A Graph-based Clustering Scheme for Identifying Related Tags in Folksonomies

A Graph-based Clustering Scheme for Identifying
Related Tags in Folksonomies
Symeon Papadopoulos, Yiannis Kompatsiaris, Athena Vakali

Bilbao, Spain
30 Aug – 3 Sep

CERTH ITI AUTH

overview

• tag clustering / intro

• existing solutions - limitations

• hybrid graph clustering (HGC)
– core set detection
– (μ,ε)-space exploration
– core set expansion

• evaluation

• conclusions

Symeon Papadopoulos (CERTH-ITI, AUTH) 2

overview




• evaluation

• conclusions


tag clustering
• starting point:
folksonomy, i.e. annotation scheme produced by the set of users,
resources, tags of a social tagging system, e.g. delicious, flickr,
BibSonomy (Mika, 2005)

• observation I:
folksonomies  a direct encoding of the views of users on how content
items should be organized through a flexible annotation scheme

• observation II:
tags used to describe the same resources  tags related to each other
(meaningful semantic association)

Mika, P.: Ontologies are us: A unified model of social networks and semantics. ISWC 2005, LNCS 3729, 522-536,
Springer-Verlag (2005)


why is tag clustering useful?
• information exploration and navigation (Begelman et al., 2006;
Simpson, 2008)
• automatic content annotation (Brooks, 2006)
• user profiling (Gemmell, 2008)
• content clustering (Giannakidou, 2008)
• tag sense disambiguation (Au Yeung, 2009)
Begelman, G., Keller, P., Smadja, F.: Automated Tag Clustering: Improving search and exploration in the tag space. Online article:
http://www.pui.ch/phred/automated_tag_clustering (2006)
Simpson, E.: Clustering Tags in Enterprise and Web Folksonomies. Technical Report HPL-2008-18 (2008)
Au Yeung, C. M., Gibbins, N., Shadbolt., N.: Contextualising Tags in Collaborative Tagging Systems. Proceedings of 20th ACM
Conference on Hypertext and Hypermedia, pages 251-260, Turin, Italy, 29 June - 1 July, ACM (2009)
Giannakidou, E., Koutsonikola, V. A., Vakali, A., Kompatsiaris, Y.: Co-Clustering Tags and Social Data Sources. Proceedings of WAIM
2008: 9th International Conference on Web-Age Information Management. IEEE, 317-324 (2008)
Brooks, C. H., Montanez, N.: Improved annotation of the blogosphere via autotagging and hierarchical clustering.
Proceedings of WWW '06: 15th international Conference on World Wide Web. ACM, New York, NY, 625-632 (2006)
Gemmell, J., Shepitsen A., Mobasher B., Burke, R.: Personalizing Navigation in Folksonomies Using Hierarchical Tag
Clustering. Data Warehousing and Knowledge Discovery 5182, 196-205 (2008)


overview




• evaluation

• conclusions


existing solutions (i) :: conventional clustering

• conventional clustering schemes
represent tags in some feature space and employ
standard clustering method, e.g.:
• k-means (Giannakidou et al., 2008)
• hierarchical agglomerative clustering (HAC)
(Brooks et al., 2006; Gemmell et al., 2008)

Giannakidou, E., Koutsonikola, V. A., Vakali, A., Kompatsiaris, Y.: Co-Clustering Tags and Social Data Sources. Proceedings
of WAIM 2008: 9th International Conference on Web-Age Information Management. IEEE, 317-324 (2008)
Brooks, C. H., Montanez, N.: Improved annotation of the blogosphere via autotagging and hierarchical clustering.
Proceedings of WWW '06: 15th international Conference on World Wide Web. ACM, New York, NY, 625-632 (2006)
Gemmell, J., Shepitsen A., Mobasher B., Burke, R.: Personalizing Navigation in Folksonomies Using Hierarchical Tag
Clustering. Data Warehousing and Knowledge Discovery 5182, 196-205 (2008)


existing solutions (i) :: conventional clustering

• problems with conventional clustering

• needs number of clusters to be defined: very hard to even
estimate it in large-scale tagging systems

• not easily scalable:
– k-means (Lloyd’s): O(I  C  n  D)
– HAC: O(n2  logn)
n: number of tags, I: number of iterations, C: number of clusters, D:
number of dimensions
HAC is hardly applicable since it requires n2 memory for storing the
dissimilarity matrix


existing solutions (ii) :: community detection

• use of community detection methods on tag graphs (derived
from folksonomies) to find groups of tags that are more
densely connected to each other than to the rest of the graph
• community detection methods largely address shortcomings
of conventional clustering (Begelman et al., 2006; Simpson,
2008; Au Yeung et al., 2009) schemes
– efficient: complexity O(n  logn)
– do not require number of communities to be provided as input
(typically use modularity maximization)

Begelman, G., Keller, P., Smadja, F.: Automated Tag Clustering: Improving search and exploration in the tag space.
Online article: http://www.pui.ch/phred/automated_tag_clustering (2006)

Au Yeung, C. M., Gibbins, N., Shadbolt., N.: Contextualising Tags in Collaborative Tagging Systems. Proceedings of
20th ACM Conference on Hypertext and Hypermedia, pages 251-260, Turin, Italy, 29 June - 1 July, ACM (2009)


existing solutions (ii) :: community detection

• existing community detection schemes also suffer
from problems
– modularity maximization typically leads to highly skewed
cluster size distribution (Simpson, 2008):
few gigantic clusters and numerous small ones 
gigantic clusters (representing even half the number of
objects) are not useful for IR
– not possible to leave noisy objects out of cluster structure
– not possible to have overlap among clusters (which is
useful in the context of tag clustering)


overview




• evaluation

• conclusions


hybrid graph clustering
• our solution is based on a structure-connected community
detection approach (Xu et al., 2007) that is based on the
concept of structural similarity and (μ,ε)-cores:
– nodes on the graph are structurally similar when they have many
neighbors in common
– a (μ,ε)-core is a node that has at least μ neighboring nodes with which
it has structural similarity at least ε

• extended in two ways:
– parameter space exploration  raises the need for setting
parameters
– core community expansion  permits overlap among communities

Xu, X., Yuruk, N., Feng, Z., Schweiger, T. A.: SCAN: A Structural Clustering Algorithm for Networks. Proceedings of KDD
'07: 13th international Conference on Knowledge Discovery and Data Mining, ACM, New York, NY, 824-833 (2007)


hybrid graph clustering

• hybrid scheme:
– (μ,ε)-core identification and structure connected
cluster extraction (original approach)
– (μ,ε)-parameter space exploration
makes scheme completely parameter-free
– cluster expansion
increases coverage, permits overlap among clusters


structure connected cluster extraction
• structural similarity between nodes u, w on a graph G = {V, E}:

• ε-neighborhood:
• (μ,ε)-core:
• direct structure reachability of w w.r.t. to core u:

• cluster extraction (Xu et al., 2007):
starting from a (μ,ε)-core node grow the cluster to contain all nodes that
are directly structure reachable to it or reachable through a chain of
nodes that are directly structure reachable to each other
Xu, X., Yuruk, N., Feng, Z., Schweiger, T. A.: SCAN: A Structural Clustering Algorithm for Networks. Proceedings of KDD
'07: 13th international Conference on Knowledge Discovery and Data Mining, ACM, New York, NY, 824-833 (2007)


structure connected cluster extraction

• edge labels denote structural similarity values between nodes
• blue nodes are (μ, ε)-cores for μ = 5 and ε = 0.65
• gray nodes are directly structure reachable from (μ, ε)-cores
• the rest of nodes are left out of the cluster structure


parameter space exploration

• original approach needs parameter setting that is
troublesome for complex datasets
• parameter interpretation:
– μ: a high value for μ will lead to fewer and larger clusters,
i.e. only nodes with degree of at least μ will be considered
to be cores
– ε: a high value for ε will make the cluster extraction
process stricter, i.e. less nodes will be assigned to clusters
• in fact, a single (μ,ε) parameter pair is unlikely to
discover all interesting clusters


parameter space exploration
• search for clusters at multiple parameter pairs
• identify the highest quality clusters (high μ, high ε), then
proceed to less profound clusters
• exclude nodes that have
already been assigned
to a cluster from being
re-assigned  makes
process faster
• log-sampling along μ
axis for faster
exploration


cluster expansion
• the original structure connected approach may be too strict and thus
leave too many nodes out of the clustering structure
• an expansion process attempts to mitigate this weakness
• for each extracted core cluster, a local expansion process is conducted
that attaches neighboring nodes
• the expansion is based on a simple greedy maximization of a local cluster
density measure called subgraph modularity (Luo et al., 2006):

• nodes with very high degree (belonging to the top 10 percentile of the
degree distribution) are not considered in this process in order to make
the expansion process more efficient

Luo, F., Wang, J. Z., Promislow, E.: Exploring Local Community Structures in Large Networks. Proceedings of the
2006 IEEE/WIC/ACM international Conference on Web Intelligence. IEEE Computer Society, 233-239 (2006)


cluster expansion

(a) before attaching node 11 (b) after attaching node 11
M(S) = 1.429 M(S) = 2.4


overview




• evaluation

• conclusions


evaluation :: overview
goal: compare the quality of tag clusters produced by our method (HGC) with
the one produced by state-of-the-art, namely:
(a) modularity-maximization method by Clauset et al., 2004 (CNM)
(b) original structure connected graph clustering by Xu et al., 2007 (SCAN)

two kinds of evaluation:

• direct small-scale evaluation
subjective assessment of the produced tag clusters by eyeballing to see
whether tags belonging to the same cluster are related

• indirect large-scale evaluation
evaluate how useful the produced cluster structure is for some IR task, namely
tag recommendation  if tag clusters are good, performance of tag
recommendation based on them will be good as well


evaluation :: datasets

• three different folksonomy datasets of various sizes:

• resulting tag graphs (large component)

average degree

average clustering coefficient

direct evaluation (i)
examples of unrelated tags placed in the same gigantic community by CNM


direct evaluation (ii)
examples of interesting HGC communities


indirect evaluation :: setup (i)

• process
– simple tag recommender based on tag clusters:
• input tag
• find containing community
• recommend most frequent tags of the same community
naïve technique, but fair for comparing the effectiveness of the used
tag cluster structure
– the three competing tag cluster structures (CNM, SCAN, HGC) were
used by the recommender
– historic tagging data were used as ground truth
• for each user one tag was used as input and the rest were considered as
the “correct” output
• very frequent tags (top 5%) were left out of this process in order not to
allow trivial (very generic) recommendations to mask the actual results

indirect evaluation :: setup (ii)

• measures
– RTP: number of correct recommendations per
recommender instance
– UTP: number of unique correct recommendations
– P: precision, i.e. ratio of correct recommendations over
total recommendation per recommender instance
– R: recall, i.e. ratio of correct recommendations of a
recommender instance over all correct tags according to
ground truth
– F-measure
– P@1, P@5: Precision in the top-1/top-5 recommendations


indirect evaluation :: results

• for SCAN, we used the (μ,ε)-pair that yielded the highest F-measure
• both SCAN and HGC perform considerably better than CNM
• HGC results in more unique correct recommendations and higher recall
• the cluster expansion step was responsible for the largest increase in recall and
corresponding drop in precision

conclusion: given the task and the evaluation setup, we would prefer HGC,
since: (a) it is parameter free, (b) it leads to more correct recommendations


overview




• evaluation

• conclusions


conclusions
contributions:
• efficient tag clustering scheme that addresses several shortcomings of previous
approaches
– no need for setting the number of clusters
– no gigantic communities
– noisy tags left out of cluster structure
– possibility for overlap among communities

caveats:
• despite being efficient compared to conventional clustering schemes, the method
is still much slower than the original SCAN (Xu et al., 2007)
• the fact that previously assigned nodes are not taken into account when a new
(μ,ε) pair is explored, distorts the actual clustering results

future work:
• investigate means of making parameter exploration more efficient
• evaluate the value of permitting overlap among communities


questions


A Graph-based Clustering Scheme for Identifying Related Tags in Folksonomies

More Related Content

Similar to A Graph-based Clustering Scheme for Identifying Related Tags in Folksonomies (20)

More from Symeon Papadopoulos (20)

Recently uploaded (20)

A Graph-based Clustering Scheme for Identifying Related Tags in Folksonomies