Presentation Journée Thématique - Fouille de Grands Graphes

Point of View BasedClustering ofSocio-Semantic NetworksJuan David CRUZ1

Cécile BOTHOREL1

François POULET2JT – FGG2010

Outline

1 IntroductionSocial Networks and Points of ViewSome Previous Work

2 The Point of View of Social Networks

3 Influencing the Community Detection with the Point of ViewPhase 1Phase 2

4 Preliminary Experiments and Results

5 Conclusion

page 2 Cruz,Bothorel,Poulet Clustering by Point of View

MotivationIntroduction I Social Networks and Points of View

� Socio–semantic networks are contains both:• The social graph• Semantic information represented by the features of thevertices and the edges.

� Given this information, how to identify communities derivedfrom the conjoint use of it?

� It is necessary to measure the quality of the partitions foundusing this information from two perspectives:

• The quality of the graph clustering• The quality of the semantic information within thecommunities


Social Networks and Points of ViewIntroduction I Social Networks and Points of View

It is possible to obtain different partitions from different pointsof view


Quality MeasuresIntroduction I Some Previous Work

Type Objective Examples

Similarity

Reduce the distance betweenthe members of the samegroup while the distance

between groups is increased.

Manhattan L1Euclidean L2Chebyshev L∞

Quality

Increase the number of edgeswithin each community whilethe number of edges betweencommunities is reduced. In

general: index (C) = f (C)+g(C)N(G)

[Gae05]

CoverageConductancePerformanceModularity


Graph Clustering AlgorithmsIntroduction I Some Previous Work

Several graph clustering algorithms have been developed, amongothers:� Newman [New01] (Modularity optimization)� Fast unfolding [BGLL08] (Modularity optimization)� Maximal cliques enumeration and kernel generation [DWP+07](Modularity optimization)

� Genetic algorithm for detecting communities in large graphs[LM09] (Fitness function based on modularity)

� Genetic algorithm for detecting overlapped communities[Piz09] (Fitness function based on internal edges vs. outgoingedges)


The Representation of a Point of ViewThe Point of View of Social Networks

� The semantic information can be described by a set of features� The point of view is PoVi = si ×V

Point of ViewNodes Feature 1 Feature 2 . . . Feature fNode 1 1 0 . . . 0Node 2 0 1 . . . 1

......

......

Node n 1 0 . . . 1

� The assignation of features to each node in the network


General ArchitectureInfluencing the Community Detection with the Point of View

� Guide the community detection algorithm according tosemantic information.

� Use of clustering techniques from different domains.


Semantic ClusteringInfluencing the Community Detection with the Point of View I Phase 1


Semantic ClusteringInfluencing the Community Detection with the Point of View I Phase 1

� Clustering of the defined point of view: search nodes withsimilar instances of features.

� Use of Self–Organizing Maps (SOM): non–supervised machinelearning method [Koh97].

� The proximity between the input vector (instance) and theweight vector of the network is measured with the Euclideandistance.

� The SOM algorithm will find some number of groups.


Weights Assignation and CommunityDetection

Influencing the Community Detection with the Point of View I Phase 2


Weights Assignation and CommunityDetection

Influencing the Community Detection with the Point of View I Phase 2

� Given the trained SOM network N and a graph G (V ,E ):� For each e (i , j) ∈ E , the weight will be changed according to:

wij = 1+αd (Nij)δij

where α ≥ 1 is constant parameter, d (Nij), is the distancebetween the node i and the node j in the SOM network andδij = 1 if i , j belong to the same group in the SOM network.

� After the weights are set, a classic graph clustering algorithm(the fast unfolding algorithm [BGLL08]) is used.


Experiments Configuration IPreliminary Experiments and Results

� In each experiment three algorithm were compared:• SOM• Fast unfolding• Our method

� Performed in two levels:• The final modularity: to measure the quality of the partition.• The average intra–cluster Euclidean distance: to measure thequality of the semantic clustering.

� The experiment were executed using a graph of 5389 nodesand 27347 edges extracted from a Twitter data set. The initialmodularity of this graph is −2.5192×10−3


Experiments Configuration IIPreliminary Experiments and Results

� The experiments were performed using two different points ofview.

� Point of View 1:• Composed of 33 features. Each feature represents a time zonefrom the Twitter data set.

• A feature will be set to 1 if the node has at least one friend inthe time zone represented by the feature.

• Distances vary from 0 to√32

� Point of View 2:• Composed of 4 features representing the messaging profile ofeach user.

• The first feature is set to 1 if the user has more friends thanfollowers.


Experiments Configuration IIIPreliminary Experiments and Results

• The next three features indicate the user behavior according tothe number of messages sent: below the mean, between themean plus three standard deviations and, over mean plus threestandard deviations.

• Distances vary from 0 to√3


Case Twitter – Point of View 1Preliminary Experiments and Results

Experiment Final Q Avg. Intracluster DistanceSOM Graph −7.5×10−3 0.3697

Graph based clustering 0.5728 1.8091PoV based Clustering 0.5747 1.1947

� The average intracluster distance found by our proposedmethod is less than the average intracluster distance found bythe graph based algorithm.

� The modularity obtained is very similar: the point of view usesinformation associated with the localization of people’s friends.

� The modularity of the graph from the SOM clustering is notvery different from the modularity of the original graph.

� SOM groups are close to the structure of the non–clusteredgraph.


Case Twitter – Point of View 2Preliminary Experiments and Results

Experiment Final Q Avg. Intracluster DistanceSOM Graph -0.2991 0

Classic clustering 0.5728 0.7100PoV based Clustering 0.6351 0.5507

� The SOM clustered the nodes into six groups, each oneexpressing one of the possible instances described above. Thisexplains the average distance found.

� Creating a graph from the SOM clustering will produce bettersemantic clusters, however, the modularity is worst than theone from the original graph.

� The SOM groups are totally unrelated with the structure ofthe graph.

� Regarding the modularity and the average intracluster distance,the performance of the PoV based algorithm was better.


Conclusion IConclusion

� The classic community detection algorithms do not take intoaccount the semantic information to influence the clusteringprocess.

� Changing the weights according to the results of the semanticclustering, the semantic information is included into thecommunity detection process.

� The two types of informations are merged to find andvisualize a social network from a selected point of view.


Conclusion IIConclusion

� Outlook• Make tests over the obtained partitions: rand index, robustnesstests...

• Study the case of overlapping communities.• Include the features of the edges into the point of viewgeneration.

• Development of a visualization algorithm for representing thePoV and the transition between two points of view.


Cruz,Bothorel,Poulet Clustering by Point of View

nav

Thanks for your attentionAppendix I Appendix

Questions?


For Further Reading IAppendix I Appendix I For Further Reading

Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte,and Etienne Lefebvre.Fast unfolding of communities in large networks.Journal of Statistical Mechanics: Theory and Experiment,2008(10):P10008 (12pp), 2008.

Nan Du, Bin Wu, Xin Pei, Bai Wang, and Liutong Xu.Community detection in large-scale social networks.In WebKDD/SNA-KDD ’07: Proceedings of the 9th WebKDDand 1st SNA-KDD 2007 workshop on Web mining and socialnetwork analysis, pages 16–25, New York, NY, USA, 2007.ACM.


For Further Reading IIAppendix I Appendix I For Further Reading

Marco Gaetler.Network Analysis: Methodological Foundations, chapterClustering, pages 178 – 215.Springer Berlin / Heidelberg, 2005.

Teuvo Kohonen.Self-Organizing Maps.Springer, 1997.


For Further Reading IIIAppendix I Appendix I For Further Reading

Marek Lipczak and Evangelos Milios.Agglomerative genetic algorithm for clustering in socialnetworks.In GECCO ’09: Proceedings of the 11th Annual conference onGenetic and evolutionary computation, pages 1243–1250, NewYork, NY, USA, 2009. ACM.

M. E. Newman.Scientific collaboration networks. ii. shortest paths, weightednetworks, and centrality.Phys Rev E Stat Nonlin Soft Matter Phys, 64, July 2001.


For Further Reading IVAppendix I Appendix I For Further Reading

Clara Pizzuti.Overlapped community detection in complex networks.In GECCO ’09: Proceedings of the 11th Annual conference onGenetic and evolutionary computation, pages 859–866, NewYork, NY, USA, 2009. ACM.


Quality Measures SummaryAppendix I Appendix I For Further Reading

Summary of quality measures for graph clustering.


Presentation Journée Thématique - Fouille de Grands Graphes

Technology

Transcript of Presentation Journée Thématique - Fouille de Grands Graphes