Aaa ped-17-Unsupervised Learning: Dimensionality reduction

Unsupervised Learning:
Dimensionality reduction
AAA-Python Edition

Plan
●
1- Dimensionality reduction
●
2- Some Math
●
3- PCA
●
4- PCA in scikit-learn
●
5- Manifold Learning
●
6- Manifold Examples

3
1-Dimensionality
reduction
[By Amina Delali]
ObjectiveObjective
●
Dimensionality reduction in machine learning is reducing the
number of features of the training dataset.
●
This reduction is necessary to:
➢
Eliminate the noise from the data
➢
Visualize the data in 2 or 3 dimensions
➢
Speed up the learning process
➢
Enhance the learning results by eliminating correlated features.
➢
Eliminate unnecessary features.
➢
Compress the data size.
●
Two main approaches to dimensionality redcution are:
➢
Projection : project the data into a lower dimensional space.
➢
Manifold: suppose that the data in the higher dimension is just
a manifold of a representation of the data in the lower
dimension.

4
1-Dimensionality
reduction
[By Amina Delali]
ProjectionProjection
●
Sometimes the degree of the variation of the data is diferent
from one dimension to an other. So, for some features, the values
can be very diverse, an for others, they can barely change.
●
So we project the data into a lower dimension in order to keep
only the most infuential information ==> we defne a mapping
between the original data from the higher dimension to new data
in a lower dimension.
●
The most used technique to defne this mapping, is PCA (Principal
Component Analysis) and its variations:
➢
Incremental PCA
➢
Randomized PCA
➢
Kernel PCA

5
1-Dimensionality
reduction
[By Amina Delali]
ManifoldManifold
●
Like we said earlier, we make the hypothesis that our data is
created from a manifold of a data in a lower dimension. So,
reducing it to this low dimension is like straightening up this
manifold (or unrolling it).
●
The diferent techniques used, are:
➢
MDS: Multidimensional Scaling. Tries to preserve the distances
between instances.
➢
LLE: Locally Linear Embedding. Tries to preserve the relationship
between a sample and its closets points.
➢
Isomap: the samples will represent nodes of a graph. These
nodes are connected to their closets neighbors. The algorithm
tries to preserve the number of nodes in the shortest path
connecting two nodes.

6
2-SomeMath
[By Amina Delali]
Singular value decompositionSingular value decomposition
● It is the the decomposition of a matrix M (m,n)
into 3 matrices:
U(m,m)
, S(m,n)
, and V(n,n) .
Considering only real values, we have the
following characteristics:
➢
M = U .S . VT
( VT
is the transpose matrix of V : value at i,j
becomes at j,i )
➢ U . UT
= UT
. U = I(m,m)
(the identity matrix)
➢ V . VT
= VT
. VT
= I (n,n)
➢
The diagonal (values with the same row and column indices) of S
are the Singular values of M
➔
Singular values are the square roots of eigenvalues
➔
The other values of S are zeros.
➢
The columns of U are the eigenvectors of M . MT
.
➢ The columns of V are the eigenvectors of MT
. M .

7
2-SomeMath
[By Amina Delali]
Eigenvectors, EigenvaluesEigenvectors, Eigenvalues
● Given A (n,n)
a square matrix:
➔ If A . V(n)
= . V(n)
then: V is an eigenvector and is its
corresponding eigenvalue.
➔
The above equation can be rewritten as follow: (A- I). V= 0
➔
Several can solve the equation. For each lambda value,
an eigenvector is computed.
●
Example:
➢
If
●
Its eigenvalues will be: 1 , 3
●
And their corresponding eigenvectors will be: and
λλ λ
λλ
λ λ
λ
A=[2 1
1 2]
[ 1
−1] [1
1]

8
2-SomeMath
[By Amina Delali]
Standard DeviationStandard Deviation
●
The standard deviation measures how data is spread (or distant
from the mean) . It is the square root of the variance.
●
The variance is computed as follow:
➢
●
And the standard deviation:
●
To project data on new axis, we select the axis that preserve the
maximum possible variance of the data. This way, most of the
information is preserved.
variance=
∑
i=1
N
(xi−μ)2
N
σ
σ=√variance

9
3-PCAinscikit-learn
[By Amina Delali]
DefnitionDefnition
●
It is a linear dimensionality reduction technique that project data
using orthogonal axes (components) that preserve the maximum
variance possible. One of the method used is singular value
decomposition of the mean centered training data.
●
As stated before the decomposition leads to 3 matrices. The
vectors of the matrix VT
will be used to project the data. They are
the “principal components”.
●
Each component will conserve a certain amount of variance. The
variance obtained after projection is the accumulation of the
variances obtained by each component
●
To project, we select a sufcient number of component to preserve
the maximum of variance, then we apply the transformation (the
projection), using only this number of vectors.
●
The number of vectors will determine the dimension of the
projection.

10
3-PCAinscikit-learn
[By Amina Delali]
ExampleExample
●
Center the data to the
mean, before
applying the
decomposition
The
decomposition
To project, we multiply the
centered data by the first
selected component==> we will
have a 3 dimensions projection

11
3-PCAinscikit-learn
[By Amina Delali]
ResultsResults
●
Since our data was
originally labeled (we
don’t use those label for
decomposition), we used
them for colorizing the
data.
And what is obvious, is
that the data is clustered
according to its classes.
Which proofs:
●
that the clustering can
in certain cases classify
data.
●
the decomposition
preserved the most
important amount of
information.
3D projection
2D projection

12
4-ProcessingData
[By Amina Delali]
With matplotlibWith matplotlib
●
It tells to only
center the
data, and to
not
standardize
It will drop all the axis with
variance ratio < minfrac.
In this case, it will only
keep 2 axis.
Same results as in our
previous implementation

13
4-ProcessingData
[By Amina Delali]
With sklearnWith sklearn
●
We have to select
the number of
components before
transforming the
data
Comparing with matplotlib we see
that the directions are inverted
The reason of
this inversion is
that sklearn
flip the
eigenvector’s
sign before the
projection : it
apply the
method
svd_flip on the
vectors U and
V in the fitting
methods
As in matplotlib, we don’t
have to center the data

14
4-ProcessingData
[By Amina Delali]
Explained variance ratioExplained variance ratio
●
The correct number of components can be defned by the
explained variance ratio of each component.
●
It is computed by the value of explained variance divided by the
sum of all variances.
●
The ratio of each component are summed up until a certain
percentage is obtained.
●
The variances can be computed from the square of the singular
values in S

15
5-ManifoldLearning:
LLE
[By Amina Delali]
AlgorithmAlgorithm
●
LLE for Locally Linear Embeeding. The algorithm consist of 3
major steps:
● Step 1 - identifying the neighbors for each sample xi
from
the data X(N,D)
(for N samples and D features) :
➢ Compute the distances of the other samples from xi
➢ Select the k smallest distances.
● Step 2 - for each sample xi
compute its neighbors weights:
➢ Create the matrix Z(k,D)
with the k samples rows from X(N,D)
corresponding to the neighbors of xi
➢ Subtract xi
values from each row of Z(k,D)
➢ Compute C(k,k)
= Z(k,D)
. ZT
(D,k)
(
in the original page it is inverted because of X and Z are transposed)
➢ Compute the row i of the matrix W(N,N)
with:
➔ Compute the weights in the one column vector w(k,1)
that solve
the equation C(k,k)
. w(k,1)
= 1(k,1)
(1 is a column vector with only 1 as
values)

16
5-ManifoldLearning:
LLE
[By Amina Delali]
Algorithm (Suite)Algorithm (Suite)
➔ For the samples j that do not belong to each xi,
neighbors, set
the weights to 0.
➔ For each neighbor b of xi
set the weight to: w(p)
/sum(w(k,1)
). Where p is the indices in w corresponding to the
b neighbor of xi.
●
Step 3 – reduce the dimensionality to d < D in a new matrix
Y(N,d)
:
➢ Compute the matrix M(N,N)
= ( I(N,N)
– W(N,N)
)T
. (I(N,N)
– W(N,N)
)
➢ Select the d+1 eigenvectors of M(N,N)
corresponding to the d+1
smallest eigenvalues. Order these eigenvectors according to the
corresponding eigenvalues sorted in a decreasing order.
➢
For each column q in Y set the values equal to the values of the
q+1 smallest eigenvector counting from the bottom (to discard
the last eigenvector corresponding to the eigenvalue 0)

17
5-ManifoldLearning:
LLE
[By Amina Delali]
ExampleExample
●
N== 1500, D == 3
LLE : k == 12, d == 2

18
6-PolynomialRegression
[By Amina Delali]
AlgorithmAlgorithm
●
There are two types of Multidimensional Scaling: classical (or
metric) that tries to reproduce the original distances. The second
one is non-metric (NMDS) that tries to reproduces only the rank of
the distances.
●
We will describe the algorithm of the classical method using the
euclidean distance:
➢
Compute the distances between all points, and form a matrix of
those distances in a matrix D.
➢
Compute the matrix A as follow: A(i,j) = -1/2 * D(i,j)2
➢
Compute the matrix B as follow: B(i,j)= A(i,j)- A(i,.) - A(.,j) +A(.,.)
where: A(i,.) is the average of all A(i,j) for a selected i
A(.,j) is the average of all A(.,j) for a selected j
A(.,.) is the average of all values of A
➢
Find the p (the new dimension, lesser than the original
dimension ) largest eigenvalues of B:
and their corresponding normalized eigenvectors L1
,L2
, …, L p
so that Li
T
. Li
=
λ1>λ2>...>λp
λi

19
[By Amina Delali]
Algorithm (suite)Algorithm (suite)
➢ Form the matrix L as follow: L = (L1
, L2
, …, Lp
). The new values
(coordinates) are the rows of L.
●
This method minimizes the value of the Stress
●
The stress is a measure that can be used to fnd the optimal
lower dimension. It is computed as follow:
●
stress =
●
where: is the matrix of the distances of the new
matrix L
➢
A stress with a value < 0.05 is acceptable, below 0.01 is
considered to be good.
√
∑
i< j
(D(i, j)−Δ(i, j))2
∑
i< j
D(i , j)
2
Δ

20
[By Amina Delali]
Example in Scikit-learnExample in Scikit-learn
●
The results are completely
different from the previous
manifold technique. We
see here, the goal is to
keep the same original
distances values as much
as possible.

References
●
Aurélien Géron. Hands-on machine learning with Scikit-Learn and
Tensor-Flow: concepts, tools, and techniques to build intelligent
systems. O’Reilly Media, Inc, 2017.
●
J. D. Hunter. Matplotlib: A 2d graphics environment. Computing In
Science & Engineering, 9(3):90–95, 2007.
●
NCSS Statistical Software. Multidimensional Scaling, ncss, llc edition.
●
Scikit-learn.org. scikit-learn, machine learning in python. On-line at
https://scikit-learn.org/stable/. Accessed on 03-11-2018.
●
Jake VanderPlas. Python data science handbook: essential tools for
working with data. O’Reilly Media, Inc, 2017.
●
web.mit.edu. Singular value decomposition (svd) tutorial. On-line at
https://web.mit.edu/be.400/www/SVD/Singular_Value_Decompositio
n.htm. Accessed on 28-12-2018.
●
wikipedia.org. Wikipedia, the free encyclopedia. On-line at
https://www.wikipedia.org/. Accessed on 25-12-2018.

Aaa ped-17-Unsupervised Learning: Dimensionality reduction

More Related Content

What's hot (15)

Similar to Aaa ped-17-Unsupervised Learning: Dimensionality reduction (20)

More from AminaRepo (20)

Recently uploaded (20)

Aaa ped-17-Unsupervised Learning: Dimensionality reduction