Network collection and visualisation may reveal how there are several subgroups in a network. For example, in Figure 4.5 there are two clear subgroups: nodes A-D and nodes F-I. Social scientists benefit from understanding subgroups when they form clusters or communities - that is, nodes that are highly connected within the same subnetwork but less connected to other subgroups. For example, using network analysis, Zachary (1977) was able to identify two factions in a Karate club by exploring members' social connections - before the club was split into two factions. Similarly, Cross et al. (2002, Exhibit 1) visualise how an expert consulting group is divided into two subgroups, domain and technology experts. They identify only one individual as a bridge between these two expert communities. In both of these, the network analysis is used to study the internal organisation of the group and identify communities, bridges and structural holes in the network. These techniques can be used for any kind of network, including social networks like above but also other types of networks discussed throughout this chapter.
The cases above focused on small networks, and the existence of subgroups was obvious from the network plots. Therefore, they did not require a formal method to detect these subgroups. However, larger networks require more systematic approaches. While the approach is different from unsupervised machine learning (see Section 3.4), the outcome from these algorithms is the same: some groups or clusters that stand out from other groups. In network analysis, these communities (or clusters) are a set of nodes that have a high density of ties within the set and lower likelihood of ties for nodes outside the cluster. Thus, clusters are groups of nodes that are more connected within the group and less connected with other nodes in the network.
There are several ways to detect these communities, and within the community the discussion about which methods to use is actively debated (e.g. Newman, 2006). Within social sciences, people often recommend using modularity-based network analysis. However, it has recently been under criticism as it might not capture groups in all contexts (Guerra et al., 2013). Similar to visualisation, it seems that different approaches provide different analysis results. To illustrate, we now present a few approaches to show how they might work and what kinds of results may emerge from them.
The Girvan-Newman approach takes benefit from the betweenness to explore communities (see Figure 4.8a). It removes the highest betweenness ties from the network, thus alternating the original network. This process is continued, removing one tie at a time. Thus, communities emerge or become visible in this process as ties that connect different groups are removed.
Another approach for community detection is hierarchical clustering (see Figure 4.8b). Unlike dividing the network as in the Girvan-Newman approach, the communities develop from merging potential communities to larger ones. The algorithm measures which nodes are similar, such as which nodes have a similar type of ties with other nodes. Each time, these nodes are paired together to communities with two members.
The ability to determine communities from a network - through any of the algorithms available for the task - allows answering many relevant questions for social sciences. Examples include social media analysis, where clusters have been used to unravel groups that are more likely to interact among themselves; bubbles of discussion, both for politically oriented differences and non-political bubbles (e.g., Himelboim et al., 2013; Maireder and Schlögl, 2014; Ausserhofer and Maireder, 2013; Highfield et al., 2013); or examination of scientific disciplines through their citation practices and clusters that emerge in them (Edelmann et al., 2020; Velden and Lagoze, 2013).