Page 337 - Contributed Paper Session (CPS) - Volume 6
P. 337
CPS1950 Paolo G. et al.
2. Methodology
In this section we start by reviewing RKM and FKM. Later, the new
clustering procedure is illustrated in detail.
a. Reduced K-Means (RKM):
Let X be the data matrix of order (I × J) containing the scores of I observation
units with respect to J variables. The Reduced K-means (RKM) analysis (De
Soete & Carroll, 1994) can be formulated as:
X = UFA + E (1)
where U, of order (I × K), is the membership matrix with elements equal to 0
or 1 expressing for each observation units the membership to one of the K
clusters. Note that U is row-stochastic, that is, its row-wise sum is equal to 1.
A is the component weight matrix of order (J × Q). It is column-wise
orthonormal, i.e., AA = IQ, being IQ the identity matrix of order Q, and every
column expresses the weights of the variables on the corresponding
component. Finally, F is the centroid matrix of order (K × Q) such that every
row refers to a cluster centroid. The centroids lie in the reduced subspace
spanned by the columns of A. Finally, E is the residual matrix having the same
order of X. The optimal parameter matrices U, F and A are obtained in the
least square sense by minimizing the residual sum of squares:
2
2
fRKM = || E || = || X UFA || , (2)
being || · || the Frobenius norm of matrices. Suitable Alternating Least Squares
(ALS) algorithms can be adopted for the minimization of (2). The RKM solution
is not unique. Equally fitting solutions can be found up to rotational
indeterminacy for the weights in A and the centroids in F and cluster label
switching. Given an orthonormal rotation matrix R of order (Q × Q) and a
permutation matrix P of order (K × K), letting A* = AR, U* = UP and F* = PFR,
we have U*F*A* = UFA.
When Q = J, i.e., when the variable space is not reduced through PCA (A = IJ,
where IJ is the identity matrix of order J), RKM boils down to the standard KM
algorithm.
b. Factorial K-Means (FKM):
The RKM loss function in (2) is a proxy of the within-cluster sum of squares
in the reduced space. In fact, it is the sum of the squared distances between
the observation units in the (J-dimensional) observed space and the centroids
in the (Q-dimensional) reduced space. This represents a sort of idiosyncrasy
because it appears more reasonable to compute the within-cluster sum of
squares in the reduced space by considering not only the centroids but also
the observation units lying in the reduced space. This motivation leads to the
so-called Factorial K-Means (FKM) procedure developed by Vichi & Kiers
(2001). The FKM model is expressed as
XAA = UFA + E. (3)
326 | I S I W S C 2 0 1 9