Transcript Diapositiva 1

Non Parametric Methods
Pattern Recognition and Machine
Learning
Debrup Chakraborty
Nearest Neighbor classification
Given:
Given a labeled sample of n feature vectors ( call X)
A distance measure (say the Euclidian Distance)
To find:
The class label of a given feature vector x
which is not in X
Nearest Neighbor classification
(contd.)
The NN rule:
Find the point y in X which is nearest to x
Assign the label of y to x
Nearest Neighbor classification
(contd.)
This rule allows us to partition the feature space into
cells consisting of all points closer to a given training
point x
All points in such cells are labeled by the class of the
training point. This partitioning is called a Voronoi
Tesselation
Nearest Neighbor classification (contd.)
Voronoi Cells in 2d
Nearest Neighbor classification
Complexity of the NN rule
Distance calculation
Finding the minimum distance
Nearest Neighbor classification
Nearest Neighbor Editing
X= Data set, n= no of training points, j=0
Construct the full Voronoi diagram for X
Do j=j+1, for each point xj in X
find Voronoi neighbors of xj
If any neighbor is not from the same class as xj
then mark xj
Until j==n
Discard all points that are not marked.
k nearest neighbor classification
Given:
Given a labeled sample of N feature vectors ( call X)
A distance measure (say the Euclidian Distance)
An integer k (generally odd)
To find:
The class label of a given feature vector x which
is not in X
k-NN classification (contd.)
Algorithm:
Find out the k nearest neighbors of x in X
Call them x1, x2 ,..., xk
Out of the k samples, let ki of them belong to class ci .
Choose that ci to be the class of x for which ki is
maximum
K-nn Classification
Class 1
Class 2
z
Class 3
k-NN classification (contd.)
Distance weighted nearest neighbor
xi , f ( xi )
n
i 1
Training set
In case x=xi,
return f(xi)
Given an instance x to be classified
Let x1, x2 ,..., xk be the nearest neighbors of x
Return
k
^
f ( x)  arg max  wi (c, f ( xi ))
cC
1 if a  b
 (a, b)  
0 otherwise
i 1
wi 
1
x  xi
2
Remarks on k-NN classification
•The distance weighted kNN is robust to noisy training data
and is quite effective when it is provided a sufficiently large
set of training examples.
•One drawbak of kNN method is that, it defers all
computation till a new querry point is presented. Various
methods have been developed to index the training
examples so that the nearest neighbor can be found with
less search time. One such indexing method is kd-tree
developed by Bently 1975
•kNN is a lazy learner
Locally Weighted Regression
In the linear regression problem, to find h(x) at a point x we
would do the following:
1. Minimize
2. Output
T x
1 n T
2
(

x

y
)
 i i
2 i 1
Locally Weighted Regression
In the llocally weighted regression problem we would do
the following
1. Minimize
2. Output
1 n
T
2
w
(

x

y
)
 i i i
2 i 1
T x
A standard choice of weights is

 xi  x 2 

wi  exp 
2

2



is called the bandwidth parameter
Clustering
Is different from Classification
Classification is partitioning
the feature space
whereas
Clustering is partitioning
the data into
“homogeneous groups”
Clustering is Unsupervised!!
K-means Clustering
Given: A data set


X  x1, x2 ,..., xn  R
p
Fix the number of clusters K
Let zk represent the i-th cluster center
i
(prototype) at the k-th iteration
k
S
Let j represent the j-th cluster at the k-th
iteration
K-means Clustering
Steps
1. Choose the initial cluster centers

Z 1  z11 , z21 ,..., z1K
2. At the k-th iterative step distribute the points in X in K
cluster using:
k
xi  S j
if
k
xi  z j

k
xi  zl
l  1,2,..., K
3. Compute
zkj 1

1
x
S kj xS jk
1
k
k 1
z

z
 th then the procedure has
4. If
 j
j
K j
converged else repeat from 2.
