Retweet networks of the European Parliament: evaluation of the community structure

The European Parliament (EP) functions analogously to national parliaments in traditional parliamentary democracies. It is elected every five years directly by the citizens of the EU countries. Member states are allocated a number of seats which roughly reflect the state’s population. The EP members are elected on a national basis, but sit in the EP according to political groups they belong to. They can address the EP in any of the 24 official languages of the EU, but data shows that they primarily use English.

Our work focuses on the period between October 1, 2014 and September 30, 2015. This period falls within the 8th EP which was elected on July 1, 2014. During this period, the EP consisted of 8 political groups:

We acquired the list of the EP members form the official site of the EP². The list consists of 750 EP members. Their distribution according to political groups is presented in Table 1 (column EP seats). The official Twitter account of the EP, Europarl_EN, provides a list of Twitter accounts of the EP members³. We matched the EP members to the Twitter accounts and excluded Twitter accounts of former EP members; the result is a manually verified list of 546 Twitter accounts of all the EP members which have one. The distribution of the EP members with Twitter accounts according to political groups is given in Table 1 (column Twitter accounts).

We monitored the activity related to the official accounts of the EP members through the Twitter Streaming API⁴. For each member, we acquired all their tweets as well as all the replies and retweets.

Within the period of our analysis—between October 1, 2014 and September 30, 2015—the EP members have posted 561,255 tweets, of which 295,395 (53 %) are originally authored and the other 265,860 (47 %) are retweets. On average, all EP members together posted 1,538 tweets per day, and each active member posted on average 3.1 tweets per day (Fig. 1).

Together with each tweet, Twitter provides metadata about the tweet which includes the identified language of the tweet. We use this information to determine the language an EP member uses on Twitter. For each EP member, we compute the distribution of languages across all of the tweets posted by her/him. We determine the most commonly used language and use this language in the analysis. In some cases, Twitter is unable to determine the language of a tweet and reports the language as ‘undetermined’. If most of the tweets of an EP member are categorized as ‘undetermined’, we consider him/her as using ‘undetermined’ as their preferred language.

The collected tweets described in the previous section are used to construct retweet networks. A retweet network is a directed weighted graph, where nodes represent Twitter users and edges represent the retweet relation. The direction of an edge corresponds to the direction of information spreading or influence; the weight of the edge is the number of times one user retweets the other. We construct two retweet networks: (i) the core network, containing as nodes only the EP members and (ii) the extended network, containing as nodes the EP members and all other users which have retweeted or have been retweeted by an EP member.

The core network consists of 459 nodes and 4,441 edges. The distribution of nodes according to political groups is in Table 1 (column Core network). The extended network consists of 498,103 nodes, of which 507 are the EP members, and 808,505 edges. The distribution of the member nodes according to political groups is also in Table 1 (column Extended network). Note that there are more EP members in the extended network (507) than in the core network (459) since 48 EP members (the difference) have been retweeted only by non-EP members. The extended network is much larger, and at the same time much sparser than the core network. The sparsity of the extended network is expected. While the core network represents interactions between the EP members, the interactions between the EP members in the extended network are dwarfed by the interactions between the general public on Twitter and the EP members. These interactions are mostly unidirectional, represented as retweets by the general public of the tweets posted by the EP members. An overview of the size and modularity of both networks is in Table 2.

In Fig. 2, we present the core network with a force-directed layout, where the color of the node identifies the political group of the EP member. There is an intuitive visual grouping of the EP members according to political groups. In Fig. 3, the extended network with the detected communities is presented.

The goal of most community detection algorithms, implicit or explicit, is to find the best trade-off between a large intra-cluster density and a small inter-cluster density. Community detection algorithms perform maximization of modularity (Newman 2006). A good partitioning of a network in communities is one in which there are fewer than expected edges between the communities. The modularity is, up to a multiplicative constant, the number of edges falling within groups minus the expected number in an equivalent network with edges placed at random. Previous work on roll-call votes suggests that the result of modularity optimization should find groups and coalitions in a parliament (Dal Maso et al. 2014).

We perform community detection using the well established Louvain algorithm (Blondel et al. 2008). The Louvain method is a computationally very efficient algorithm that is well suited for large networks. It optimizes modularity through an iterative heuristic approach that consists of two repeating phases. In the first phase, modularity is optimized by allowing only local changes in communities; in the second, a new network is build that consists of one node for each previously found community. The algorithm repeats the iterations until the first phase can make no further improvements in modularity.

To asses how closely the detected communities correspond to actual groups, we use the B ³ measure (Bagga and Baldwin 2008), which is considered the most appropriate measure for extrinsic evaluation of clustering (Amigó et al. 2009). The B ³ measure decomposes the evaluation into calculating the precision and recall associated with each node in the network. Let N be the set of all nodes in the network. For each node n∈N, we denote as L(n) the set of nodes which have the same label as n, i.e., members of the same actual group. With C(n), we denote the set of all nodes which are members of the same community as n. The B ³ precision of a node n, P(n), is computed as the fraction of nodes which have the same label and are in the same community as n, from all the nodes which are in the same community as n. Similarly, the recall of a node n, R(n), is computed as the fraction of nodes with the same label and in the same community, from all the nodes with the same label as n.

The precision and recall can be further combined into an F ₁ score, which is a harmonic mean of the precision and recall:

The F ₁ score is a special case of Van Rijsbergen’s effectiveness measure (Van Rijsbergen 1979), where precision and recall can be combined with different weights.

The precision reflects the homogeneity of the community. The lower the number of actual groups in the community, the higher the precision. Conversely, the recall reflects to compactness of the actual group. The lower the number of detected communities in an actual group, the higher the recall. The F ₁ score balances the precision and recall.

Furthermore, to quantify how well an actual group is reflected in the community structure of the network, we calculate the mean precision, recall, and F ₁ of the EP members for each group. Let {L ₁,L ₂,…,L _k } be the partitioning of the nodes according to actual labels. The precision, recall, and F ₁ score of the set of the nodes L _i are computed as:

Figure 5 shows the correspondence between the political groups and detected communities. First, it points out how members of different political groups are distributed across communities. Generally, most of the members of one group are located in a single community. The EP members from S&D, however, are divided almost evenly into two communities (C0 and C4); the members from EFDD, are also divided into two communities (C2 and C5). Second, Fig. 5 also shows the composition of communities with respect to political groups. In general, the communities consist mostly of members of a single group. Notable exceptions are community C2 which contains many EP members from both GUE-NGL and EFDD, and community C5 which contains EP members from ECR and EFDD.

We calculate the mean precision, recall, and F ₁ score for the core network to characterize how well the community structure reflects the political groups of the EP members. The results are shown in Fig. 4 (rows Groups). The precision is high, 0.843, which reflects the fact that most of the communities, with the exception of C2 and C5, are dominated by a single political group. The recall is moderately high, 0.764, which reflects the fact that most of the political groups, with the exception of S&D and EFDD, are predominantly contained within a single community. The F ₁ score is also moderately high, 0.777. In comparison, a random partitioning of the graph into 8 communities has (on average over 1000 random partitionings) precision of 0.195, which is over 4 times lower, recall of 0.140, which is over 5 times lower, and F ₁ score of 0.149, which is also over 5 times lower than the scores obtained with the partitioning into communities.

The correspondence measures for the political groups are presented in Fig. 6. GUE-NGL has an average precision (0.566) and moderately high recall (0.744), which corresponds to its members being dispersed in several groups, in only one of which they constitute the majority. S&D has a very high precision (0.954) and the lowest recall (0.489), as a result of being almost perfectly split into two communities where its members are an overwhelming majority. Greens-EFA has a high precision (0.813) and recall (0.870) because its members are mostly contained in a single community where they are a majority. ALDE has the highest precision (0.956) and recall (0.957) due to the fact that almost all of its members are contained in a single community which contains very few others. EPP has a very high precision (0.951) and recall (0.935) since it members are predominantly contained in one large community with only a few other members. ECR has above average precision (0.645) and high recall (0.899) as a consequence of being contained predominantly in a single community which contains almost as many members from other groups. EFDD has the lowest precision (0.273) and an average recall (0.506) resulting from the fact that it is evenly divided into two communities, in none of which its members are a majority. And finally, NA has very high precision (0.916) and recall (0.917) corresponding to the largest part of its members being in a single community which contains only one member from another political group.

We also evaluate the correspondence between the EU countries and the detected communities. We again calculate the mean precision, recall, and F ₁ score for the network to characterize how well the community structure reflects the country membership of the EP members. The results are shown in Fig. 4 (rows Countries). Both precision and recall are low, 0.191 and 0.321 respectively, which shows that communities in the core network and not organized along country membership. In comparison, a random partitioning of the graph into 8 partitions has (on average over 1000 random partitionings) precision of 0.093, recall of 0.178, and F ₁ score of 0.106, which are all only around 2 times lower than the scores obtained with the partitioning into communities. We do not further investigate the differences between individual countries since the overall correspondence between the countries and the communities is low.

Lastly, we evaluate the correspondence between languages and communities. The results are shown in Fig. 4 (rows Languages). Again, precision and recall are both low, 0.212 and 0.263 respectively, which shows that communities in the core network and not organized along the language in which EP members post tweets. In comparison, a random partitioning of the graph into 8 partitions has (on average over 1000 random partitionings) precision of 0.138, recall of 0.167, and F ₁ score of 0.125, which are all only 1.5 times lower than the scores obtained with the partitioning into communities. Again, we do not further investigate the differences between individual languages since the overall correspondence between the languages and the communities is low.

Analogously to the core network, we analyze how closely the partitioning in communities corresponds to the partitioning in political groups.

The mean precision, recall, and F ₁ score for the extended network, which characterize how well the community structure reflects the political group membership, are presented in Fig. 7 (rows Groups). Both precision and recall (and subsequently F ₁) are low. In comparison, a random partitioning of the graph into 16 partitions has (on average over 1000 random partitionings) precision which is almost 2 times lower, recall which is 3 times lower, and F ₁ score which is around 2 times lower than the ones obtained with the partitioning into communities. These values are still substantially lower than the ones obtained for the core network with respect to political groups.

We next investigate how the country of origin of the EP members is reflected in the community structure. To this end, we evaluate the matching between the 16 communities and the 28 EU countries. Figure 8 illustrates how members from different EU countries are spread across the detected communities. Most countries have their members contained within only a few communities. Moreover, for the majority of countries, one community contains the prevailing number of the its members.

The evaluation results for the partitioning in countries are presented in Fig. 7 (rows Countries). In comparison to the partitioning in political groups, they are substantially higher. We also evaluated the average random partitioning, which has precision, recall, and F ₁ score that are around 5.5 times lower than the ones obtained with the partitioning into communities. These values are comparable with those for the partitioning in political groups of the core network.

Figure 9 shows the mean precision, recall, and F ₁ score for each EU country. The F ₁ scores for the different countries vary substantially, ranging from 0.028 for Estonia to 0.899 for Poland.

Finally, we investigate how the language of Twitter posts of the EP members is reflected in the community structure. Hence, we evaluate the correspondence between the 16 detected communities and the 23 languages used on Twitter. We rely on the Twitter language recognition which is not perfect. Figure 10 illustrates how different languages used by the EP members are spread across the detected communities. The size of each language group in Fig. 10 is proportional to the number of the EP members who primarily use that language. Note that the community identifiers are the same as in Fig. 8.

The evaluation measures for the partitioning by language are presented in Fig. 7 (rows Languages). In comparison to the partitioning by countries, they are slightly lower. This is an interesting result since one would expect that common language used on Twitter is more important than the country of origin.