"Miracles"

of Machine Learning


Dmitry Efimov


American University of Sharjah

March, 5, 2015

Self-driving car

Crawling robot

Spam detection

spamdetection

People you may know

facebook

Dubizzle example

This is our "training data":

Age (years)
Mileage (km)
Brand
Price (AED)
2 40 000 Toyota Fortuner
75 000
3 35 000 Nissan Patrol
90 000
3 75 000 Toyota Fortuner
65 000
4 45 000 Nissan Patrol
40 000
1 10 000 Toyota Fortuner
95 000
2 60 000 Nissan Patrol
85 000


We want to "predict" a car price:

Age (years)
Mileage (km)
Brand
Price (AED)
4 100 000 Toyota Fortuner
???

Analogy:

Math Quiz - Teacher's Answer Key

1)   2     4     5   =   3 5)   6     2     2   =   10
2)   5     2     8   =   2 6)   3     1     1   =   2
3)   2     2     1   =   3 7)   5     3     4   =   11
4)   4     2     2   =   6 8)   1     8     1   =   7


Somebody erased the arithmetic symbols!

Let's recover it!

Dubizzle example solution

Age
Mileage
Brand
Price
My Guess
2 40 000 Toyota Fortuner
75 000
78 221
3 35 000 Nissan Patrol
90 000
71 639
3 75 000 Toyota Fortuner
65 000
59 208
4 45 000 Nissan Patrol
40 000
52 876
1 10 000 Toyota Fortuner
95 000
97 185
2 60 000 Nissan Patrol
85 000
90 053



\[ \begin{align*} \mbox{MyGuess} = &-18663.59 \times \mbox{Age} - 0.01 \times \mbox{Mileage}\\ &- 12031.49 \times \mbox{Brand} + 127980.03 \end{align*} \]

A dumb way to solve Dubizzle problem

Age
Mileage
Brand
Price
My Guess
2 40 000 Toyota Fortuner
75 000
40 003
3 35 000 Nissan Patrol
90 000
35 003
3 75 000 Toyota Fortuner
65 000
75 004
4 45 000 Nissan Patrol
40 000
45 004
1 10 000 Toyota Fortuner
95 000
10 002
2 60 000 Nissan Patrol
85 000
60 002


\[ \begin{align*} \mbox{MyGuess} = & 1.0 \times \mbox{Age} + 1.0 \times \mbox{Mileage}\\ & + \; 1.0 \times \mbox{Brand} \end{align*} \]
Try to guess weights... :(

A smart way to solve Dubizzle problem

Age
Mileage
Brand
Price
My Guess
2 40 000 Toyota Fortuner
75 000
40 003
3 35 000 Nissan Patrol
90 000
35 003
3 75 000 Toyota Fortuner
65 000
75 004
4 45 000 Nissan Patrol
40 000
45 004
1 10 000 Toyota Fortuner
95 000
10 002
2 60 000 Nissan Patrol
85 000
60 002


\[ \mbox{Cost} = \dfrac{1}{6}\sum\limits_{i=1}^{6} (\mbox{MyGuess}(i) - \mbox{Price}(i))^2 \]

\[ \mbox{MyGuess} = \mathbf{a} \times \mbox{Age} + \mathbf{b} \times \mbox{Mileage} \]

Graph the cost equation

\[ \Tiny \mbox{Cost} = \dfrac{1}{6}\sum\limits_{i=1}^{6} (\mathbf{a} \times \mbox{Age}(i) + \mathbf{b} \times \mbox{Mileage}(i) - \mbox{Price}(i))^2 \]

plot of chunk example-fig1

Machine Learning algorithms


  • Multivariate linear regression

  • Neural Networks

  • Decision Trees

  • Support Vector Machines

  • Factorization Machines

  • Boosting and bagging

  • Clustering methods

  • Bayesian learning

  • ...

What about our "miracles"?


  • Self-driving car

  • Crawling robot

  • Spam detection

  • People you may know



ALL OF THEM ARE MACHINE LEARNING PROBLEMS!

Thank you!

Dmitry Efimov
defimov@aus.edu

Questions?

questions