CUBELSI: AN EFFECTIVE AND EFFICIENT METHOD FOR SEARCHING RESOURCES IN SOCIAL TAGGING SYSTEMS

Bin Bi, Sau Dan Lee, Ben Kao, Reynold Cheng

The University of Hong Kong

{bbi, sdlee, kao, ckcheng}@cs.hku.hk

SOCIAL TAGGING SYSTEMS

Tags

2

SEARCH IN SOCIAL TAGGING SYSTEMS

Two Problems:

1.Tag Inconsistency2.A Multitude of Aspects

3

TAG INCONSISTENCY

car? automobile?

car, automobile

car, Benz

car

car, automobileautomobile

Audi

car

4

A MULTITUDE OF ASPECTS

moon,worm moon,Perigee moon,lunar

cherry blossoms,Sakura,cherry

blossom

Nikon,astrophotograph

y,D40 5

SOLUTION

LSI(Latent

Semantic Indexing)

CubeLSI

SVD(Singular

Value Decomposition

)

Tucker Decompositio

n

Taggers

Analyzing semantic relations among tags by taking into account the role of taggers

6

PROPOSED RANKING FRAMEWORK

CubeLSI Algorithm:Input: tag assignmentsOutput: pairwise tag semantic distances

7

CONCEPT DISTILLATION

Tags with pairwise distances

mp3

music

photo

photos

video

movie

photophotos

musicmp3

videomovie

Concepts/Clusters8

PROPOSED RANKING FRAMEWORK

9

BAG-OF-CONCEPTS REPRESENTATION

Distilled Concepts

10

PROPOSED RANKING FRAMEWORK

11

PROPOSED RANKING FRAMEWORK

12

RANKING SEARCH RESULTS

x

y

z

Query

Search results are sorted in descending order of their Cosine similarity scores.

Resource 1

Resource 2

13

PROPOSED RANKING FRAMEWORK

CubeLSI Algorithm:Input: tag assignmentsOutput: pairwise tag semantic distances

14

CUBELSI

Tensor

Second-order Tensor

Third-order Tensor

15

REPRESENTING DATA AS A THIRD-ORDER TENSOR

16

PAIRWISE TAG DISTANCE

Two sources of noise:

1. may not result from user considering tag to be irrelevant to 2.Tagging is a casual and ad-hoc activity

17

TUCKER DECOMPOSITION

Tag

Resource

User

1 2 3Tag

Resource

User

core tensor

original tensor

purified tensor

factor matrices

Purified Tag Distance:

18

SPACE & TIME COSTS

Last.fm dataset (3897 users, 3326 tags, 2849 resources)

36.9 billion entries

11.1 million entries

Computing the Frobenius-norm for EACH tag pair requires 11.1 million subtractions, squaring and additions.

There are a total of 5.5 million tag pairs for 3326 tags !

The amount of computations needed would be prohibitively huge!!!

19

• The new formula depends only on core tensor and factor matrix• There is no need to compute any entries of purified tensor• The relatively low dimensions of and implies much fewer

computations needed

SHORT-CUT TO EVALUATING

impractical

is a matrix that can be readily computed from the core tensor

20

EXPERIMENTAL RESULTS

Dataset statistics#users #tags #resource

s#records

21

SAMPLE TAG CLUSTERS

22

OTHER RANKING METHODS

Freq: Resources are ranked in descending order of # of users who annotate the resource with query tags.

BOW (Bag-of-Words) : Use IR; each resource is a document and each tag is a word.

FolkRank [Hotho et al. 2006]: A modified version of PageRank. It follows the assumption that votes cast by important users with important tags would make the annotated resources important.

23

OTHER RANKING METHODS

LSI: This method projects the third-order tensor onto a 2D tag-resource matrix, and then applies traditional LSI on the tag-resource matrix using SVD.

CubeSim: This method is similar to CubeLSI except that it computes the distance between two tags and directly from the original tensor by

24

RANKING QUALITY

Evaluation Metric Normalized Discounted Cumulative Gain (NDCG) NDCG rewards more heavily to relevant

resources that are top-ranked than those that appear lower down in the list.

where denotes that the metric is evaluated only on the resources that are ranked top in the list, is the relevance level of the resource ranked in the list, and is a normalization factor that is chosen so that the optimal ranking’s NDCG score is 1.

16 users, each

proposing 8 queries

25

RANKING QUALITY (Delicious)

26

RANKING QUALITY (Bibsonomy)

27

RANKING QUALITY (Last.fm)

28

EFFICIENCY

Offline: pre-processing times (hours)

Online: query processing times (seconds)

Storage size:

29

RELATED WORK

Matrix Factorization Our work differs from MF in two ways:

We aim at capturing semantic relations among tags. We deal with a three-dimensional tensor.

Hotho et al. 2006 Our work differs from FolkRank in that our approach

performs offline semantic analysis, which allows online query processing to be efficiently done.

Wu et al. 2006 Our approach is technically different from that work.

Bi et al. 2009 Our approach scales to large social tagging databases,

which the previous work is unable to handle.30

CONCLUSIONS

We introduce a novel tag-based framework for searching resources in social tagging systems.

We study the role of taggers in search quality for social tagging systems.

We propose CubeLSI, which is a 3D extension of LSI, for semantic analysis over the third-order tensor of resources, taggers, and tags.

We present a comprehensive empirical evaluation of CubeLSI against a number of ranking methods on real datasets.

31

THANK YOU!

Bin Bi, Sau Dan Lee, Ben Kao, Reynold Cheng

The University of Hong Kong

{bbi, sdlee, kao, ckcheng}@cs.hku.hk32