Information Retrieval in Folksonomies Nikos Sarkas Social Information Systems Seminar DCS,...

Post on 12-Jan-2016

214 views 0 download

Tags:

Transcript of Information Retrieval in Folksonomies Nikos Sarkas Social Information Systems Seminar DCS,...

Information Retrieval in Folksonomies

Nikos Sarkas

Social Information Systems Seminar

DCS, University of Toronto, Winter 2007

Social Resource Sharing

The del.icio.us paradigm. Users store links to web pages of interest along

with arbitrary, user-specified tags in a server. The model is independent of the resource

being shared. Music (Last.fm) Photos (Flickr) Publications (CiteULike) …

Folksonomies

Folk+taxonomy. Taxonomies are rigid, carefully engineered

structures. Folksonomies are flexible, time-variant

structures that result from the converging use of the same vocabulary.

Interesting Problems

A wealth of interest problems in this setting: Search result ranking Personalization Recommendation Trend detection Community extraction …

Keyword Search

Result ranking is currently naïve. Resources associated with tags matching the

keywords are returned in reverse chronological order.

TF/IDF not useful in this context. What about PageRank™?

PageRank Algorithm

Let be a collection of web pages. Then

Many alternatives in interpreting the

PageRank of a web page. Iterative computation

1,..., nP P

( )

( )( )

( )j i

ji

P M P j

PR PPR P

L P

1 (1 )t tw dAw d p

Formalism

Entities of a Folksonomy Users U Tags T Resources R Assignments Y

Representation Tripartite undirected hypergraph G=(V,E), V=UUTUR, E={ (u,t,r) | (u,t,r) in Y }

Adapted PageRank

Flatten the Folksonomy graph.

Apply PageRank. A resource tagged with

important tags by important users becomes important. Symmetrically for tags and users.

2 1

1

11

U1

U2 T2

T1 R1

R2

U1

U2

T1

T2

R1

R2

12

1

1 1

Adapted PageRank

Important! The flat Folksonomy graph is undirected.

Part of the weight that goes through an edge at time t, will flow back at time t+1.

Results are similar to an edge degree ranking. They are identical for d=1.

FolkRank

Topic specific ranking in Folksonomies. A topic is defined through preference vector A topic can be defined through tags,

resources or users. Let be the Adapted PageRank vector for

d=1. Let be the Adapted PageRank vector for

d<1 and a specified preference vector. The FolkRank vector is .

p

0w

1w

1 0:w w w

Results

Adapted PageRank, d=1

Results

Adapted PageRank vs FolkRank

Extensions

Resource recommendation. Similar tag suggestion. User introduction. Trend detection.