"Who is with whom" and "did the structure reorganise, and when" are different questions. This one compares each bin's partition with another bin's by a label-invariant statistic, so it says how much changed without needing to know which community became which.
Because every statistic here is invariant to relabelling,
match_communities() is not a prerequisite. That is a common
misunderstanding and it is worth stating plainly: matching is needed to
follow a community, not to measure change.
Arguments
- x
A
dynet_communitiesframe fromtemporal_communities().- measure
One or more of
"nmi","ami","uami","iami","ari","vi","split_join","jaccard"and"omega_index". The default asks for the first several; name them explicitly to choose.- against
"previous"compares each bin with the one before it,"first"with the opening bin,"all"with every other bin.
Value
A dynet_metric at graph level. For "previous" and "first",
one row per bin per measure with columns session, time, measure,
value; for "all", one row per ordered pair with an other column
giving the bin compared against, in the same long shape similarity()
uses so both plot with the same code. The n_compared attribute records
how many vertices each comparison could actually use.
Details
Every statistic is built from the contingency table of the two labellings,
restricted to the vertices active in both bins. A vertex that merely
goes quiet must not read as a reorganisation, so it is dropped from both
sides rather than counted as having left its community. When fewer than two
vertices are shared, nothing is defined and the value is NA – a real
absence, not a zero.
The first bin has no predecessor, so under against = "previous" its value
is NA by construction.
What each measures, briefly. "nmi" is shared information normalised by
the mean entropy, in \([0, 1]\), high for agreement; it is not
chance-corrected, so two unrelated partitions score above zero. "ami" is
the same quantity with the chance level subtracted, and is the one to reach
for. "ari" is chance-corrected pair agreement, centred on zero for
unrelated partitions. "vi" is a true metric in nats, zero for agreement
and unbounded above. "split_join" counts the vertices that would have to
move, so it is an integer in \([0, 2N]\). "jaccard" scores agreement
over co-classified pairs. "omega_index" is chance-corrected pair
agreement; for the disjoint partitions this package produces it coincides
exactly with "ari", and it is offered because it is what multinet
reports and because it generalises to overlapping communities.
"uami" and "iami" are Zhong et al.'s (2025) answer to a problem every
temporal network has: the vertex set moves between bins, and ordinary
mutual information assumes it does not. "iami" compares over the
intersection, the vertices both bins have. "uami" compares over the
union, giving each partition one extra virtual community holding the
vertices it does not have, so a vertex arriving or leaving is itself
information about the change rather than something discarded. The two
answer different questions: "iami" asks whether the vertices that stayed
kept their company, "uami" asks whether the whole structure held. Since
every other measure here is already restricted to the shared vertices,
"iami" is numerically identical to "ami" in this package, and is
offered under both names because the literature uses both.
Two notes on the sources, because both cost time to discover. Adjusted
mutual information admits more than one normaliser: this package divides by
the arithmetic mean of the two entropies, matching "nmi" above,
sklearn's default, and the stated intent of Zhong et al.; aricode::AMI
divides by the maximum instead, so the two disagree by convention and
not by error. And the expression printed for the adjusted measures in Zhong
et al. subtracts the chance term once in a denominator that needs it twice,
which scores two identical partitions at 0.949 rather than 1; the corrected
form is used here.
References
Danon, L., Diaz-Guilera, A., Duch, J., & Arenas, A. (2005). Comparing community structure identification. Journal of Statistical Mechanics, P09008.
Hubert, L., & Arabie, P. (1985). Comparing partitions. Journal of Classification, 2, 193-218.
Meila, M. (2007). Comparing clusterings – an information based distance. Journal of Multivariate Analysis, 98(5), 873-895.
van Dongen, S. (2000). Performance criteria for graph clustering and Markov cluster experiments. CWI Technical Report INS-R0012.
Collins, L. M., & Dent, C. W. (1988). Omega: a general formulation of the Rand index of cluster recovery suitable for non-disjoint solutions. Multivariate Behavioral Research, 23(2), 231-242.
Gates, A. J., & Ahn, Y.-Y. (2019). Element-centric clustering comparison unifies overlaps and hierarchy. Scientific Reports, 9, 8574.
Vinh, N. X., Epps, J., & Bailey, J. (2010). Information theoretic measures for clusterings comparison. JMLR, 11, 2837-2854.
Zhong, P., Ba, C., Mondragon, R., & Clegg, R. (2025). Quantifying community evolution in temporal networks. Scientific Reports, 15, 45373.
Examples
dn <- dynet(school_contacts)
found <- temporal_communities(dn, step = 5, window = 5, seeds = 1:5)
community_change(found)
#> # Community change (graph-level)
#> # 5 time points, 5 per bin | time in step
#> # measures: nmi, ami, ari, vi, split_join, jaccard, omega_index, uami, iami
#> # compared against the previous bin, over vertices active in both
#> time measure value
#> 0 nmi NA
#> 0 ami NA
#> 0 ari NA
#> 0 vi NA
#> 0 split_join NA
#> 0 jaccard NA
#> 0 omega_index NA
#> 0 uami NA
#> 0 iami NA
#> 5 nmi 1
#> 5 ami 1
#> 5 ari 1
#> # 33 more rows. summary() aggregates them; plot() draws them.
community_change(found, measure = c("ami", "ari"))
#> # Community change (graph-level)
#> # 5 time points, 5 per bin | time in step
#> # measures: ami, ari
#> # compared against the previous bin, over vertices active in both
#> time measure value
#> 0 ami NA
#> 0 ari NA
#> 5 ami 1
#> 5 ari 1
#> 10 ami 1
#> 10 ari 1
#> 15 ami 1
#> 15 ari 1
#> 20 ami 1
#> 20 ari 1
community_change(found, measure = c("iami", "uami"))
#> # Community change (graph-level)
#> # 5 time points, 5 per bin | time in step
#> # measures: iami, uami
#> # compared against the previous bin, over vertices active in both
#> time measure value
#> 0 iami NA
#> 0 uami NA
#> 5 iami 1
#> 5 uami 1
#> 10 iami 1
#> 10 uami 1
#> 15 iami 1
#> 15 uami 1
#> 20 iami 1
#> 20 uami 1