Setup
=====
1. Parameters
Different parametrizations of LLGC were tested in this experiment setup:
- setup/weka29a
uses 0.01 for alpha
- setup/weka29b
uses 0.5 for alpha
- setup/weka29c
uses 0.99 for alpha (the one used in the paper)
all of these setups try also all permutations of these parameters:
- sigma: 0.25, 1, 4, 16
- including the # of attributes in the normalization of the affinity matrix: no, yes
In addition, IB1 and IBK (with option -K 9 and -X) are run for comparison.
2. Datasets
the datasets are split (randomly, stratified) with the ratio 5% training and 95% test
data.
Comparison
==========
The comparison spreadsheets contain the results of the three experiments
(alpha=[0.01, 0.5, 0.99]), the overview contains the best configuration, which
was with each alpha the first configuration (see Key "(1)").
The "Maxmimum" column at the end in the overview represents the maximum
accuracy that was achieved among the three configurations. The binary column
"Maximum?" just states, whether the classifier achieved the maximum accuracy or
not.
Evaluation
==========
In contrast to the paper, which uses 0.99 as alpha throughout the experiments,
a value of 0.01 seems to perform slightly better (reached the maximum 15 times,
in comparison to 13 times; see "comparison.html") on the binarized (i.e., only
instances with the 2 most common class labels were retained) UCI datasets. This
alpha value is in combination with sigma=0.25 and not including the number of
attributes in the normalization of the affinity matrix.
Reference(s)
============
"D. Zhou, O. Bousquet, T. N. Lal, J. Weston, and B. Schoelkopf. Learning
with local and global consistency. In 18th Annual Conf. on Neural
Information Processing Systems, 2003."
http://www.kyb.mpg.de/publication.html?publ=2293
http://eprints.pascal-network.org/00000476/01/LLGC.pdf
Key
===
(1) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 0.25 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance'
(2) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 0.25 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance -include-atts'
(3) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 1.0 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance'
(4) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 1.0 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance -include-atts'
(5) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 4.0 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance'
(6) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 4.0 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance -include-atts'
(7) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 16.0 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance'
(8) collective.functions.LLGC '-S 1 -folds 5 -alpha 0.01 -sigma 16.0 -limit 0 -repeats 0 -N 0 -distance weka.core.EuclideanDistance -include-atts'
(9) lazy.IB1 ''
(10) lazy.IBk '-K 9 -W 0 -X -A \"weka.core.LinearNN -A weka.core.EuclideanDistance\"'
Note: the key is the same for alpha 0.5 and 0.99 |