Learning

+49 (0) 30 / 314 – 74 128

+49 (0) 30 / 314 – 74 003

benjamin.kille@dai-labor.de

Discussion

How to optimize the approach?

Are there other ways to introcude more diverse

recommendations?

How to evaluate diversity in the context of recommender

system?

27.02.2012 CC IRML Folie 15