Archives par étiquette : Network analysis

Un réseau géomédiatique à partir du Corpus Géomédia

Peut-on obtenir une vue cohérente du monde géopolitique à partir des flux RSS récoltés dans le cadre de l’ANR Géomedia ? Un travail, que nous avons présenté à la conférence ECTQG 2015, montre que oui. Nous avons construit un réseau de co-citations géomédiatiques en utilisant un extrait des flux récoltés en 2014. Le principe est simple : nous établissons une liste d’entités géographiques (comme des pays) et, pour deux entités données, nous regardons combien de fois elles ont été citées ensemble dans la base RSS. La combinaison de ces éléments est un objet appelé graphe.

Ce type d’objet inclut beaucoup d’informations mais est généralement trop dense pour être analysé « à la main ». Une stratégie classique dans ce cadre est de rechercher des blocs d’entités très connectées entre elles et faiblement connectées avec le reste du réseau (on parle de « détection de communautés » dans les réseaux complexes). Le but est de simplifier le réseau en utilisant ces blocs. On peut également chercher à savoir si ces groupes ont du sens d’un point de vie thématique et/ou géographique.

 consensus_graph_component2

Nous avons utilisé différentes méthodes pour rechercher ces blocs et comparer les résultats. Bien que ces derniers varient, on trouve une cohérence d’un point de vue thématiques dans les blocs détectés. Pour rendre compte des résultats et offrir une image utile du réseau, nous avons filtré les liens entre entités géographiques n’étant pas à l’intérieur d’un bloc quelque soit l’algorithme utilisé (voir image ci-dessus).

Le Corpus Géomédia pourrait ainsi être utilisé pour étudier l’évolution de la représentation géopolitique du monde selon différents critères, par exemple en fonction de l’origine géographique des médias ou selon différentes époques. Vous trouverez les liens vers les jeux de données ci-dessous, n’hésitez pas à les utiliser !

Plus de détails : F. Queyroi, L. Beauguitte et H. Pecout « RSS Flows, World Structure & Community Detection », ECTQG 2015

Papier : queyroi2015_communities_geomedia

Dataset : ectqg2015_cocitations_dataset

Do International News Reflect World Structure? A Network Approach

Paper from Laurent Beauguitte (GIS CIST), Marta Severo (Geriico, Université Lille 3) and Hugues Pecout (GIS CIST) at the 1st European Conference on Social Networks, Barcelona, July 2014

The presentation made in Barcelona had one main objective: to compare different types of newspapers from different countries and to launch network analysis on states co-occurrences in order to see if it could provide valuable information regarding world structure. At the beginning, we had in mind the following theoretical model (Figure below): raw number of state occurrence would give a proxy of hierarchy among nations; state co-occurrences would allow to distinguish two main types of relations: cooperation and conflict.

model

As our approach was exploratory, we began with a small sample: two French and two Australian newspapers and, in each country, a reference newspaper (Le Figaro in France and The Australian… in Australia) and a more popular one (Le Parisien in France and The Daily Telegraph in Australia). It must be noted that The Daily Telegraph is a tabloid while Le Parisien is a popular newspaper and they are not strictly comparable. We collected for these four newspapers RSS flows about international news from January 1st 2014 until February 28th 2014.

We had in total the following number of items per newspaper: The Australian, 1160; The Daily Telegraph, 1103; Le Figaro, 608 and Le Parisien 643 items. A R program tagged all items automatically and was able to gather state occurrence and co-occurrence present in each item. Consider for instance the following item (which was not in our corpus): Abe’s war shrine visit stokes tensions ; JAPANESE PM Shinzo Abe has visited a controversial war shrine in a move that has further inflamed already tense relations with China and South Korea. 26/12/13 08:37. Our tag dictionary recognizes the following terms and add a column with the corresponding state: Abe, JAPANESE, China and South Korea. The item would then give three co-occurrences (or, in network analysis vocabulary, 3 links): Japan-China, Japan-South Korea and China-South Korea.

A first result concerns the great polarisation on a few number of states: depending on newspapers, from 5 to 7 states count for 50% of all states occurrences. The following table gives the top 5 per journal and the percentage of all state occurrence in the newspaper during the whole period. If polarization is greater in Australia for the first rank, the dominant national character of international news in France must be underlined: France is the most quoted state and if Central African Republic is on the third position for both Le Figaro and Le Parisien, it can clearly be explained by the fact that French army intervened in the country at that period…

The Australian The Daily Telegraph Le Figaro Le Parisien
USA     19 USA        24 France 13 France 17.4
Syria    7.5 UK          9.7 USA 11.47 USA 11.48
UK       6.61 Australia 6.2 Centrafrique 6.43 Centrafrique 7.16
France  5.06 Russia     4 Ukraine 6.43 Ukraine 6.62
Russia  4.77 China      3.75 UK 5.03 Russia 6.48

If the number of occurrences is able to reveal main powers and hot spots regarding international news, co-occurrences are much harder to interpret. In this example above, the links Japan-China and Japan-South Korea were conflictual but what about the China-South Korea link? And the vast majority of items is much harder to interpret and to code: when a tourist from a country A gets trouble in a country B, can it seriously be considered as a conflict between two states?… We made several tests on different corpus and in each case, the majority of co-occurrences were coded « others » and many differences of interpretation between coders appeared. So we had to give up our theoretical model: if the world system involves the co-presence of cooperative and conflictual relations between actors, RSS items do not allow to perceive them.

Nevertheless, a promising and unexpected result appears. We get for our four newspapers the list of states quoted and the number of quotations. If we launch a rank Spearman correlation on these numbers, we get the following table:

The Daily Tel. Le Figaro Le Parisien
The Australian 0.90 0.69 0.59
The Daily Tel. 0.61 0.57
Le Figaro 0.92

What does it mean? That even if the nature of information is quite different in Le Parisien and in Le Figaro, and really different in The Australian compared to The Daily Telegraph, in each case, the information deals with the same states, even when the national bias is removed (Australia and France were deleted from tables during the analysis). The result is of course temporary as we need to test it on a bigger sample to validate it.

Presentation made in Barcelona