Structure of media attention

The structure of media attention

V.A. Traag, R. Reinanda, J. Hicks, G. Van Klinken

KITLV, Leiden, the Netherlandse-Humanities, KNAW, Amsterdam, the Netherlands

September 30, 2014

eRoyal Netherlands Academy of Arts and SciencesHumanities

Background

Research focus

• Study elite (network) behaviour.

• Relation with political developments.

• Data: newspaper articles. How can we use them?

Data

• Current corpus: Joyo/Indonesian News Service, 2004–2012.

• Contains about 140 263 articles.

Network

Building the network

1 Detect names automatically .I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

2 Disambiguate names.I Susilo Bambang Yudhoyono or Dr. Yudhoyono , etc. . .

3 Co-occurrence in sentence (record frequency).I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

K

B Y

1

1

1

Network

Building the network

1 Detect names automatically .I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

2 Disambiguate names.I Susilo Bambang Yudhoyono or Dr. Yudhoyono , etc. . .

3 Co-occurrence in sentence (record frequency).I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

K

B Y

1

1

1

Network

Building the network

1 Detect names automatically .I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

2 Disambiguate names.I Susilo Bambang Yudhoyono or Dr. Yudhoyono , etc. . .

3 Co-occurrence in sentence (record frequency).I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

K

B Y

1

1

1

Network

Building the network

1 Detect names automatically .I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

2 Disambiguate names.I Susilo Bambang Yudhoyono or Dr. Yudhoyono , etc. . .

3 Co-occurrence in sentence (record frequency).I “ Budhisantoso would ask Kalla to team up with Yudhoyono .”

K

B Y

1

1

1

Strength

100 101 102 103

100

101

Degree

Average

weigh

t

Joyo

100 101 102 103 104

Degree

NYT

Data

Hubs co-occur more frequently.

Strength

100 101 102 103

100

101

Degree

Average

weigh

t

Joyo

100 101 102 103 104

Degree

NYT

Data Bipartite

Hubs co-occur more frequently.

Clustering

100 101 102 10310−3

10−2

10−1

100

Degree

Clustering

Joyo

100 101 102 103 104

Degree

NYT

Data

Hubs tend to cluster less.

Clustering

100 101 102 10310−3

10−2

10−1

100

Degree

Clustering

Joyo

100 101 102 103 104

Degree

NYT

Data Bipartite

Hubs tend to cluster less.

Clustering

100 101 102 103

10−1

100

Degree

Weigh

tedClustering

Joyo

100 101 102 103 104

Degree

NYT

Data

Hubs tend to cluster less (also weighted).

Clustering

100 101 102 103

10−1

100

Degree

Weigh

tedClustering

Joyo

100 101 102 103 104

Degree

NYT

Data Bipartite

Hubs tend to cluster less (also weighted).

Neighbour degree

100 101 102 103101

102

103

Degree

Neigh

bou

rDegree

Joyo

100 101 102 103 104

Degree

NYT

Data

Hubs tend to connect to low degree nodes.

Neighbour degree

100 101 102 103101

102

103

Degree

Neigh

bou

rDegree

Joyo

100 101 102 103 104

Degree

NYT

Data Bipartite

Hubs tend to connect to low degree nodes.

Weighted Neighbour degree

100 101 102 103

102

103

Degree

Weigh

tedNeigh

bou

rDegree

Joyo

100 101 102 103 104

Degree

NYT

Data

But hubs connect much stronger to other hubs.

Weighted Neighbour degree

100 101 102 103

102

103

Degree

Weigh

tedNeigh

bou

rDegree

Joyo

100 101 102 103 104

Degree

NYT

Data Bipartite

But hubs connect much stronger to other hubs.

Predict weight

100 101 102 103 104

100

101

102

103

104

Weight

PredictedWeigh

t

Joyo

100 101 102 103 104

Weight

NYT

Data

wij ∼ Jγij exp(α(si sj)β)

Predict weight

100 101 102 103 104

100

101

102

103

104

Weight

PredictedWeigh

t

Joyo

100 101 102 103 104

Weight

NYT

Data Bipartite

wij ∼ Jγij exp(α(si sj)β)

Core-periphery

Summary Results

• Hubs attract much more weight.

• Most of the weight between hubs.

• Low degree node connect to hubs.

• Low degree nodes cluster locally.

Consistent with core-periphery structure. But, seems also presentin bipartite randomisation. Largest deviations, empirically:

• Degree is lower, average weight is higher.

• Weighted neighbour degree increases.

Model

Simple model to overcome deviations:

1 Create empty sentence

2 Add certain number of nodes

1 Either random node (with PA)2 Or random neighbour (with PA)

Probability (ki + 1)−β .

3 Repeat

Degree & Weight

Empirical Bipartite Model

JoyoAvg. Degree 12.4 22.1 12.2Avg. Weight 2.9 1.2 2.8

NYTAvg. Degree 22.3 45.2 22.6Avg. Weight 2.01 1.11 1.31

Strength

100 101 102 103

100

101

Degree

Average

weigh

t

Joyo

100 101 102 103 104

Degree

NYT

Data Model

Weight increases more in the model.

Weighted neigbhour degree

100 101 102 103

102

103

Degree

Weigh

tedNeigh

bou

rDegree

Joyo

100 101 102 103 104

Degree

NYT

Data Bipartite

Weighted neighbour degree increases in the model.

Conclusions

Results:

• Network looks like core-periphery.

• Probably due to bipartite structure.

• But also to repetitive interaction.

Further research:

• Basis for comparing elite networks.

• Compare networks across time and space.

• Dynamical, temporal aspects.

Thank you! Questions?

Presentation: SlideSharePaper: arXiv:1409.1744

Dynamics of network: arXiv:1409.2973

http://www.traag.net • @vtraag