The workflow supports social network analysts in exploring and presenting the structure of social networks that are globally sparse but locally dense (small-world networks). It addresses three core analytical tasks: identifying communities (T1), identifying central/bridging actors (T2), and analysing roles and connection patterns (T3). The approach combines a node-link diagram for global structure with adjacency matrices for dense community detail in a single hybrid representation, progressively constructed through interactive aggregation.
The input is a social network — a graph where vertices are actors (e.g., researchers) and edges are relationships (e.g., co-authorship), with associated attributes (e.g., citation counts, collaboration strength). The system computes an initial energy-based layout (LinLog) designed to give prominence to clusters, producing a node-link diagram where densely connected communities appear as tight groups while the global sparse structure is visible.
The analyst inspects the initial node-link diagram to identify the global network structure: recognising community clusters (tight groups of nodes), spotting hub nodes with many connections, and identifying the overall topology (connected components, bridging actors). The LinLog layout provides visual cues through spatial grouping — clusters become candidate communities for subsequent matrix aggregation.
The analyst enters an iterative cycle of community aggregation and pattern analysis. They lasso-select a cluster of nodes and aggregate it into an adjacency matrix, which immediately reveals intra-community connection patterns (dense blocks, cross patterns, missing edges). They inspect the matrix for collaboration patterns: a "cross pattern" (one actor connected to all others, who do not collaborate among themselves — indicating a supervisor-student relationship), a "block pattern" (near-clique with most actors collaborating — indicating a tight research group), or intermediate patterns. Inter-community links remain visible as edges connecting matrices, revealing bridging actors.
The analyst manipulates the hybrid representation: moving actors in and out of matrices to test community boundaries and identify articulation points (e.g., extracting Ed Chi from the PARC matrix reveals that the community becomes disconnected — confirming his bridging role); merging matrices to investigate potential larger communities; splitting matrices to explore sub-communities; adjusting link thickness via slider to switch between overview mode (how are communities linked?) and detail mode (who links them?); and adjusting matrix size and axis labels for readability. After each manipulation, the analyst assesses whether the current representation adequately reveals the community structure or whether further aggregation, splitting, or rearrangement is needed.
Once the analyst has identified all communities, central actors, and collaboration patterns, they compose a final NodeTrix layout optimised for communication: adjusting positions, reducing matrix axis labels for compactness, and choosing appropriate visual variable mappings (colour for citations on axes, intensity for collaboration strength in cells). The result is a summary visualisation that simultaneously communicates global structure and community-level detail without the readability problems of either pure node-link or pure matrix representations.