2. Machine Learning: Intro
What is Machine Learning?
[Wikipedia]: a branch of artificial intelligence
that allows the construction and the study of
systems that can learn from data
3. Machine Learning: Intro
Some approaches:
- Regression analysis
- Similarity and metric learning
- Decision tree learning
- Association rule learning
- Artificial neural networks
- Genetic programming
- Support vector machines
(classification and regression analysis)
- Clustering
- Bayesian networks
5. Machine Learning: Regression analysis
Regression Analysis
A statistical technique for estimating
the relationships among a dependent
variable and independent variables
8. Machine Learning: Regression analysis
Prediction of house prices
Hypothesis:
h θ (x )=θ0 + θ1 x
Cost function for linear regression:
m
1
J (θ 0, θ1 )=
(h θ (x (i) )− y(i ) )2
∑
2m i=1
9. Machine Learning: Regression analysis
Prediction of house prices
Hypothesis:
h θ (x )=θ0 + θ1 x
Cost function for linear regression:
m
1
J (θ 0, θ1 )=
(h θ (x (i) )− y(i ) )2
∑
2m i=1
Gradient Descent
repeat until convergence :
m
1
(i )
(i )
θ 0=θ 0−α ∑ (hθ ( x )− y )
m i =1
m
1
θ1 =θ1 −α ∑ [(h θ (x (i) )− y(i )) x (i) ]
m i =1
10. Machine Learning: Regression analysis
Prediction of house prices
Iterative minimization of cost function
with gradient descent
15. Machine Learning: Similarity and metric learning
Euclidean distance
euclidean distance (p , q )=
√
n
∑ (p i −q i )2
i =1
16. Machine Learning: Similarity and metric learning
Manhattan distance
n
manhattan distance (p , q )=∑ ∣(p i −q i )∣
i =1
17. Machine Learning: Similarity and metric learning
Pearson's correlation
n
n
∑ pi ∑ qi
n
∑ (p i q i )− i =1
Pearson ' s correlation ( p , q )=
i =1
√
n
n
2
i
(∑ p −
i =1
i =1
n
n
2
(∑ p i )
i =1
n
2
n
(∑ qi )
i =1
n
)( ∑ q 2 −
i
i =1
)
18. Machine Learning: Similarity and metric learning
Collaborative filtering
Searches a large group of users for finding a
small subset that have tastes like yours.
Based on what this subset likes or dislikes
the system can recommend you other items.
Two main approaches:
- User based filtering
- Item based filtering
19. Machine Learning: Similarity and metric learning
User based filtering
- based on ratings given to
the items, we can measure
the distance among users
- we can recommend to the
user the items that have
the highest ratings among
the closest users
21. Machine Learning: Similarity and metric learning
Is user based filtering good for
- scalability?
- sparse data?
- quickly changing data?
22. Machine Learning: Similarity and metric learning
Is user based filtering good for
- scalability?
- sparse data?
- quickly changing data?
No, it's better to use item
based filtering
23. Machine Learning: Similarity and metric learning
Euclidean distance for item based filtering:
nothing has changed!
- based on ratings got from
the users, we can measure
the distance among items
- we can recommend an
item to a user, getting the
items that are closer to
the highest rated by the
user
25. Machine Learning: Bayes' classifier
Bayes' theorem
P ( A∣B )=
P (B∣A)P (A )
P (B )
Example: given a company where 70% of developers use Java and 30%
use C++, and knowing that half of the Java developers always use
enhanced for loop, if you look at the snippet:
for (int j=0; j<100; j++) {
t = tests[j];
}
which is the probability that the developer who wrote it uses Java?
26. Machine Learning: Bayes' classifier
Bayes' theorem
P ( A∣B )=
P (B∣A)P (A )
P (B )
Example: given a company where 70% of developers use Java and 30%
use C++, and knowing that half of the Java developers always use
enhanced for loop, if you look at the snippet:
for (int j=0; j<100; j++) {
t = tests[j];
}
which is the probability that the developer who wrote it uses Java?
Hint:
A = developer uses Java
B = developer writes old for loops
27. Machine Learning: Bayes' classifier
Bayes' theorem
P ( A∣B )=
P (B∣A)P (A )
P (B )
Example: given a company where 70% of developers use Java and 30%
use C++, and knowing that half of the Java developers always use
enhanced for loop, if you look at the snippet:
for (int j=0; j<100; j++) {
t = tests[j];
}
which is the probability that the developer who wrote it uses Java?
Solution:
A = developer uses Java
B = developer writes old for loops
P(A) = prob. that a developer uses Java = 0.7
P(B) = prob. that any developer uses old for loop = 0.3 + 0.7*0.5 = 0.65
P(B|A) = prob. that a Java developer uses old for loop = 0.5
P (B∣A)P (A) 0.5⋅0.7
P (A∣B )=
=
=0.54
P (B )
0.65
28. Machine Learning: Bayes' classifier
Naive Bayes' classifier
- supervised learning
- trained on a set of known classes
- computes probabilities of elements to be in a class
- smoothing required
n
∏ P (c∣w i )
P c (w 1 , .... , w n )=
i =1
n
n
i =1
i =1
∏ P (c∣w i )+ ∏ (1−P (c∣w i ))
29. Machine Learning: Bayes' classifier
Naive Bayes' classifier
Example
- we want a classifier for Twitter messages
- define a set of classes: {art, tech, home, events,.. }
- trains the classifier with a set of alreay classified tweets
- when a new tweet arrives, the classifier will (hopefully)
tell us which class it belongs to
31. Machine Learning: Bayes' classifier
Sentiment analysis
- define two classes: { +, - }
- define a set of words: { like, enjoy, hate, bore, fun, …}
- train a NBC with a set of known +/- comments
- let NBC classify any new comment to know if +/- performance is related to quality of training set
33. Machine Learning: Clustering
K-Means clustering
K-Means aims at identifying
cluster centroids, such that an
item belonging to a cluster X,
is closer to the centroid of
cluster X than to the centroid
of any other cluster.
34. Machine Learning: Clustering
K-Means clustering
The algorithm requires a
number of clusters to start, in
this case 3. The centroids are
placed in the item space,
typically in random locations.
35. Machine Learning: Clustering
K-Means clustering
The algorithm will then assign
to each centroid all items that
are closer to it than to any
other centroid.
41. Machine Learning: Clustering
K-Means clustering
Another iteration occurs,
taking into account the new
centroid positions. Note that
this
time
the
cluster
membership did not change.
The cluster centers will not
move anymore.
50. Machine Learning: Neural networks
Neural networks
Backpropagation
Phase 1: Propagation
- Forward propagation of a training pattern's input
through the neural network in order to generate
the propagation's output activations
- Backward propagation of the propagation's output
activations through the neural network using the
training pattern target in order to generate the
deltas of all output and hidden neurons
Phase 2: Weight update
- Multiply its output delta and input activation to
get the gradient of the weight
- Bring the weight in the opposite direction of the
gradient by subtracting a ratio of it from the weight
53. Machine Learning: Genetic algorithms
Genetic algorithms
GA is a programming technique that mimics
biological evolution as a problem-solving strategy
Steps
- maps the variables of the problem into a sequence of
bits, a chromosome
Chromosome
- creates a random population of chromosomes
- let evolve the population using evolution laws:
- the higher the fitness, the higher the chance of breeding
- crossover of chromosomes
- mutation in chromosomes
- if otpimal solution is found or after n steps the
process is stopped