Identifying Saxophonists from Their Playing Styles - Intelligent Music Information Systems: Tools and Methodologies - page 114

of the predicted cluster, the more the score of the
interpreter is increased). Finally, the interpreter
with the higher score is returned.
Clearly, the classifiers CL i play a central role
in the output of classifier MC . For each inter-
preter, CL i is trained with data extracted from the
interpreter's performance recordings. We have
explored different classifier induction methods
for obtaining each classifier CL i . The methods
we have considered are:

linear in a higher dimension space). Thus,
they provide a flexible prediction, but with
a higher computational cost necessary to
perform all the computations in the higher
dimensional space. The classification accu-
racy of SVM largely depends on the choice
of the kernel evaluation function and the
parameters which control the amount to
which deviations are tolerated (denoted by
epsilon). In this chapter we have explored
SVM with linear and polynomial kernels
(2nd, 3rd and 4th order) and we have set
epsilon to 0.05.
• Artificial Neural Networks (ANN): ANN
learning methods provide a robust approach
to approximating a target function. In this
chapter we apply a gradient descent back
propagation algorithm (Chauvin et al., 1995)
to tune the neural network parameters to
best fit the fMRI training set. The back
propagation algorithm learns the weights for
a multi layer network, given a network with
a fixed set of units and interconnections. We
set the momentum applied to the weights
during updating to 0.2 and the learning rate
(the amount the weights are updated) to 0.3.
We use a fully-connected multilayer neural
network with one hidden layer (one input
neuron for each attribute and one output
neuron for each class).

• K-means Clustering: Clustering techniques

apply when there is no class to be predicted
but rather when the instances are to be
divided into natural groups. In k-means
clustering (k is the number of clusters),
k points are chosen at random as cluster
centers, each instance is assigned to the
nearest cluster center, for each cluster a new
cluster center is computed by averaging over
all instances in the cluster, and the whole
process is repeated with the new cluster
centers. Iteration continues until the same
instances are assigned to each cluster in
consecutive rounds. In this chapter, we apply
fuzzy k-means clustering where Instances
can belong to several clusters with different
“degrees of membership”.

• Decision Trees: A decision tree classifier

recursively constructs a tree by selecting at
each node the most relevant attribute. This
process gradually splits up the training set
into subsets until all instances at a node have
the same classification. The selection of the
most relevant attribute at each node is based
on the information gain associated with each
node of the tree (and corresponding set of
instances). We have applied the decision tree
building algorithm C4.5 (Quinlan, 1993).

• Lazy Methods: Lazy Methods are based

on the notion of lazy learning which sub-
sumes a family of algorithms that store the
complete set of given (classified) examples
of an underlying example language and
delay all further calculations until requests
for classifying yet unseen instances are
received. In this chapter we have explored
the k-Nearest Neighbor (k-NN) algorithm
(with k ∈ {1,2,3,4,7}) which is capable of
handling noisy data well if the training set
has an acceptable size. However, k-NN does
not behave well in the presence of irrelevant
attributes.

• Support Vector Machines (SVM): SVM

(Cristiani & Shawe-Taylor, 2000) take great
advantage of using a nonlinear attribute
mapping that allows them to be able to pre-
dict non-linear models (though they remain