On Sat, Jun 23, 2007 at 10:30:32AM +0100, Le Thi Kim wrote:
I have following questions:
1_ In LM 1, there is 1 layer missing in the model. What does it means?
Attribute selection for the linear models is performed by M5 when the
tree is being pruned. Only attributes tested in sub-trees are made
available to the linear regression at a leaf. That is, during pruning,
a decision is made between keeping a sub-tree and replacing it with a
linear regression built using only the attributes tested in the
sub-tree. Furthermore, the LinearRegression class used to produce the
linear regression functions in M5 does its own internal attribute
selection. It will drop attributes that are not useful in reducing the
squared error between the actual and predicted target. Finally, the
linear regression functions that appear at the leaves of an M5 tree
are "smoothed". That is, they are weighted combinations of the linear
regression functions that are at each node on the path from the leaf
in question to the root of the tree.
2- According to the binary (less than/ greater than) condition given
above, I generate the final out put map.
However, I don't understand how the M5 MT work out to assign the weight
for each layer. from the outset I thought that
the weight of these layers are assigned based on the predicted output and
the pattern (attribute's value) of the training set.
the M5 MT will find out the best combination to produce the final output
map that best matched with the predicted output
given in the learning stage. However, I found out that some layers
in these sub models were given very high weights but
their attribute's value in these sub spaces were "0". I am wondering if
the value in these sub-spaces equal "0",
the M5 MT should assigned the weight "0" instead of very high weights
which It did not connect to any binary condition
given in previous sub-spaces. the weights of these layers in this case had
no meaning at all.
Standard least-squares linear regression (with attribute selection) is
used to produce the linear functions. The fact that smoothed models
are reported at the leaves probably explains the case where there are
non-zero coefficients for attributes that appear to have all zero
values in the sub-space of the leaf - these attributes will have
non-zero values in the sub-spaces corresponding to the nodes higher up
in the tree.
Hope this helps.
Department of Computer Science
University of Waikato