Social Network Analysis

What you’ll learn

• What Social Network Analysis (SNA) shows and why it matters
• How co-occurrence data drives the network graph
• Which tuning knobs are available and what they affect
• How community detection works and which algorithms are supported
• How cohorts filter what edges you see
• How to compare the SNA result against your catalog kinship graph

What is Social Network Analysis?

Social Network Analysis builds an interactive network graph of your population from encounter co-occurrence data. Two animals seen in the same encounter form a network edge; the more frequently they co-occur, the heavier the edge. A community-detection algorithm groups densely-connected nodes into clusters that often correspond to real social units — family groups, foraging parties, persistent associations.

This is valuable for:

• Identifying stable social groups
• Tracking changes in association patterns over time
• Discovering which individuals serve as connectors between groups
• Validating known social structure against the data

Pre-requisites

Before you can run an SNA:

1. Encounter data with multiple identified animals per encounter. Singletons produce no edges by definition.
2. An active age cohort version for the population. SNA is cohort-aware — every edge instance carries the cohort each endpoint was in at the encounter date. The submit drawer disables when no cohort version exists.
3. (Optional) An active catalog graph if you want to run the kinship overlay, validation report, side-by-side compare, or kinship-weighted bias features documented in Catalog vs SNA.

Submitting a new analysis

The New analysis drawer collects four blocks of input:

Date range and name

Pick a name and the encounter window the analysis should cover. Edges are constructed only from encounters whose date falls in the window.

Age cohort version

The cohort version anchors how each edge instance is labelled by sex/age. The version is snapshotted into the result — later changes to the cohort definition don’t retroactively affect this analysis. See Age Cohorts for details.

Network parameters

Three tuning knobs let you trade noise for signal:

Knob	What it does	Default
Min sightings per individual	Drop animals seen in fewer than N distinct encounters before edge construction. Removes long-tail singletons that distort the layout.	1 (no filter)
Min encounter size	Ignore encounters with fewer than N identified individuals. Singletons make no edges anyway; raising this drops “two-animal-only” sightings if you don’t trust them as social signal.	2
Community detection algorithm	Which method partitions the graph (see below).	Louvain

Bias toward known kinship (optional)

When enabled, finwave fetches the active catalog’s mother-of, father-of, sibling-of, and mate-of edges and the worker adds a per-edge weight bonus to those pairs before community detection. Use this when sparse data is fragmenting known matrilines. Full details: Kinship-weighted SNA.

Community detection

Three algorithms are supported. They all partition nodes into communities; they differ in how they decide where to cut.

Algorithm	Best for	Notes
Louvain (default)	Most populations. Fast, scales well, good modularity on dense networks.	Greedy modularity optimisation.
Edge betweenness	Smaller, more deeply-structured networks where the boundary between groups matters more than raw modularity.	Slower (O(VE) per partition step). Closer to the classical Girvan–Newman approach.
Label propagation	Quick exploratory partitions. Useful as a sanity check against Louvain.	Non-deterministic — different runs can produce different communities.

Communities render as colour-coded clusters in the visualization. Each community represents a group of animals that associate more frequently with each other than with the rest of the population. Modularity (a single number reported under the graph) measures how cleanly-separated those communities are: higher = stronger community structure, with values above ~0.4 typically considered meaningful.

Tip

Compare detected communities against your existing knowledge of the population. Unexpected groupings can reveal associations you hadn’t documented; expected groups that fail to appear may indicate gaps in observation data.

The chart area

When the analysis completes, the chart area renders below the jobs table.

Top row controls

• Node opacity / Edge opacity sliders — fade either the nodes or the edges to focus on the other layer.
• Fit to view — recenter the camera on the full graph.
• Compare — swap the chart for the side-by-side catalog comparison view.
• PNG / SVG export — save the current rendering to disk.

Show known kinship overlay

The Show known kinship toggle in the chart header adds catalog-curated kinship edges as a styled second edge layer. See Kinship overlay for the full reference.

Cohort filter

When the analysis has cohort data, a unified Show dropdown sits above the graph:

Mode	What it shows
All edges	Every edge in the network, regardless of which cohorts are selected. The cohort chips are a colour legend only.
Edges touching any selected cohort	Edges where at least one encounter had at least one endpoint in the currently-selected cohort chips. Surfaces boundary edges between selected cohorts and the rest.
Edges between selected cohorts	Edges where at least one encounter had both endpoints in selected cohorts. Surfaces within-group structure.

The chip selection is the dropdown’s argument: in All edges mode the chips are disabled (legend only); in the other two modes they’re toggleable filters. A help line below the row explains the current state (“Showing every edge — switch the mode to filter by cohort” / “No cohorts selected — graph and matrix are empty”).

Mixing matrix

The matrix below the graph shows raw or row-normalised co-occurrence counts between every cohort pair. The matrix follows the cohort filter — it counts only the same instances that the rendered graph shows, so the two views can never disagree. A small subtitle on the matrix header reads either “All co-occurrences” or “Matches the current edge filter” so the coupling is unambiguous.

The Row-normalized toggle switches between absolute co-occurrence counts and per-row fractions (what fraction of cohort X’s co-occurrences were with cohort Y).

Diagnostics strip

Under the matrix:

• Nodes / Edges counts follow the filter; a faint / total suffix shows when the filter is restricting the visible count.
• Communities / Modularity are computed at analysis time on the full network and don’t recompute under client-side filtering. They display with a small info_outline icon and tooltip explaining this.

Catalog vs SNA card

When the kinship overlay is on, a validation report card appears beneath the matrix with per-type recovery numbers and an out-of-graph relationships list.

Reviewing past analyses

The jobs table below the chart lists every previous run. Click a job’s name to render its result. Click the row’s actions menu to view details (parameters, completion time) or delete the job.

When you click View details on an existing job, the submit drawer reopens with the job’s parameters pre-populated — useful for cloning a run with one parameter changed. The drawer is read-only for completed jobs; there’s no in-place edit since the result is already pinned to a specific parameter set.

Access

Social Network Analysis is a Workbench tool that requires finwave Pro. Your population administrator controls which roles can access it through the Workbench Access settings. By default, administrators and professionals have access; experts and novices do not unless explicitly granted.

Related

• Catalog vs SNA — kinship overlay, validation report, side-by-side compare, snapshots, and kinship-weighted bias
• Age Cohorts — the cohort version is required input
• Catalog Builder — where kinship data comes from
• Composition Analysis — demographic breakdown
• Discovery Curves — tracking new individual discovery rates
• Capture History Export — capture history matrices for mark-recapture
• Workbench Overview — all Workbench tools and how access works