Unsupervised Learning

吉建民

USTC
[email protected]

2023 年 6 月 4 日

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Used Materials

Disclaimer: 本课件采用了 S. Russell and P. Norvig’s Artificial

Intelligence –A modern approach slides, 徐林莉老师课件和其他网
络课程课件，也采用了 GitHub 中开源代码，以及部分网络博客
内容

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Table of Contents

Unsupervised Learning
Clustering
Principle Component Analysis

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Supervised learning has many successes

▶ Document classification
▶ Protein prediction
▶ Face recognition
▶ Speech recognition
▶ Vehicle steering etc.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
However...

▶ Labeled data can be rare or expensive in many real

applications

- Speech
- Medical data
- Protein
- ···

▶ Unlabeled data is much cheaper and abundant

Question: Can we use unlabeled data to help?

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Unsupervised learning

Learning from unlabeled data (without supervision)

▶ What can we predict from unlabeled

data?
▶ Groups or clusters in the data

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Unsupervised learning

Learning from unlabeled data (without supervision)

▶ What can we predict from unlabeled

data?
▶ Groups or clusters in the data
▶ Density estimation (密度估计)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Unsupervised learning

Learning from unlabeled data (without supervision)

▶ What can we predict from unlabeled

data?
▶ Groups or clusters in the data
▶ Density estimation (密度估计)
▶ Low-dimensional structure
▶ Principal Component Analysis 主元分
析 (PCA) (linear)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Unsupervised learning

Learning from unlabeled data (without supervision)

▶ What can we predict from unlabeled

data?
▶ Groups or clusters in the data
▶ Density estimation (密度估计)
▶ Low-dimensional structure
▶ Principal Component Analysis 主元分
析 (PCA) (linear)
▶ Manifold learning 流行学习
(non-linear)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Table of Contents

Unsupervised Learning
Clustering
Principle Component Analysis

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Clustering

▶ Are there any “groups” in the data ?

▶ What is each group ?
▶ How many ?
▶ How to identify them?

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Clustering
▶ Group the data objects into subsets
or “clusters”:
▶ High similarity within clusters
▶ Low similarity between clusters

▶ A common and important task

that finds many applications in
Science, Engineering, information
Science, and other places
▶ Group genes that perform the
same function
▶ Group individuals that has similar
political view
▶ Categorize documents of similar
topics
▶ Identify similar objects from
pictures

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Clustering

▶ Input: training set of input point

Dtrain = {x1 , . . . , xn }
▶ Output: assignment of each point to a cluster

( C(1), . . . , C(n) ) where C(i) ∈ { 1, . . . , k }

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means clustering

Create centers and assign points to centers to minimize sum of

squared distance

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means objective

▶ Each cluster is represented by a centroid µ

▶ Encode each point by its cluster center, pay a cost for
deviation
▶ Loss function based on reconstruction
∑
n
2
Losskmeans µC(j) − xj
j=1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means algorithm

∑n 2
▶ Goal: minµ minC j=1 µC(j) − xj

▶ Strategy: alternating minimization

▶ Step 1: if know cluster centers µ, can find best C
▶ Step 2: if know cluster assignments C, can find best cluster
centers

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means algorithm
Optimize loss function Loss(µ, C)

∑
n
2
min min µC(j) − xj
µ C
j=1

(1) Fix µ, optimize C

∑
n
2
min µC(j) − xj
C(1),C(2),...,C(n)
j=1

Assign each point to the nearest cluster center

(2) Fix C, optimize µ

∑
n
2
min µC(j) − xj
µ(1),µ(2),...,µ(k)
j=1

Solution: average of points in cluster i, exactly second step

(re-center) .
.
.
.
.
. . . . .
. . . .
. . . .
. . . .
. . . .
. . . . .
.
.
.
.
.
.
.
.
.
K-Means

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means clustering: Example

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means clustering: Example

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means clustering: Example

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means clustering: Example
Repeat until convergence

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Properties of K-means algorithm

▶ Guaranteed to converge in a finite number of iterations

▶ To a local minimum
▶ The objective is non-convex, so coordinate descent on is not
guaranteed to converge to the global minimum
▶ Running time per iteration: simple and efficient
▶ Assign data points to closest cluster center

O(KN)

▶ Change the cluster center to the average of its assigned points

O(N)

▶ Different initialization will lead to different results

▶ K-means problem is NP-hard （之前公式的最优解）
▶ Not robust to noise and outliers
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means convergence

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means getting stuck

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
K-means not able to properly cluster

Changing the features (distance function) can help

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Table of Contents

Unsupervised Learning
Clustering
Principle Component Analysis

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Principle component analysis

▶ What is dimensionality reduction?

▶ Why dimensionality reduction?
▶ Principal Component Analysis (PCA)
▶ Nonlinear PCA using Kernels

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
What is dimensionality reduction?

▶ Dimensionality reduction refers to the mapping of the original

high-dimensional data onto a lower-dimensional space.
- Criterion for dimensionality reduction can be different based on
different problem settings.
▶ Unsupervised setting: minimize the information loss
最近重构性：样本点到这个超平面的距离都足够近
▶ Supervised setting: maximize the class discrimination
最大可分性：样本点在这个超平面上的投影能尽可能分开
▶ 对样本进行中心化处理以后，两者等价

▶ Given a set of data points of d dimension variables

{x1 , x2 , . . . , xn }
▶ Compute the linear transformation (projection)

P ∈ Rd×m : x ∈ Rd → y = P⊤ x ∈ Rm (m << d)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
What is dimensionality reduction?

P ∈ Rd×m : x ∈ Rd → y = P⊤ x ∈ Rm

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
High-dimensional data

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Why dimensionality reduction?

▶ Most machine learning and data mining techniques may

not be effective for high-dimensional data
▶ Curse of Dimensionality
▶ Query accuracy and efficiency degrade rapidly as the dimension
increases.
▶ The intrinsic dimension may be small.
▶ For example, the number of genes responsible for a certain
type of disease may be small.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Curse of Dimensionality (维数灾难)
▶ When dimensionality increases, data
becomes increasingly sparse in the space
that it occupies
▶ Definitions of density and distance
between points, which is critical for
clustering and outlier detection, become
less meaningful
▶ If N1 = 100 represents a dense sample for
a single input problem, then N10 = 10010
is the sample size required for the same ▶ Randomly generate
sampling density with dimension 10. 500 points
▶ The proportion of a hypersphere (超球面) ▶ Compute difference
with radius r and dimension d, to that of a between max and
hyercube (超立方体) with sides of length min distance
2r and dimension d converges to 0 as d between any pair of
goes to infinity —nearly all of the points
high-dimensional space is “far away” from
the center . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
High dimensional spaces are empty

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lost in space
Let’s consider a hypersphere of radius r inscribed in a hypercube
with sides of length 2r. Then take the ratio of the volume (体积)
of the hypersphere to the hypercube. We observe the following
trends.
▶ in 2 dimensions:
V(S2 (r)) πr2
= 2 = 78.5%
V(H2 (2r)) 4r

▶ in 3 dimensions:
4
V(S3 (r)) πr3
= 3 3 = 52.4%
V(H3 (2r)) 8r

▶ when the dimensionality d increases

asymptotically

V(Sd (r)) π d/2

lim = lim d d →0
d→∞ V(Hd (2r)) d→∞ 2 Γ( + 1)
2
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Why dimensionality reduction?

▶ Visualization: projection of high-dimensional

data onto 2D or 3D.
▶ Data compression: efficient storage and retrieval
▶ Noise removal: positive effect on query accuracy.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Application of feature reduction

▶ Face recognition
▶ Handwritten digit recognition
▶ Text mining
▶ Image retrieval
▶ Microarray data analysis
▶ Protein classification
▶ ···

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
What is Principal Component Analysis?

▶ Principal component analysis (PCA)

- Reduce the dimensionality of a data set by finding a new set of
variables, smaller than the original set of variables
- Retains most of the sample’s information.
- Useful for the compression and classification of data.
▶ By information we mean the variation present in the sample,
given by the correlations between the original variables.
▶ The new variables, called principal components (PCs), are
uncorrelated, and are ordered by the fraction of the total
information each retains.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Principal components (PCs)

Given n points in a d dimensional space, for large d, how does one

project on to a low dimensional space while preserving broad trends
in the data and allowing it to be visualized?

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Geometric picture of principal components

▶ Given n points in a d dimensional space, for large d, how does

one project on to a 1 dimensional space

▶ Choose a line that fits the data so the points are spread out
well along the line

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Let us see it on a figure

PCA 希望降维后信息损失最小，可以理解为投影后的数据尽可能的分
开，这种分散程度可以用方差来表示 (µ 为均值)：

1∑
n
Var(a) = (ai − µ)2
n i=1

对数据进行中心化后，即 µ = 0：

1∑ 2
n
Var(a) = a
n i=1 i
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Geometric picture of principal components
对数据进行中心化:

1∑
n
x̄ = xi ,
n
i=1
x′i = xi − x̄, 1 ≤ i ≤ n.

对于中心化以后的数据，即 x̄′ = 0，以下说法等价: Find a line

that
▶ maximize the variance of the projected data
▶ maximize the sum of squares of data samples’ projections on
that line
▶ minimize the sum of squares of distances to the line

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic Interpretation — 1D

▶ Minimizing sum of squares of distances to the line is the same

as maximizing the sum of squares of the projections on that
line, thanks to Pythagoras (毕达哥拉斯).

投影长度为: x⊤ ∥w∥
w

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic Interpretation — 1D

投影长度为: x⊤ u = u⊤ x subject to u⊤ u = 1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Geometric picture of principal components

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Geometric picture of principal components

▶ the 1st PC u1 is a minimum distance fit to a line in X space

▶ the 2nd PC u2 is a minimum distance fit to a line in the
plane perpendicular (垂直于) to the 1st PC
PCs are a series of linear least squares fits to a sample, each
orthogonal (垂直于) to all the previous.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic derivation of PCs

▶ Given a sample of n observations on a vector of d variables

{ x1 , x2 , . . . , xn } ∈ Rd

▶ First project the data onto a one-dimensional space with a d

-dimensional vector u1 (where u⊤ 1 u1 = 1):
{ }
u⊤ ⊤ ⊤
1 x1 , u1 x2 , · · · , u1 xn

▶ Find u1 to maximize the variance the projected data:

1 ∑( ⊤ )2
n
u1 x i − u⊤
1 x̄ = u⊤
1 Su1
n
i=1
∑n ∑n
Where x̄ = 1
n i=1 xi and S = 1
n i=1 (xi − x̄) (xi − x̄)⊤

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic derivation of PCs
▶ To solve maxu1 u⊤ ⊤
1 Su1 subject to u1 u1 = 1
▶ Let λ1 be a Lagrangian multiplier (拉格朗日乘子)
( )
L = u⊤1 Su 1 + λ1 1 − u ⊤
1 u1

∂L
= Su1 − λ1 u1 = 0
∂u1
Su1 = λ1 u1
⇒ u1 is an eigenvector (特征向量)
u⊤
1 Su1 = λ1
⇒ u1 corresponds to the eigenvector with the largest eigenvalue λ1

▶ 即，maxu1 u⊤ ⊤
1 Su1 subject to u1 u1 = 1 的结果就是矩阵 S
的最大特征值
▶ 矩阵 S 特征值计算方法：构造特征多项式 | S − λI | = 0 (I 为
单位矩阵)，特征值为线性方程组的解
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic derivation of PCs

▶ To find the second component u2

▶ Solve the following

max u⊤ ⊤ ⊤
2 Su2 subject to u2 u2 = 1 & u1 u2 = 0
u2

- u2 is the eigenvector with the second largest eigenvalue λ2

···

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic derivation of PCs
▶ Main steps for computing PCs
▶ Calculate the covariance matrix S

1∑
n
⊤
S= (xi − x̄) (xi − x̄)
n i=1

or first center the data: { x′1 , x′2 , . . . , x′n } and x̄′ = 0

[ ] 1
let X = x′1 , x⊤ ′
2 , . . . , xn ∈ R
d×n
; then S = XX⊤
n
m
▶ Find the first m eigenvectors {ui }i=1
▶ Form the projection matrix

P = [u1 u2 · · · um ] ∈ Rd×m

▶ A new test point can be projected as:

xnew ∈ Rd → P⊤ xnew ∈ Rm
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Algebraic derivation of PCs

y = P⊤ x ∈ Rm

▶ Getting the old data back?

- If P is a square matrix (方阵), we can recover x by
( )−1
x = P⊤ y = Py = PP⊤ x

注：u⊤ ⊤ ⊤
i ui = 1 and ui uj = 0 for i ̸= j, then P P = Im (where
⊤ −1
m = n) and (P ) = P
▶ Here P is not full (m << d), but we can still recover x by
x = Py = PP⊤ x, and lose some information
▶ Objective:
▶ Lose least amount of information

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Optimality property of PCA

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Optimality property of PCA
Main theoretical result:
The matrix P consisting of the first m eigenvectors of the
covariance matrix S solves the following min problem:

Notice that, for a matrix A m × n and B n × m,

∑m ∑n
trace(AB) = trace(BA) = i=1 j=1 aij bji
∑d ∑n ∑d ∑n
arg minP i=1 j=1 (xij − x′ij )2 is equivalent to arg maxP i=1 j=1 xij x′ij ,
∑d ∑n
as i=1 j=1 x′ij = trace((PPT X)T PPT X) = trace(XT PPT X)
2

PCA projection minimizes the reconstruction error among all linear

projections of size m. .
. .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
PCA for image compression

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Nonlinear PCA using Kernels

Rewrite PCA in terms of dot products

∑
▶ Assume the data has been centered, i.e., i xi =0
▶ 1∑ ⊤
The covariance matrix S can be written as S = n i xi xi
▶ If u is an eigenvector of S corresponding to nonzero eigenvalue
1∑ ⊤ 1 ∑( ⊤ )
Su = xi xi u = λu ⇒ u = xi u xi
n nλ
i i

▶ Eigenvectors of S lie in the space spanned by all data points

Kernel methods:
▶ denote the representation of x as φ(x)
▶ define the kernel function k : X × X → R by
k(xi , xj ) = φ(xi )⊤ φ(xj )
▶ define the kernel matrix K: Kij = k(xi , xj )
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Nonlinear PCA using Kernels

1∑ ⊤ 1 ∑( ⊤ )
Su = xi xi u = λu ⇒ u = xi u xi
n nλ
i i

The covariance matrix can be written in matrix form

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Nonlinear PCA

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Comments on PCA

▶ Linear dimensionality reduction method

▶ Can be kernelized
▶ Many nonlinear dimensionality reduction methods (Isomap,
graph Laplacian eigenmap, and locally linear embedding/LLE)
can be described as kernel PCA with a special kernel

▶ Non-convex optimization problem

▶ But easy to solve…

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Want to Learn More?
▶ Machine Learning: a Probabilistic Perspective, K. Murphy
▶ Pattern Classification, R. Duda, P. Hart, and D. Stork.
Standard pattern recognition textbook. Limited to
classification problems. Matlab code.
http://rii.ricoh.com/~stork/DHS.html
▶ Pattern recognition and machine learning. C. Bishop
▶ The Elements of statistical Learning: Data Mining, Inference,
and Prediction. T. Hastie, R. Tibshirani, J. Friedman,
Standard statistics textbook. Includes all the standard
machine learning methods for classification, regression,
clustering. R code. http://www-stat-class.stanford.
edu/~tibs/ElemStatLearn/
▶ Introduction to Data Mining, P.-N. Tan, M. Steinbach, V.
Kumar. AddisonWesley, 2006
▶ Principles of Data Mining, D. Hand, H. Mannila, and P.
Smyth. MIT Press, 2001
▶ 统计学习方法，李航 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Machine Learning in AI

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Machine Learning History

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Summary

▶ Supervised learning
▶ Learning Decision Trees
▶ K Nearst Neighbor Classfier
▶ Linear Predictions
▶ Support Vector Machines
▶ Unsupervised learning
▶ Clustering
▶ Principle Component Analysis

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
作业

▶ K-means 算法是否一定会收敛? 如果是，给出证明过程；如

果不是，给出说明。

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .