MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
50<br />
where M is the number of datapoints, i.e. x={x 1 , x 2 ,..., x M }, ( , )<br />
measure, e.g. the Euclidean distance.<br />
d x y is a distance<br />
Update Step (M-step): Each cluster’s parameters µ<br />
k, αk and σ are adjusted to<br />
k<br />
match the data points for which the cluster is responsible:<br />
σ<br />
2<br />
k<br />
=<br />
µ =<br />
k<br />
∑<br />
i<br />
∑<br />
i<br />
rx<br />
k<br />
i<br />
r<br />
k<br />
i<br />
∑r<br />
i<br />
N⋅∑<br />
i<br />
∑ ri<br />
i<br />
k<br />
∑∑<br />
α =<br />
( x − µ ) 2<br />
k<br />
i i k<br />
The α represents a measure of the likelihood that the Gaussian k (or cluster k)<br />
k<br />
generated the whole dataset.<br />
k<br />
i<br />
k<br />
i<br />
r<br />
k<br />
i<br />
r<br />
k<br />
i<br />
(3.7)<br />
(3.8)<br />
(3.9)<br />
This fits a mixture of spherical Gaussians. “Spherical” means that the variance of the Gaussians<br />
is the same in all directions. In other words, in the case of multidimensional dataset, the<br />
covariance matrix of the Gaussian is diagonal and isotropic (i.e. all the elements on the diagonal<br />
are equal and all elements on the off-diagonal are zero). This algorithm is still not good at<br />
modeling datasets spread along two elongated clusters, as shown in Figure 3-9. If we wish to<br />
model the clusters by axis-aligned Gaussians, we replace the assignment rule given by Equations<br />
(3.6) and (3.8) with the following:<br />
r<br />
k<br />
i<br />
=<br />
i=<br />
1<br />
k<br />
( 2πσ<br />
) i<br />
k '<br />
( 2πσi<br />
)<br />
⎛⎛ N<br />
⎜⎜ d k<br />
⎜⎜−∑<br />
⎜⎜ i=<br />
1 2<br />
⎝⎝<br />
( µ , x )<br />
k<br />
( σi<br />
)<br />
2 ⎞⎞<br />
i ⎟⎟<br />
2 ⎟⎟<br />
⎟⎟<br />
⎠⎠<br />
(3.10)<br />
⎛⎛ N<br />
2 ⎞⎞<br />
⎜⎜ d( µ k'<br />
, xi)<br />
⎟⎟<br />
⎜⎜−∑<br />
k<br />
2 ⎟⎟<br />
1<br />
⎜⎜ i=<br />
1 2( σi<br />
) ⎟⎟<br />
⎝⎝<br />
⎠⎠<br />
∑αk<br />
e<br />
N<br />
∏<br />
k '<br />
α<br />
k<br />
N<br />
∏<br />
i=<br />
1<br />
k<br />
j<br />
1<br />
µ =<br />
∑<br />
i<br />
∑<br />
i<br />
rx<br />
k<br />
i<br />
r<br />
e<br />
k<br />
i<br />
i,<br />
j<br />
(3.11)<br />
© A.G.Billard 2004 – Last Update March 2011