Machine

 Learning  Crash  Course:  

Part  I  
Ariel  Kleiner  
August  21,  2012  
 Machine  learning  
exists  at  the  intersec<on  of  
computer  science  and  sta<s<cs.  
 Examples  
•  Spam  filters  
•  Search  ranking  
•  Click  (and  clickthrough  rate)  predic<on  
•  Recommenda<ons  (e.g.,  NeJlix,  Facebook  friends)  
•  Speech  recogni<on  
•  Machine  transla<on  
•  Fraud  detec<on  
•  Sen<ment  analysis  
•  Face  detec<on,  image  classifica<on  
•  Many  more  
 A  Variety  of  Capabili<es  
•  Classifica<on   •  Collabora<ve  filtering  
•  Regression   •  Ac<ve  learning  and  
•  Ranking   experimental  design  
•  Clustering   •  Reinforcement  learning  
•  Dimensionality   •  Time  series  analysis  
reduc<on   •  Hypothesis  tes<ng  
•  Feature  selec<on   •  Structured  predic<on  
•  Structured  probabilis<c  
modeling  
 For  Today  

Classifica<on  
Clustering  

(with  emphasis  on  

implementability  and  scalability)  
Typical  Data  Analysis  Workflow  
Obtain  and  load  raw  data  

Data  explora<on  

Preprocessing  and  featuriza<on  

Learning  

Diagnos<cs  and  evalua<on  

 Classifica<on  
•  Goal:  Learn  a  mapping  from  en<<es  to  
discrete  labels.  
–  Refer  to  en<<es  as  x  and  labels  as  y.  

•  Example:  spam  classifica<on  

–  En<<es  are  emails.  
–  Labels  are  {spam,  not-­‐spam}.  
–  Given  past  labeled  emails,  want  to  predict  
whether  a  new  email  is  spam  or  not-­‐spam.  
 Classifica<on  
•  Examples  
–  Spam  filters  
–  Click  (and  clickthrough  rate)  predic<on  
–  Sen<ment  analysis  
–  Fraud  detec<on  
–  Face  detec<on,  image  classifica<on  
 Classifica<on  
Given  a  labeled  dataset  (x1,  y1),  ...,  (xN,  yN):  
1.  Randomly  split  the  full  dataset  into  two  disjoint  
parts:  
–  A  larger  training  set  (e.g.,  75%)  
–  A  smaller  test  set  (e.g.,  25%)  
2.  Preprocess  and  featurize  the  data.  
3.  Use  the  training  set  to  learn  a  classifier.  
4.  Evaluate  the  classifier  on  the  test  set.  
5.  Use  classifier  to  predict  in  the  wild.  
 Classifica<on  

training
classifier
full set
dataset

test set new entity

accuracy prediction
 Example:  Spam  Classifica<on  

From: [email protected]

"Eliminate your debt by spam

giving us your money..."

From: [email protected]

"Hi, it's been a while! not-spam

How are you? ..."
 Featuriza<on  
•  Most  classifiers  require  numeric  descrip<ons  
of  en<<es  as  input.  
•  Featuriza1on:  Transform  each  en<ty  into  a  
vector  of  real  numbers.  
–  StraighJorward  if  data  already  numeric  (e.g.,  
pa<ent  height,  blood  pressure,  etc.)  
–  Otherwise,  some  effort  required.    But,  provides  an  
opportunity  to  incorporate  domain  knowledge.  
 Featuriza<on:  Text  
•  Ofen  use  “bag  of  words”  features  for  text.  
–  En<<es  are  documents  (i.e.,  strings).  
–  Build  vocabulary:  determine  set  of  unique  words  
in  training  set.    Let  V  be  vocabulary  size.  
–  Featuriza1on  of  a  document:  
•  Generate  V-­‐dimensional  feature  vector.  
•  Cell  i  in  feature  vector  has  value  1  if  document  contains  
word  i,  and  0  otherwise.  
 Example:  Spam  Classifica<on  

From: [email protected]

"Eliminate your debt by Vocabulary

giving us your money..."
been
debt
eliminate
giving
how
From: [email protected] it's
money
"Hi, it's been a while! while
How are you? ..."
 Example:  Spam  Classifica<on  
0 been

1 debt

1 eliminate
From: [email protected]
1 giving
"Eliminate your debt by
0 how
giving us your money..."
0 it's

1 money

0 while
 Example:  Spam  Classifica<on  
•  How  might  we  construct  a  classifier?  
•  Using  the  training  data,  build  a  model  that  will  
tell  us  the  likelihood  of  observing  any  given  (x,  y)  
pair.  
–  x  is  an  email’s  feature  vector  
–  y  is  a  label,  one  of  {spam,  not-­‐spam}  
•  Given  such  a  model,  to  predict  label  for  an  email:  
–  Compute  likelihoods  of  (x,  spam)  and  (x,  not-­‐spam).  
–  Predict  label  which  gives  highest  likelihood.  
 Example:  Spam  Classifica<on  
•  What  is  a  reasonable  probabilis<c  model  for  
(x,  y)  pairs?  
•  A  baseline:  
–  Before  we  observe  an  email’s  content,  can  we  say  
anything  about  its  likelihood  of  being  spam?  
–  Yes:  p(spam)  can  be  es<mated  as  the  frac<on  of  
training  emails  which  are  spam.  
–  p(not-­‐spam)  =  1  –  p(spam)  
–  Call  this  the  “class  prior.”    Wrinen  as  p(y).  
 Example:  Spam  Classifica<on  
•  How  do  we  incorporate  an  email’s  content?  
•  Suppose  that  the  email  were  spam.    Then,  
what  would  be  the  probability  of  observing  its  
content?  
 Example:  Spam  Classifica<on  
•  Example:  “Eliminate  your  debt  by  giving  us  your  
money”  with  feature  vector  (0,  1,  1,  1,  0,  0,  1,  0)  
•  Ignoring  word  sequence,  probability  of  email  is  
 p(seeing  “debt”  AND  seeing  “eliminate”  AND  
seeing  “giving”  AND  seeing  “money”  AND  not  
seeing  any  other  vocabulary  words  |  given  that  
email  is  spam)  
•  In  feature  vector  nota<on:  
 p(x1=0,  x2=1,  x3=1,  x4=1,  x5=0,  x6=0,  x7=1,  x8=0  |  
given  that  email  is  spam)  
 Example:  Spam  Classifica<on  
•  Now,  to  simplify,  model  each  word  in  the  
vocabulary  independently:  
–  Assume  that  (given  knowledge  of  the  class  label)  
probability  of  seeing  word  i  (e.g.,  eliminate)  is  
independent  of  probability  of  seeing  word  j  (e.g.,  
money).  
–  As  a  result,  probability  of  email  content  becomes  
 p(x1=0  |  spam)  p(x2=1  |  spam)  ...  p(x8=0  |  spam)  
 rather  than  
 p(x1=0,  x2=1,  x3=1,  x4=1,  x5=0,  x6=0,  x7=1,  x8=0  |  spam)  
 Example:  Spam  Classifica<on  
•  Now,  we  only  need  to  model  the  probability  of  
seeing  (or  not  seeing)  a  par<cular  word  i,  
assuming  that  we  knew  the  email’s  class  y
(spam  or  not-­‐spam).  
–  But,  this  is  easy!  
–  To  es<mate  p(xi  =  1  |  y),  simply  compute  the  
frac<on  of  emails  in  the  set  {emails  in  training  set  
with  label  y}  which  contain  the  word  i.  
 Example:  Spam  Classifica<on  
•  Pusng  it  all  together:  
–  Based  on  the  training  data,  es<mate  the  class  prior            
p(y).  
•  i.e.,  es<mate  p(spam)  and  p(not-­‐spam).  
–  Also  es<mate  the  (condi<onal)  probability  of  seeing  
any  individual  word  i,  given  knowledge  of  the  class  
label  y.  
•  i.e.,  es<mate  p(xi  =  1  |  y)  for  each  i  and  y  
–  The  (condi<onal)  probability  p(x  |  y)  of  seeing  an  
en<re  email,  given  knowledge  of  the  class  label  y,  is  
then  simply  the  product  of  the  condi<onal  word  
probabili<es.  
•  e.g.,  p(x=(0,  1,  1,  1,  0,  0,  1,  0)  |  y)  =  
                                               p(x1=0  |  y)  p(x2=1  |  y)  ...  p(x8=0  |  y)  
 Example:  Spam  Classifica<on  
•  Recall:  we  want  a  model  that  will  tell  us  the  
likelihood  p(x,  y)  of  observing  any  given  (x,  y)  pair.  
•  The  probability  of  observing  (x,  y)  is  the  
probability  of  observing  y,  and  then  observing  x  
given  that  value  of  y:  
p(x,  y)  =  p(y)  p(x  |  y)  
•  Example:  
 p(“Eliminate  your  debt...”,  spam)  =  
       p(spam)  p(“Eliminate  your  debt...”  |  spam)  
 Example:  Spam  Classifica<on  
•  To  predict  label  for  a  new  email:  
–  Compute  log[p(x,  spam)]  and  log[p(x,  not-­‐spam)].  
–  Choose  the  label  which  gives  higher  value.  
–  We  use  logs  above  to  avoid  underflow  which  
otherwise  arises  in  compu<ng  the  p(x  |  y),  which  
are  products  of  individual  p(xi  |  y)  <  1:  
       log[p(x,  y)]  =  log[p(y)  p(x  |  y)]  
                         =  log[  p(y)  p(x1  |  y)  p(x2  |  y)  ...]  
                         =  log[p(y)]  +  log[p(x1  |  y)]  +  log[p(x2  |  y)]  +  ...  
 Classifica<on:  Beyond  Text  
•  You  have  just  seen  an  instance  of  the  Naive  
Bayes  classifier.  
•  Applies  as  shown  to  any  classifica<on  problem  
with  binary  feature  vectors.  
•  What  if  the  features  are  real-­‐valued?  
–  S<ll  model  each  element  of  the  feature  vector  
independently.  
–  But,  change  the  form  of  the  model  for  p(xi  |  y).  
 Classifica<on:  Beyond  Text  
•  If  xi  is  a  real  number,  ofen  model  p(xi  |  y)  as  
2
Gaussian  with  mean  µ        iy
       and  variance    σ
     iy
     :  
� x −µ �2
1 − 12 i iy
p(xi |y) = √ e σiy

σiy 2π
•  Es<mate  the  mean  and  variance  for  a  given  i,y  as  
the  mean  and  variance  of  the  xi  in  the  training  set  
which  have  corresponding  class  label  y.  
•  Other,  non-­‐Gaussian  distribu<ons  can  be  used  if  
know  more  about  the  xi.  
 Naive  Bayes:  Benefits  
•  Can  easily  handle  more  than  two  classes  and  
different  data  types  
•  Simple  and  easy  to  implement  
•  Scalable  
 Naive  Bayes:  Shortcomings  
•  Generally  not  as  accurate  as  more  sophis<cated  
methods  (but  s<ll  generally  reasonable).  
•  Independence  assump<on  on  the  feature  vector  
elements  
–  Can  instead  directly  model  p(x  |  y)  without  this  
independence  assump<on.  
•  Requires  us  to  specify  a  full  model  for  p(x,  y)  
–  In  fact,  this  is  not  necessary!  
–  To  do  classifica<on,  we  actually  only  require  p(y  |  x),  
the  probability  that  the  label  is  y,  given  that  we  have  
observed  en<ty  features  x.  
 Logis<c  Regression  
•  Recall:  Naive  Bayes  models  the  full  ( joint)  
probability  p(x,  y).  
•  But,  Naive  Bayes  actually  only  uses  the  
condi<onal  probability  p(y  |  x)  to  predict.  
•  Instead,  why  not  just  directly  model  p(y  |  x)?  
–  Logis1c  regression  does  exactly  that.  
–  No  need  to  first  model  p(y)  and  then  separately          
p(x  |  y).  
 Logis<c  Regression  
•  Assume  that  class  labels  are  {0,  1}.  
•  Given  an  en<ty’s  feature  vector  x,  probability  that  
label  is  1  is  taken  to  be  
1
p(y = 1|x) = −b Tx
1+e
 where  b  is  a  parameter  vector  and  bTx  denotes  a  
dot  product.  
•  The  probability  that  the  label  is  1,  given  features  
x,  is  determined  by  a  weighted  sum  of  the  
features.  
 Logis<c  Regression  
•  This  is  libera<ng:  
–  Simply  featurize  the  data  and  go.  
–  No  need  to  find  a  distribu<on  for  p(xi  |  y)  which  is  
par<cularly  well  suited  to  your  sesng.  
–  Can  just  as  easily  use  binary-­‐valued  (e.g.,  bag  of  
words)  or  real-­‐valued  features  without  any  
changes  to  the  classifica<on  method.  
–  Can  ofen  improve  performance  simply  by  adding  
new  features  (which  might  be  derived  from  old  
features).  
 Logis<c  Regression  
•  Can  be  trained  efficiently  at  large  scale,  but  
not  as  easy  to  implement  as  Naive  Bayes.  
–  Trained  via  maximum  likelihood.  
–  Requires  use  of  itera<ve  numerical  op<miza<on  
(e.g.,  gradient  descent,  most  basically).  
–  However,  implemen<ng  this  effec<vely,  robustly,  
and  at  large  scale  is  non-­‐trivial  and  would  require  
more  <me  than  we  have  today.  
•  Can  be  generalized  to  mul<class  sesng.  
 Other  Classifica<on  Techniques  
•  Support  Vector  Machines  (SVMs)  
•  Kernelized  logis<c  regression  and  SVMs  
•  Boosted  decision  trees  
•  Random  Forests  
•  Nearest  neighbors  
•  Neural  networks  
•  Ensembles  

See  The  Elements  of  Sta4s4cal  Learning  by  Has<e,  

Tibshirani,  and  Friedman  for  more  informa<on.  
 Featuriza<on:  Final  Comments  
•  Featuriza<on  affords  the  opportunity  to  
–  Incorporate  domain  knowledge  
–  Overcome  some  classifier  limita<ons  
–  Improve  performance  
•  Incorpora<ng  domain  knowledge:  
–  Example:  in  spam  classifica<on,  we  might  suspect  that  
sender  is  important,  in  addi<on  to  email  body.  
–  So,  try  adding  features  based  on  sender’s  email  
address.  
 Featuriza<on:  Final  Comments  
•  Overcoming  classifier  limita<ons:  
–  Naive  Bayes  and  logis<c  regression  do  not  model  
mul<plica<ve  interac<ons  between  features.  
–  For  example,  the  presence  of  the  pair  of  words  
[eliminate,  debt]  might  indicate  spam,  while  the  
presence  of  either  one  individually  might  not.  
–  Can  overcome  this  by  adding  features  which  explicitly  
encode  such  interac<ons.  
–  For  example,  can  add  features  which  are  products  of  
all  pairs  of  bag  of  words  features.  
–  Can  also  include  nonlinear  effects  in  this  manner.  
–  This  is  actually  what  kernel  methods  do.  
 Classifica<on  
Given  a  labeled  dataset  (x1,  y1),  ...,  (xN,  yN):  
1.  Randomly  split  the  full  dataset  into  two  disjoint  
parts:  
–  A  larger  training  set  (e.g.,  75%)  
–  A  smaller  test  set  (e.g.,  25%)  
2.  Preprocess  and  featurize  the  data.  
3.  Use  the  training  set  to  learn  a  classifier.  
4.  Evaluate  the  classifier  on  the  test  set.  
5.  Use  classifier  to  predict  in  the  wild.  
 Classifica<on  

training
classifier
full set
dataset

test set new entity

accuracy prediction
 Classifier  Evalua<on  
•  How  do  we  determine  the  quality  of  a  trained  
classifier?  
•  Various  metrics  for  quality,  most  common  is  
accuracy.  
•  How  do  we  determine  the  probability  that  a  
trained  classifier  will  correctly  classify  a  new  
en<ty?  
 Classifier  Evalua<on  
•  Cannot  simply  evaluate  a  classifier  on  the  
same  dataset  used  to  train  it.  
–  This  will  be  overly  op<mis<c!  
•  This  is  why  we  set  aside  a  disjoint  test  set  
before  training.  
 Classifier  Evalua<on  
•  To  evaluate  accuracy:  
–  Train  on  the  training  set  without  exposing  the  test  set  
to  the  classifier.  
–  Ignoring  the  (known)  labels  of  the  data  points  in  the  
test  set,  use  the  trained  classifier  to  generate  label  
predic<ons  for  the  test  points.  
–  Compute  the  frac<on  of  predicted  labels  which  are  
iden<cal  to  the  test  set’s  known  labels.  
•  Other,  more  sophis<cated  evalua<on  methods  
are  available  which  make  more  efficient  use  of  
data  (e.g.,  cross-­‐valida<on).