cluster_diagnostics() collects the quality measures of a clustering in
one net_cluster_diagnostics object, laid out the same way whichever
route produced the clusters: a distance clustering from
build_clusters(), a mixed Markov model from build_mmm() or
cluster_mmm(), the netobject_group that cluster_network() or
build_network() builds from one of them, or the net_mmm_clustering
attribute of such a group. It packages what the fit already computed and
adds only the per-cluster summaries the fit does not store (silhouette
means and mean within-cluster distances for a distance clustering, the
classification error within each cluster for a mixture); it does not
refit.
For a distance clustering it reports the average silhouette width, overall and per cluster, and each cluster's mean within-cluster distance. For a mixture it reports each cluster's average posterior probability (AvePP) and mixing proportion, the overall AvePP, the normalised entropy of the posterior probabilities (0 when every sequence is assigned with certainty, 1 when all assignments are equally likely; the reverse of the relative entropy often reported for latent class models), the classification error (the expected share of sequences misclassified when each is assigned to its most likely cluster: one minus the largest posterior, averaged over all sequences and over each cluster's members, so it equals one minus the AvePP; Vermunt & Magidson, 2005), and the log-likelihood with BIC, AIC and ICL.
These measures describe how cleanly one solution separates the
sequences; none of them picks the number of clusters. Compare solutions
with cluster_choice() for distance clustering or compare_mmm() for
mixtures, and build the clusters' networks with build_network() on the
clustering to see how they differ, with bootstrap_network() for how
stable their edges are.
Arguments
- x
The clustering: a
net_clustering, anet_mmm, anetobject_groupcarrying a clustering (built bycluster_network()or bybuild_network()from anet_clusteringornet_mmm), or anet_mmm_clustering. Anetobject_groupwithout a clustering, and any other object, raise an error. For theprint()andplot()methods: anet_cluster_diagnostics.- ...
Not used: an extra argument raises an error, here and in the
print()andsummary()methods. Forplot(): further arguments for the clustering's ownplot()method, such ascombined.- digits
For
print(): decimal places of the silhouette widths, distances, AvePP, entropy and information criteria, a non-negative whole number. Default3. Percentages are always shown to one decimal.- type
For
plot(): the view, passed to the clustering's ownplot()method. DefaultNULL, which draws"silhouette"for a distance clustering and"posterior"for a mixture. A distance clustering also takes"mds","heatmap"and"predictors"(seebuild_clusters()); a mixture takes"covariates", with"predictors"accepted for it when the source is anet_mmm_clustering(seebuild_mmm()).- object
For the
summary()method: anet_cluster_diagnostics.
Value
A net_cluster_diagnostics object. print() shows the numbers
of sequences and clusters (with the method and dissimilarity of a
distance clustering, or the number of states of a mixture), the overall
quality measures, the information criteria of a mixture, one row per
cluster, and the covariates of the fit, if any; plot() draws the clustering's own views. The object's
fields:
- family
"distance"or"mmm".- k, n, sizes
Number of clusters, number of sequences, and the cluster sizes.
- per_cluster
Data frame with one row per cluster; its columns depend on the family. Distance:
cluster,size,pct,mean_within_dist,sil_mean. Mixture:cluster,size,pct,mix_pct,avepp,class_err_pct.pct,mix_pctandclass_err_pctare in percent.- overall
List of the family's summary measures:
silhouettefor distance;avepp_overall,entropyandclassification_error(a proportion) for a mixture.- ics
For a mixture, a list with
BIC,AIC,ICLandlog_likelihood; NULL for distance.- metadata
The settings of the fit. Distance:
method,dissimilarity,weighted,lambda,medoids,covariates. Mixture:states,n_states,iterations,converged,covariates.- source
The clustering object itself, which
plot()draws without recomputing anything.
For print(): the input, invisibly. For plot(): what the
clustering's own plot() method returns, a ggplot object (invisibly)
or, for the covariate views with combined = FALSE, a list of
ggplot objects named by cluster (invisibly). For summary():
the per-cluster table described under per_cluster, a data frame with
one row per cluster.
References
Vermunt, J. K., & Magidson, J. (2005). Latent GOLD 4.0 User's Guide. Statistical Innovations.
See also
build_clusters, build_mmm,
cluster_choice,
compare_mmm for k-sweep model selection (MMM only).
Examples
seqs <- data.frame(V1 = sample(c("A","B","C"), 30, TRUE),
V2 = sample(c("A","B","C"), 30, TRUE))
cl <- build_clusters(seqs, k = 2, method = "ward.D2")
cluster_diagnostics(cl)
#> Cluster Diagnostics (distance) [ward.D2 / hamming]
#> Sequences: 30 | Clusters: 2
#> Quality: silhouette = 0.463
#>
#> Cluster N Mean within-dist Silhouette
#> 1 14 (46.7%) 1.132 0.283
#> 2 16 (53.3%) 0.608 0.620
# \donttest{
fit <- cluster_mmm(seqs, k = 2, n_starts = 1, max_iter = 20, seed = 1)
cluster_diagnostics(fit)
#> Cluster Diagnostics (mmm) [k = 2]
#> Sequences: 30 | Clusters: 2 | States: 3
#> Quality: AvePP = 0.946 | Entropy = 0.287 | Class.Err = 5.4%
#> ICs: LL = -56.481 | BIC = 170.782 | AIC = 146.962 | ICL = 174.190
#>
#> Cluster N Mix% AvePP Class.Err%
#> 1 21 (70.0%) 67.5% 0.943 5.7%
#> 2 9 (30.0%) 32.5% 0.952 4.8%
summary(cluster_diagnostics(fit))
#> cluster size pct mix_pct avepp class_err_pct
#> 1 1 21 70 67.48869 0.9428695 5.713045
#> 2 2 9 30 32.51131 0.9519734 4.802660
# }