Detect communities of co-associated modalities

Applies graph community detection to a catmodgraph object and writes the resulting community membership onto the graph vertices. Communities here are groups of modalities (factor levels) that tend to co-associate across different variables. This is a category co-association analysis, not a respondent-segmentation method; for the latter, see the poLCA or FactoMineR packages.

Usage

cluster_modalities(x, method = c("louvain", "walktrap"), signed = TRUE)

Arguments

x

A catmodgraph object.

method

Character. Community detection method. One of "louvain" (default) or "walktrap".

signed

Logical. If TRUE, only edges with positive standardised Pearson residual (attraction: modalities co-occurring more than expected under independence) are used for community detection, so communities are defined solely by positive co-association. Repulsion edges (negative std_resid) are excluded from the clustering graph but are retained on the original graph for downstream plotting with plot(mg, signed = TRUE). Only supported for method = "louvain". Default TRUE (changed from v0.9.0 which defaulted to FALSE).

Value

The input catmodgraph object with additional vertex attribute cluster, and a new component membership giving the cluster label for each modality node.

Details

Why signed = TRUE is now the default. The weight edge attribute stores the absolute phi coefficient so that edge thickness in plots scales with association strength regardless of direction. When unsigned Louvain uses these absolute weights, pairs of modalities with large negative residuals (strong repulsion, e.g. smoking_status=current and lung_disease=no) are pulled into the same community — the opposite of what subject-matter knowledge expects. Restricting clustering to positive-residual edges produces communities defined by genuine co-occurrence surplus, which are substantively more interpretable.

The std_resid edge attribute is retained on all edges in the original graph, so plot(mg, signed = TRUE) still shows both attraction (green) and repulsion (red) edges even when communities were detected on the positive-only subgraph.

Signed Louvain is a pragmatic adaptation. Modern igraph rejects negative edge weights in cluster_louvain(), so signed = TRUE simply drops repulsion edges rather than implementing a proper signed algorithm. For principled signed community detection (Traag & Bruggeman, 2009), see the signnet package.

signed = TRUE requires the std_resid edge attribute, which is attached by build_modality_graph and preserved through prune_modality_edges.

References

Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of communities in large networks. Journal of Statistical Mechanics, 2008(10), P10008. doi:10.1088/1742-5468/2008/10/P10008

Traag, V. A., & Bruggeman, J. (2009). Community detection in networks with positive and negative links. Physical Review E, 80(3), 036115. doi:10.1103/PhysRevE.80.036115

Examples

df <- expand_table(Titanic)
mg <- build_modality_graph(df)

# Default: positive-only Louvain (recommended)
mg <- cluster_modalities(mg)
table(mg$membership)
#> 
#> 1 2 
#> 6 4 

# Legacy unsigned behaviour: abs(phi) drives community detection
mg_abs <- cluster_modalities(mg, signed = FALSE)
table(mg_abs$membership)
#> 
#> 1 2 
#> 7 3