data-intensive distributed computing - university of waterloo

Data-Intensive Distributed Computing

Part 6: Data Mining (1/4)

This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 United StatesSee http://creativecommons.org/licenses/by-nc-sa/3.0/us/ for details

CS 431/631 451/651 (Winter 2022)

Ali Abedi

These slides are available at https://www.student.cs.uwaterloo.ca/~cs451

1

1

Structure of the Course

“Core” framework features

and algorithm design

Analy

zin

gT

ext

Analy

zin

gG

raphs

Analy

zin

g

Rela

tional D

ata

Data

Min

ing

2

2

Supervised Machine Learning

The generic problem of function induction given sample instances of input and output

Classification: output draws from finite discrete labels

Regression: output is a continuous value

This is not meant to be an exhaustive treatment of machine learning!

3

3

Source: Wikipedia (Sorting)

Classification

4

4

Applications

Spam detection

Sentiment analysis

Content (e.g., topic) classification

Link prediction

Document ranking

Object recognition

And much much more!

Fraud detection

5

5

Applications

Spam detection

Sentiment analysis

Content (e.g., topic) classification

Link prediction

Document ranking

Object recognition

And much much more!

Fraud detection

6

6

training

Model

Machine Learning Algorithm

testing/deployment

?

Supervised Machine Learning

7

7

Objects are represented in terms of features:

“Dense” features: sender IP, timestamp, # of recipients, length of message, etc.

“Sparse” features: contains the term “Viagra” in message, contains “URGENT” in subject, etc.

Feature Representations

8

8

Applications

Spam detection

Sentiment analysis

Content (e.g., genre) classification

Link prediction

Document ranking

Object recognition

And much much more!

Fraud detection

9

9

Components of a ML Solution

DataFeaturesModel

Optimization

10

Data matters the most. If we throw a lot of data at almost any algorithm it performs

good.

10

(Banko and Brill, ACL 2001)

(Brants et al., EMNLP 2007)

No data like more data!

11

We have seem these examples before. For example stupid backoff outperforms

other algorithms when it’s trained on a lot of data.

11

Limits of Supervised Classification?

Why is this a big data problem?Isn’t gathering labels a serious bottleneck?

SolutionsCrowdsourcing

Bootstrapping, semi-supervised techniquesExploiting user behavior logs

12

12

Supervised Binary Classification

Restrict output label to be binaryYes/No

1/0

Binary classifiers form primitive building blocks for multi-class problems…

13

13

Binary Classifiers as Building BlocksExample: four-way classification

One vs. rest classifiers

A or not?

B or not?

C or not?

D or not?

A or not?

B or not?

C or not?

D or not?

Classifier cascades

14

14

The Task

Given:

(sparse) feature vector

label

Induce:Such that loss is minimized

loss function

Typically, we consider functions of a parametric form:

model parameters15

15

Key insight: machine learning as an optimization problem!

16

16

Because the closed form solutions generally not possible …

Gradient Descent

17

Old weights

Update based on gradient

New weights

Intuition behind the math…

18

Gradient Descent: Preliminaries

18

Linear Regression + Gradient Descent

19

Gradient Descent: Preliminaries

Rewrite:

Compute gradient:“Points” to fastest increasing “direction”

So, at any point:

20

20

Gradient Descent: Iterative Update

Start at an arbitrary point, iteratively update:

We have:

21

21

Gradient Descent: Iterative Update

Start at an arbitrary point, iteratively update:

Lots of details:Figuring out the step size

Getting stuck in local minimaConvergence rate

…

We have:

22

22

Repeat until convergence:

Gradient Descent

Note, sometimes formulated as ascent but entirely equivalent23

23

Even More Details…

Gradient descent is a “first order” optimization techniqueOften, slow convergence

Newton and quasi-Newton methods:Intuition: Taylor expansion

Requires the Hessian (square matrix of second order partial derivatives):impractical to fully compute

24

24

Logistic Regression: Preliminaries

26

Logistic Regression: Preliminaries

Given:

Define:

Interpretation:

27

27

Relation to the Logistic Function

After some algebra:

The logistic function:

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8

log

istic(z

)

z28

28

Training an LR Classifier

Maximize the conditional likelihood:

Define the objective in terms of conditional log likelihood:

We know:

So:

Substituting:

29

29

LR Classifier Update RuleTake the derivative:

General form of update rule:

Final update rule:

30

30

Want more details? Take a real machine-learning course!

Lots more details…

Regularization

Different loss functions…

31

31

32

mapper mapper mapper mapper

reducer

compute partial gradient

single reducer

mappers

update model iterate until convergence

MapReduce Implementation

33

33

Shortcomings

Hadoop is bad at iterative algorithmsHigh job startup costs

Awkward to retain state across iterations

High sensitivity to skewIteration speed bounded by slowest task

Potentially poor cluster utilizationMust shuffle all data to a single reducer

Some possible tradeoffsNumber of iterations vs. complexity of computation per iteration

E.g., L-BFGS: faster convergence, but more to compute

34

34

val points = spark.textFile(...).map(parsePoint).persist()

var w = // random initial vectorfor (i <- 1 to ITERATIONS) {

val gradient = points.map{ p =>p.x * (1/(1+exp(-p.y*(w dot p.x)))-1)*p.y

}.reduce((a,b) => a+b)w -= gradient

}

mapper mapper mapper mapper

reducer

compute partial gradient

update model

Spark Implementation

35

35