View source: R/dissimilarity.R
cl_dissimilarity  R Documentation 
Compute the dissimilarity between (ensembles) of partitions or hierarchies.
cl_dissimilarity(x, y = NULL, method = "euclidean", ...)
x 
an ensemble of partitions or hierarchies and dissimilarities,
or something coercible to that (see 
y 

method 
a character string specifying one of the builtin
methods for computing dissimilarity, or a function to be taken as
a userdefined method. If a character string, its lowercased
version is matched against the lowercased names of the available
builtin methods using 
... 
further arguments to be passed to methods. 
If y
is given, its components must be of the same kind as those
of x
(i.e., components must either all be partitions, or all be
hierarchies or dissimilarities).
If all components are partitions, the following builtin methods for measuring dissimilarity between two partitions with respective membership matrices u and v (brought to a common number of columns) are available:
"euclidean"
the Euclidean dissimilarity of the memberships, i.e., the square root of the minimal sum of the squared differences of u and all column permutations of v. See Dimitriadou, Weingessel and Hornik (2002).
"manhattan"
the Manhattan dissimilarity of the memberships, i.e., the minimal sum of the absolute differences of u and all column permutations of v.
"comemberships"
the Euclidean dissimilarity of the elements of the comembership matrices C(u) = u u' and C(v), i.e., the square root of the sum of the squared differences of C(u) and C(v).
"symdiff"
the cardinality of the symmetric set difference of the sets of coclassified pairs of distinct objects in the partitions. I.e., the number of distinct pairs of objects in the same class in exactly one of the partitions. (Alternatively, the cardinality of the symmetric set difference between the (binary) equivalence relations corresponding to the partitions.) For soft partitions, (currently) the symmetric set difference of the corresponding nearest hard partitions is used.
"Rand"
the Rand distance, i.e., the rate of distinct pairs of objects in the same class in exactly one of the partitions. (Related to the Rand index a via the linear transformation d = (1  a) / 2.) For soft partitions, (currently) the Rand distance of the corresponding nearest hard partitions is used.
"GV1"
the square root of the dissimilarity Delta_1 used for the first model in Gordon and Vichi (2001), i.e., the square root of the minimal sum of the squared differences of the matched nonzero columns of u and v.
"BA/d"
distance measures for hard partitions discussed in Boorman and Arabie (1972), with d one of A, C, D, or E. For soft partitions, the distances of the corresponding nearest hard partitions are used.
"BA/A"
is the minimum number of single element moves (move
from one class to another or a new one) needed to transform one
partition into the other. Introduced in Rubin (1967).
"BA/C"
is the minimum number of lattice moves for
transforming one partition into the other, where partitions are
said to be connected by a lattice move if one is just finer
than the other (i.e., there is no other partition between them) in
the partition lattice (see cl_meet
). Equivalently,
with z the join of x
and y
and S giving
the number of classes, this can be written as S(x) + S(y)  2
S(z). Attributed to David Pavy.
"BA/D"
is the “pairbonds” distance, which can be
defined as S(x) + S(y)  2 S(z), with z the meet of
x
and y
and S the supervaluation (i.e.,
nondecreasing with respect to the partial order on the partition
lattice) function ∑_i (n_i (n_i  1)) / (n (n  1)),
where the n_i are the numbers of objects in the respective
classes of the partition (such that n_i (n_i  1) / 2 are the
numbers of pair bonds in the classes), and n the total
number of objects.
"BA/E"
is the normalized information distance, defined as
1  I / H, where I is the average mutual information
between the partitions, and H is the average entropy of the
meet z of the partitions. Introduced in Rajski (1961).
(Boorman and Arabie also discuss a distance measure (B) based on the minimum number of set moves needed to transform one partition into the other, which, differently from the A and C distance measures is hard to compute (Day, 1981) and (currently) not provided.)
"VI"
Variation of Information, see Meila (2003). If
...
has an argument named weights
, it is taken to
specify case weights.
"Mallows"
the Mallowstype distance by Zhou, Li and
Zha (2005), which is related to the MongeKantorovich mass
transfer problem, and given as the pth root of the minimal
value of the transportation problem ∑ w_{jk} ∑_i
u_{ij}  v_{ik} ^ p with constraints w_{jk} ≥ 0,
∑_j w_{jk} = α_j, ∑_k w_{jk} = β_k,
where ∑_j α_j = ∑_k β_k. The parameters
p, α and β all default to one (in this
case, the Mallows distance coincides with the Manhattan
dissimilarity), and can be specified via additional arguments
named p
, alpha
, and beta
, respectively.
"CSSD"
the Cluster Similarity Sensitive Distance of
Zhou, Li and Zha (2005), which is given as the minimal value of
∑_{k,l} (1  2 w_{kl} / (α_k + β_l)) L_{kl},
where L_{kl} = ∑_i u_{ik} v_{il} d(p_{x;k}, p_{y;l}) with
p_{x;k} and p_{y;l} the prototype of the kth
class of x
and the lth class of y
,
respectively, d is the distance between these, and the
w_{kl} as for Mallows distance. If prototypes are matrices,
the Euclidean distance between these is used as default. Using
the additional argument L
, one can give a matrix of
L_{kl} values, or the function d. Parameters
α and β all default to one, and can be
specified via additional arguments named alpha
and
beta
, respectively.
For hard partitions, both Manhattan and squared Euclidean dissimilarity give twice the transfer distance (Charon et al., 2005), which is the minimum number of objects that must be removed so that the implied partitions (restrictions to the remaining objects) are identical. This is also known as the Rmetric in Day (1981), i.e., the number of augmentations and removals of single objects needed to transform one partition into the other, and the partitiondistance in Gusfield (2002), and equals twice the number of single element moves distance of Boorman and Arabie.
For hard partitions, the pairbonds (BoormanArabie D) distance is identical to the Rand distance, and can also be written as the Manhattan distance between the comembership matrices corresponding to the partitions, or equivalently, their symdiff distance, normalized by n (n  1).
If all components are hierarchies, available builtin methods for measuring dissimilarity between two hierarchies with respective ultrametrics u and v are as follows.
"euclidean"
the Euclidean dissimilarity of the ultrametrics (i.e., the square root of the sum of the squared differences of u and v).
"manhattan"
the Manhattan dissimilarity of the ultrametrics (i.e., the sum of the absolute differences of u and v).
"cophenetic"
1  c^2, where c is the cophenetic correlation coefficient (i.e., the productmoment correlation of the ultrametrics).
"gamma"
the rate of inversions between the ultrametrics (i.e., the rate of pairs (i,j) and (k,l) for which u_{ij} < u_{kl} and v_{ij} > v_{kl}).
"symdiff"
the cardinality of the symmetric set difference of the sets of classes (hierarchies in the strict sense) induced by the dendrograms. I.e., the number of sets of objects obtained by a split in exactly one of the hierarchies.
"Chebyshev"
the Chebyshev (maximal) dissimilarity of the ultrametrics (i.e., the maximum of the absolute differences of u and v).
"Lyapunov"
the logarithm of the product of the maximal and minimal ratios of the ultrametrics. This is also known as the “Hilbert projective metric” on the cone represented by the ultrametrics (e.g., Jardine & Sibson (1971), page 107), and only defined for strict ultrametrics (which are strictly positive for distinct objects).
"BO"
the m_δ family of tree metrics by Boorman and Olivier (1973), which are of the form m_δ = \int_0^∞ δ(p(h), q(h)) dh, where p(h) and q(h) are the hard partitions obtaining by cutting the trees (dendrograms) at height h, and δ is a suitably dissimilarity measure for partitions. In particular, when taking δ as symdiff or Rand dissimilarity, m_δ is the Manhattan dissimilarity of the hierarchies.
If ...
has an argument named delta
it is taken to
specify the partition dissimilarity δ to be employed.
"spectral"
the spectral norm (2norm) of the differences of the ultrametrics, suggested in Mérigot, Durbec, and Gaertner (2010).
The measures based on ultrametrics also allow computing dissimilarity
with “raw” dissimilarities on the underlying objects (R objects
inheriting from class "dist"
).
If a userdefined dissimilarity method is to be employed, it must be a function taking two clusterings as its arguments.
Symmetric dissimilarity objects of class "cl_dissimilarity"
are
implemented as symmetric proximity objects with selfproximities
identical to zero, and inherit from class "cl_proximity"
. They
can be coerced to dense square matrices using as.matrix
. It
is possible to use 2index matrixstyle subscripting for such objects;
unless this uses identical row and column indices, this results in a
(nonsymmetric dissimilarity) object of class
"cl_cross_dissimilarity"
.
Symmetric dissimilarity objects also inherit from class
"dist"
(although they currently do not “strictly”
extend this class), thus making it possible to use them directly for
clustering algorithms based on dissimilarity matrices of this class,
see the examples.
If y
is NULL
, an object of class
"cl_dissimilarity"
containing the dissimilarities between all
pairs of components of x
. Otherwise, an object of class
"cl_cross_dissimilarity"
with the dissimilarities between the
components of x
and the components of y
.
S. A. Boorman and P. Arabie (1972). Structural measures and the method of sorting. In R. N. Shepard, A. K. Romney, & S. B. Nerlove (eds.), Multidimensional Scaling: Theory and Applications in the Behavioral Sciences, 1: Theory (pages 225–249). New York: Seminar Press.
S. A. Boorman and D. C. Olivier (1973). Metrics on spaces of finite trees. Journal of Mathematical Psychology, 10, 26–59. doi: 10.1016/00222496(73)900035.
I. Charon, L. Denoeud, A. Guénoche and O. Hudry (2006). Maximum Transfer Distance Between Partitions. Journal of Classification, 23, 103–121. doi: 10.1007/s0035700600062.
W. E. H. Day (1981). The complexity of computing metric distances between partitions. Mathematical Social Sciences, 1, 269–287. doi: 10.1016/01654896(81)900421.
E. Dimitriadou, A. Weingessel and K. Hornik (2002).
A combination scheme for fuzzy clustering.
International Journal of Pattern Recognition and Artificial
Intelligence, 16, 901–912.
doi: 10.1142/S0218001402002052.
A. D. Gordon and M. Vichi (2001). Fuzzy partition models for fitting a set of partitions. Psychometrika, 66, 229–248. doi: 10.1007/BF02294837.
D. Gusfield (2002). Partitiondistance: A problem and class of perfect graphs arising in clustering. Information Processing Letters, 82, 159–164. doi: 10.1016/S00200190(01)002630.
N. Jardine and E. Sibson (1971). Mathematical Taxonomy. London: Wiley.
M. Meila (2003). Comparing clusterings by the variation of information. In B. Schölkopf and M. K. Warmuth (eds.), Learning Theory and Kernel Machines, pages 173–187. SpringerVerlag: Lecture Notes in Computer Science 2777.
B. Mérigot, J.P. Durbec and J.C. Gaertner (2010). On goodnessoffit measure for dendrogrambased analyses. Ecology, 91, 1850—1859. doi: 10.1890/091387.1.
C. Rajski (1961). A metric space of discrete probability distributions, Information and Control, 4, 371–377. doi: 10.1016/S00199958(61)800557.
J. Rubin (1967). Optimal classification into groups: An approach for solving the taxonomy problem. Journal of Theoretical Biology, 15, 103–144. doi: 10.1016/00225193(67)90046X.
D. Zhou, J. Li and H. Zha (2005). A new Mallows distance based metric for comparing clusterings. In Proceedings of the 22nd international Conference on Machine Learning (Bonn, Germany, August 07–11, 2005), pages 1028–1035. ICML '05, volume 119. ACM Press, New York, NY. doi: 10.1145/1102351.1102481.
cl_agreement
## An ensemble of partitions. data("CKME") pens < CKME[1 : 30] diss < cl_dissimilarity(pens) summary(c(diss)) cl_dissimilarity(pens[1:5], pens[6:7]) ## Equivalently, using subscripting. diss[1:5, 6:7] ## Can use the dissimilarities for "secondary" clustering ## (e.g. obtaining hierarchies of partitions): hc < hclust(diss) plot(hc) ## Example from Boorman and Arabie (1972). P1 < as.cl_partition(c(1, 2, 2, 2, 3, 3, 2, 2)) P2 < as.cl_partition(c(1, 1, 2, 2, 3, 3, 4, 4)) cl_dissimilarity(P1, P2, "BA/A") cl_dissimilarity(P1, P2, "BA/C") ## Hierarchical clustering. d < dist(USArrests) x < hclust(d) cl_dissimilarity(x, d, "cophenetic") cl_dissimilarity(x, d, "gamma")
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.