AI 330: Machine Learning

Fall 2023

• Dr. Wessam EL-Behaidy

• Associate Professor, Computer Science Department,
• Faculty of Computers and Artificial Intelligence,
• Helwan University.
•
Dr. Ensaf Hussein
Associate Professor, Computer Science Department,
Faculty of Computers and Artificial Intelligence,
Helwan University.
Course Map Data

Numerical Text Image/video Audio

Preprocessing Preprocessing

Feature
ML Algorithms
Extraction

Evaluation

2
 Optimization ML Algorithms

Normal
Gradient Descent Supervised Unsupervised
Equations

Dimensionality
Regression Classification Clustering
Reduction

Logistic
Linear Regression K-means PCA
regression

Univariate Multivariate KNN

ML Algorithms Map Ensemble

Models

3
 Lecture 7
k-Nearest Neighbor (kNN)

Slides of:
Machine Learning Specialization https://www.coursera.org/specializations/machine-learning-introduction
at Stanford University (Andrew Ng)

4
Grading Distribution (Update)
 Project: (20%)
• Code “github”, pitching video, presentation.
 Programming Assignment (10%)- DataCamp track
 Midterm exam (20%)
 Final exam (50%)
k-NN Intuition
 What k-NN does for you?
Before K-NN After K-NN
X2 x2

Category 2 Category 2

New data point K-N N New data point assigned to

Category 1

Category 1 Category 1

x1 x1
How did it do that?
STEP 1: Choose the number K of neighbors

STEP 2: Take the K nearest neighbors of the new data point, according to the
Euclidean distance

STEP 3: Among these K neighbors, count the number of data points in each
category

STEP 4: Assign the new data point to the category where you counted the most
neighbors

Yo u r M o d e l is R e a d y
K-NN Algorithm
• ST E P 1: Choose the number K of neighbors: K = 5
x2

Category 2

New data point

Category 1

x1
Euclidean Distance
STEP 2: Take the K nearest neighbors of the new data point, according to the
Euclidean distance
y

y2 P 2(x 2,y2)

y1
P 1(x 1,y1)

x1 x2
x
K-NN Algorithm
STEP 3: Among these K neighbors, count the number of data points in each category
STEP 4: Assign the new data point to the category where you counted
x2 the most neighbors

Category 1: 3 neighbors
Category 2
Category 2: 2 neighbors
New data point

Category 1

x1
Simple Idea …
• Store all training examples
Unknown record

− Each point is a “vector” of

attributes
• Classify new examples
based on most “similar”
training examples
− Similar means “closer” in
vector space

What’s done in training?

Two Approaches for Learning

Eager learning Lazy learning

(e.g., logistic regression) (e.g., kNN)
• Learn/Train • Learn
− Induce an abstract model from − Just store data in memory
data • Test/Predict/Classify
• Test/Predict/Classify − Compare new data to stored
− Apply learned model to new data
data • Properties
− Retains all information seen in
training
− Complex hypothesis space
− Classification can be very slow
kNN Algorithm Dataset, number of neighbors, example to predict its label

// N number of examples that we have in dataset

Example BRIGHTNESS SATURATION CLASS
40 20 Red
BRIGHTNESS SATURATION CLASS 50 50 Blue
20 35 ? 60 90 Blue
10 25 Red
Distance #1
70 70 Blue
For the first row, d1:
60 10 Red
d1 = √(20 - 40)² + (35 - 20)²
25 80 Blue
= √400 + 225
= √625
= 25 BRIGHTNESS SATURATION CLASS DISTANCE
10 25 Red 10
5 as the value of K, we'll only consider the first five rows. That is:
40 20 Red 25
50 50 Blue 33.54
The majority class within the 5 nearest neighbors to the new
entry is Red. Therefore, we'll classify the new entry as Red. 25 80 Blue 45
60 10 Red 47.17
70 70 Blue 61.03
60 90 Blue 68.01

15
 k-Nearest Neighbors (kNN)
The k-nearest neighbor classifier will take a test sample, compare it to every single
one of the training samples, and predict the label of the k closest training samples.

X X X

(a) 1-nearest neighbor (b) 2-nearest neighbor (c) 3-nearest neighbor

Variants of kNN
Weighted Voting
• Default: all neighbors have equal weight
• Extension: weight votes of neighbors by (inverse) distance
• The intuition behind weighted kNN, is to give more weight to the points
which are nearby and less weight to the points which are farther away.

Epsilon-Ball Nearest Neighbors

• Same general principle as K-NN, but change the method for
selecting which training examples vote
• Instead of using K nearest neighbors, use all examples 𝑥 such that
𝑑 𝑖 𝑠 𝑡 𝑎 𝑛 𝑐 𝑒 ( 𝑥ෝ , 𝑥 ) ≤ 𝜀
17
 Weighted k-NN Example
• The item to predict is set as [0.62, 0.35]
(colored yellow) .
• The labels can be 0, 1, or 2 (colored red,
blue, green).
• The k = 6 closest data points and their
distances are:

https://jamesmccaffrey.wordpress.com/2022/05/03/weighte
The sum of the inverse distances is 14.1421 + . . + 6.3246 = 60.8834. d-k-nearest-neighbors-classification-example-using-python/

18
Weighted k-NN Example
• The voting weights are the inverse distances divided by their sum:
class = 0 weight = 14.1421 / 60.8834 = 0.2323
class = 0 weight = 13.1306 / 60.8834 = 0.2157
class = 1 weight = 10.8465 / 60.8834 = 0.1782
class = 2 weight = 8.9443 / 60.8834 = 0.1469
class = 0 weight = 7.4953 / 60.8834 = 0.1231
class = 2 weight = 6.3246 / 60.8834 = 0.1039

• The voting for each class is:

class 0 : 0.2323 + 0.2157 + 0.1231 = 0.5711
class 1 : = 0.1782
class 2 : 0.1469 + 0.1039 = 0.2508

• Because class 0 has the most voting weight, the prediction is that the unlabeled
data item is class 0.

19
K-NN Components

20
Components of a KNN Classifier
• Distance metric
− How do we measure distance between instances?
− Determines the layout of the example space

• The k hyperparameter
− How large a neighborhood should we consider?
− Determines the complexity of the hypothesis space

Very problem-dependent.
Must try them all out and see what works best.
Distance Metrics
• We can use any distance function to select nearest
neighbors.
• Different distances yield different neighborhoods
L2 distance L1 distance
( = Euclidean distance) Max norm
( = Manhattan distance)
Manhattan Distance (L1)
• The Manhattan distance in a 2-dimensional space is given as:
𝑑1 𝑝, 𝑞 = 𝑝1 − 𝑞1 + 𝑝2 − 𝑞2
• For instance:
If we have two points A (3,6) and B (-5,4).
 L1(A,B)= |3-(-5)|+|6 – 4|= 8 + 2= 10.

• And the generalized formula for an n-dimensional space is given as:

𝑛

𝑑1 (𝑝, 𝑞) = ෍ 𝑝𝑖 − 𝑞𝑖
𝑖=1
• Where, n = number of dimensions and pi, qi = data points

23
Euclidean Distance (L2)
• The Euclidean distance in a 2-dimensional space is given as:
𝑑2 𝑝, 𝑞 = 𝑝1 − 𝑞1 2 + 𝑝2 − 𝑞2 2
• For instance:
If we have two points A (3,6) and B (-5,4).
 L2 (A,B)= [(3-(-5))2+ (6 – 4)2]1/2 = [16 + 4]1/2= 4.472.

• And the generalized formula for an n-dimensional space is given as:

𝑛

𝑑2 (𝑝, 𝑞) = ෍ 𝑝𝑖 − 𝑞𝑖 2

𝑖=1

• Where, n = number of dimensions and pi, qi = data points

24
Distance Metric to compare images
• Manhattan distance (L1): Compare the images pixel by pixel and add
up all the absolute differences.
• Given two images I1 and I2
Distance Metric to compare images
• Euclidean distance (L2): Computing the pixel-wise difference as
before, but this time we square all of them, add them up and finally
take the square root.

• Given two images I1 and I2:

 The k hyper-parameter

Tunes the complexity of the

hypothesis space
− If k = 1, every training
example has its own
neighborhood
− If k = N, the entire feature
space is one neighborhood!

• Higher k yields smoother

decision boundaries
How would you set k in practice?

27
 Setting Hyperparameters
Setting Hyperparameters

60% 20% 20%

Validation Set
• The idea is to split the training set in two sets: a slightly smaller
training set, and what we call a validation set.

• This validation set is essentially used as

• a fake test set to tune the hyper-parameters.

• After we found the best hyper-parameters on validation set,

• Stick with these values and
• Evaluate once on the actual test set (for measuring the generalization of
your classifier).
Cross-Validation
• For small datasets, sometimes we use a more sophisticated
technique for hyperparameter tuning called cross-validation.

• Instead of arbitrarily picking the first data points to be the

validation set and rest training set,
• Get a better and less noisy estimate of how well
hyperparameters work by iterating over different validation sets
and averaging the performance across these.
 Setting Hyperparameters: k fold Cross-Validation
• For example: 5-fold cross-validation
1. Split the training data into 5 equal folds (parts),
2. Use 4 of them for training, and 1 for validation.
3. Iterate over which fold is the validation fold, and evaluate the performance,
4. Finally average the performance across the different folds.
Example: Cross Validation on k
Example of
5-fold cross-validation
for the value of k.

Each point: single

outcome.

The line goes through

the mean,

Bars indicated
standard deviation
(Seems that k ~= 10
works best for this data)

Fei-Fei Li & Justin Johnson & Serena Yeung

k
K-NN Pros and Cons

33
Feature Scale
• Example:
− height of a person may vary from 1.5m to 1.8m
− weight of a person may vary from 90lb to 300lb
− income of a person may vary from $10K to $1M

What’s the problem here?

• Attributes may have to be scaled to prevent distance

measures from being dominated by one of the attributes.

35
Disadvantage of K-NN
- It is the computationally expensive in testing phase which is
impractical in industry settings. Predict O(N)
• This is bad: we want classifiers that are fast at prediction, slow in
training is ok.
- KNN can suffer from skewed class distributions.
• For example, if a certain class is very frequent in the training set, it will tend
to dominate the majority voting of the new example (large number = more
common).
- Finally, the accuracy of KNN can be severely degraded with high-
dimension data. For that, it is never used on images.
Advantage of K-NN
- Simple to understand and easy to implement.
- With zero to little training time, Train O(1)
- KNN works just as easily with multiclass data sets (more than 2
classes).
Thanks
38