I'm working through a comprehensive workflow for analyzing co-authorship networks in the IEEE VIS conference dataset, which involves detecting communities, tracking their stability over time, characterizing research topics within each group, and monitoring how communities merge, split, or dissolve across years. I'll need to design a systematic approach that includes visual inspection checkpoints at each analytical stage to validate the findings. For each time window, I'll apply a community detection algorithm like Leiden to partition the co-authorship graph, tuning the resolution parameter and visually inspecting the results. The key decision is whether to detect communities independently in each snapshot or use a temporal/dynamic approach that maintains continuity across time windows—I'm leaning toward a hybrid where I detect communities per window but then track how they evolve.

I'll use overlapping sliding windows and match communities across consecutive time periods by comparing their membership overlap with a Jaccard index, building a lineage that shows when communities merge, split, or persist. This gives the analyst interpretable intermediate results at each stage while still capturing temporal dynamics through the community evolution graph.

For the matching and evolution tracking, I'll compute similarity scores between communities in adjacent windows to identify key events like births, deaths, mergers, and splits, then visualize this as an alluvial or Sankey diagram so the evolution patterns are easy to inspect.

Now I'm tracking core members within each community lineage—identifying authors who persist across multiple time windows and ranking them by centrality metrics and citation influence to distinguish the actual leaders. I'll represent this as a heatmap showing author presence over time, sorted by persistence, or use an upset plot to show how membership overlaps across communities and periods.

For characterizing what each community is about, I'm applying text analysis techniques like TF-IDF and topic modeling to the papers in each community across time windows. ual author trajectories across communities using path-based visualizations, word clouds showing top keywords per community-time node, and a core member heatmap with persistence scores. I'm also considering overlaying citation impact metrics like node size or average citations per community, with drill-down capabilities so analysts can inspect specific communities to see their papers, authors, topics, and citation networks. Then in the validation stage, I'll have domain experts review these visualizations against known research areas like SciVis, InfoVis, and VAST to confirm whether the detected communities align with actual subfields.

I need to handle several practical challenges: isolates and sparse collaborations, weighting edges by recency to avoid outdated connections, normalizing for field growth over time using resolution parameter scaling, and running community detection multiple times to assess statistical robustness.

I should also validate against null models and randomized graphs to ensure the community structure isn't spurious. This connects to established bibliometric work on "invisible colleges" using co-authorship networks, and there's relevant prior research on the VIS community specifically using the vispubdata.org dataset from IEEE VIS papers spanning 1990 to present.

For visualizing how communities evolve over time, storyline visualization techniques are particularly well-suited—they're designed to show group mergers and splits, which directly addresses tracking community dynamics across the temporal dimension of this dataset.

Since this is likely a design workflow question, I should structure the answer around problem decomposition, pipeline stages with data transformations, algorithmic choices with justification, and then map specific visualization techniques to each analytical question. I'll also want to highlight human-in-the-loop feedback mechanisms and present the workflow as a structured progression through stages, methods, and corresponding visualizations. I'm outlining a multi-stage visual analytics pipeline that combines network analysis with text mining—using dynamic community detection on co-authorship networks alongside topic modeling, all structured as an iterative sensemaking loop where analysts validate intermediate results before moving forward. The workflow starts with data preparation: building a bipartite paper-author dataset with temporal and venue information, resolving author identities, then constructing time-windowed co-authorship networks using a sliding window approach.

For the network construction, I'm weighting edges by the number of co-authored papers within each window, optionally normalizing by author count per paper using the standard bibliometric formula. Node attributes capture cumulative publication counts, citation totals, and venue distributions. I'm planning an initial visual validation step that plots network statistics like density and component size over time as line charts, plus spot-checking a few year snapshots with force-directed layouts to ensure the windowing scheme makes sense.

Next I'm applying the Leiden algorithm to detect communities in each window graph, running it multiple times to ensure robust results.

Analytical Workflow: Community Structure & Evolution in the IEEE VIS Co-Authorship Network

Pipeline overview

Raw records → Time‑windowed co‑authorship networks → Per‑window community detection
→ Cross‑window matching (lineages) → Core/leader analysis → Topic characterisation
→ Cross‑community overlap analysis → Coordinated visual front‑end → Domain‑expert validation loop

Because the dataset is small enough (3,596 papers) for fully interactive analysis, the workflow is designed as an iterative, human‑in‑the‑loop visual analytics pipeline: every algorithmic stage produces an intermediate artefact that is inspected visually before the next stage runs, and parameters can be revised based on domain knowledge.

---

1. Network construction

• Build a paper–author table (year, sub‑conference Vis/InfoVis/VAST/SciVis, title, abstract, keywords, citations, disambiguated author ID).
• Construct overlapping sliding time windows (e.g., 5‑year windows, 1‑year step) rather than cumulative or single‑year graphs — this smooths sparse early years while keeping enough temporal resolution to detect merges/splits.
• Edge weight = co‑authored papers in the window (optionally fractional-count normalised for author‑list length).

Check (visual): density/#components/giant‑component‑size over time, plus force‑directed layout of a few sample windows — sanity check before detection.

2. Per‑window community detection

• Use Leiden with the CPM objective (not plain modularity — the field's size grows ~10× from 1990–2024, so modularity's resolution limit would bias detection toward large late‑period, small early‑period communities).
• Run multiple seeds → consensus clustering to get robust partitions; expose resolution parameter γ interactively.

Check (visual): resolution profile (γ vs. #communities/modularity); node‑link/matrix view colored by community; cross‑tab heatmap community × sub‑conference as a domain-knowledge sanity check.

3. Cross‑window matching → community lineages

• Similarity between communities in adjacent windows = Jaccard overlap of author sets.
• Classify transitions with standard rules: birth, death, continuation, growth/shrink, merge (n→1), split (1→n).
• Assign persistent lineage IDs to tracked communities.

Check (visual): alluvial/storyline diagram (node size = community size, ribbon width = shared members) — the primary artefact for answering "how many communities, do they merge/split/dissolve, when". Analyst annotates events using domain knowledge (e.g., the 2013 unification of InfoVis/VAST/SciVis into one VIS proceedings is a plausible structural break to check against).

4. Core / leadership / stability analysis

• Per lineage, build an author × window presence matrix.
• Persistence score = fraction of windows in the lineage's span where the author is present.
• Centrality (degree/eigenvector, citation‑weighted) per window.
• Core = high persistence + high centrality; leaders = top core members by citations/centrality.
• Core stability index = mean pairwise Jaccard of the core set across consecutive windows.

Check (visual): heatmap (authors × windows, sorted by persistence), small multiples across lineages; ranked leader table.

5. Cross‑community overlap (bridging authors)

• Per window, compute each author's participation coefficient across communities (bridges vs. loyal members).
• Aggregate an inter‑community network (meta‑nodes = communities, edge weight = shared/collaborating authors).

Check (visual): aggregated meta‑network (animated/slider over time); UpSet plot of author-set memberships across lineages.

6. Topic characterisation

• For papers attached to each community‑window node: TF‑IDF top terms or topic model (LDA/BERTopic) on title+abstract+keywords.
• Track topic profile along each lineage.

Check (visual): word clouds/top‑term lists attached to storyline nodes; ThemeRiver of topic proportions per lineage; UMAP projection of community‑window topic vectors colored by lineage (reveals topical convergence before merges / divergence before splits).

7. Integrated visual front‑end (coordinated views)

Linked views with brushing: storyline (evolution) ↔ network snapshot ↔ core heatmap ↔ topic panel ↔ citation‑impact overlay (bubble size/line chart). Selecting a lineage or window highlights consistently across all views; hovering a node shows top authors, keywords, and papers.

8. Validation & iteration

• Compare detected leaders/topics against known VIS researchers/subfields; check alignment with sub‑conference labels and known historical events (e.g., conference restructuring).
• If misaligned, revisit window size, γ, or matching threshold; run sensitivity analysis (small multiples of storylines under varied parameters) to confirm robustness of major lineages.

---

Mapping questions → methods → visualisations

Analyst question	Method	Key visualisation
How many communities exist?	Leiden (CPM) per window + lineage tracking	Storyline/alluvial diagram, community-count timeline
Core stability & leaders?	Persistence × centrality on author–window matrix	Sorted heatmap, ranked leader table
Main research foci & evolution?	Topic modelling per community‑window	ThemeRiver, word clouds on storyline nodes
Membership overlap between communities?	Participation coefficient, meta‑network	Aggregated community graph, UpSet plot
Merges/splits/dissolutions — when/why?	Rule‑based transition classification	Annotated alluvial diagram + topic/UMAP cross‑check