You are on page 1of 5

Multidimensional scaling

• Similar to PCA but takes


Multidimensional Scaling dissimilarity as input.
• Provide a visual representation of
the pattern of proximities among a
set of objects in a lower dimensional
space.
Violeta I. Bartolome
Senior Associate Scientist • Typically displayed on a 2-d plot
PBGB-CRIL
v.bartolome@cgiar.org

Example Multidimensional Scaling


Procedures
• The “points” that are
represented in • As long as the “distance” between
multidimensional space
can be anything. the objects can be assessed in some
• These objects may be fashion, MDS can be used to find the
genotypes, in which lowest dimensional space that still
case MDS can identify adequately capture the distances
clusters of genotypes
who are “close” versus between objects.
“distant” in terms of,
say, morphological • Once the number of dimensions is
characters. identified, a further challenge is
identifying the meaning of those
dimensions.
Multidimensional Scaling
Two types of MDS
Procedures
• Basic data representation in MDS is • Classical or metric
a dissimilarity matrix that shows the • Non-metric
distance between every possible
pair of objects.
The difference between the two is the
• The goal of MDS is to represent
stress function that is being minimized in
these distances with the lowest
the scaling procedure. Stress is a statistic
possible dimensional space.
for measuring goodness of fit.

Sample Data (used in PCA)

MDS in R
Read data to R
Distance matrix

Computes the Euclidean distances


between points.

Plot solution
Classical (metric) MDS

k is the number of
dimensions

cmdscale does not


give the stress
statistic
Non-metric MDS

• Options for distance matrix


o manhattan
o euclidean
o canberra
o bray
o kulczynski Default dimension (k) is 2
o jaccard
o gower
o altGower
o morisita
o horn
o mountford
o raup
o binomial
o chao

The non-metric fit is


based on stress values

The linear fit is the


correlation between
fitted values and
ordination distances
Large values indicate poor fit.
Thank you!

You might also like