0% found this document useful (0 votes)
3 views

Lecture15

The document discusses anomaly detection in machine learning, providing examples such as aircraft engine monitoring and fraud detection. It outlines the importance of feature selection, model training, and evaluation metrics for developing effective anomaly detection systems. Additionally, it contrasts anomaly detection with supervised learning, highlighting the challenges posed by the scarcity of positive examples in anomaly detection.

Uploaded by

S R Saini
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
3 views

Lecture15

The document discusses anomaly detection in machine learning, providing examples such as aircraft engine monitoring and fraud detection. It outlines the importance of feature selection, model training, and evaluation metrics for developing effective anomaly detection systems. Additionally, it contrasts anomaly detection with supervised learning, highlighting the challenges posed by the scarcity of positive examples in anomaly detection.

Uploaded by

S R Saini
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 37

Anomaly

 
detec-on  
Problem  
mo-va-on  
Machine  Learning  
Anomaly  detec-on  example  
Aircra9  engine  features:   Dataset:  
 =  heat  generated    
 =  vibra-on  intensity   New  engine:  
         …  

(vibra-on)  

(heat)  
Andrew  Ng  
Density  es-ma-on  

Dataset:  
Is                        anomalous?  
(vibra-on)  

(heat)  
Andrew  Ng  
Anomaly  detec-on  example  
Fraud  detec-on:  
                           =  features  of  user      ’s  ac-vi-es  
Model                  from  data.  
Iden-fy  unusual  users  by  checking  which  have    
Manufacturing  
Monitoring  computers  in  a  data  center.  
             =  features  of  machine    
             =  memory  use,                =  number  of  disk  accesses/sec,  
             =  CPU  load,                =  CPU  load/network  traffic.  
…  
 
Andrew  Ng  
Anomaly  
detec-on  
Gaussian  
distribu-on  
Machine  Learning  
Gaussian  (Normal)  distribu-on  
Say                    .  If        is  a  distributed  Gaussian  with  mean        ,  variance            .  

Andrew  Ng  
Gaussian  distribu-on  example  

Andrew  Ng  
Parameter  es-ma-on  
Dataset:  

Andrew  Ng  
Anomaly  
detec-on  
Algorithm  
Machine  Learning  
Density  es-ma-on  
Training  set:  
Each  example  is    

Andrew  Ng  
Anomaly  detec-on  algorithm  
1. Choose  features            that  you  think  might  be  indica-ve  of  
anomalous  examples.  
2. Fit  parameters  
 
 
 
 
3. Given  new  example        ,  compute                  :    

 
             Anomaly  if    
Andrew  Ng  
Anomaly  detec-on  example  

Andrew  Ng  
Anomaly  
detec-on  
Developing  and  
evalua-ng  an  anomaly  
detec-on  system  
Machine  Learning  
The  importance  of  real-­‐number  evalua-on  
When  developing  a  learning  algorithm  (choosing  features,  etc.),  
making  decisions  is  much  easier  if  we  have  a  way  of  evalua-ng  
our  learning  algorithm.  
Assume  we  have  some  labeled  data,  of  anomalous  and  non-­‐
anomalous  examples.    (                        if  normal,                          if  anomalous).  
Training  set:                  (assume  normal  examples/not  
anomalous)  
Cross  valida-on  set:  
Test  set:  

Andrew  Ng  
AircraA  engines  mo-va-ng  example  
10000    good  (normal)  engines  
20    flawed  engines  (anomalous)  

Training  set:  6000  good  engines  


CV:  2000  good  engines  (      ),  10  anomalous  (                      )  
Test:  2000  good  engines  (                      ),  10  anomalous  (                      )  

Alterna-ve:  
Training  set:  6000  good  engines  
CV:  4000  good  engines  (      ),  10  anomalous  (                      )  
Test:  4000  good  engines  (                      ),  10  anomalous  (                      )  
Andrew  Ng  
Algorithm  evalua-on  
Fit  model                    on  training  set  
On  a  cross  valida-on/test  example          ,  predict  

Possible  evalua-on  metrics:  


 -­‐  True  posi-ve,  false  posi-ve,  false  nega-ve,  true  nega-ve  
 -­‐  Precision/Recall  
 -­‐  F1-­‐score  

Can  also  use  cross  valida-on  set  to  choose  parameter    


Andrew  Ng  
Anomaly  
detec-on  
Anomaly  detec-on  
vs.  supervised  
learning  
Machine  Learning  
Anomaly  detec-on   vs.   Supervised  learning  
Very  small  number  of  posi-ve   Large  number  of  posi-ve  and  
examples  (                      ).  (0-­‐20  is   nega-ve  examples.  
common).    
Large  number  of  nega-ve  (                      )    
examples.    
Many  different  “types”  of   Enough  posi-ve  examples  for  
anomalies.  Hard  for  any  algorithm   algorithm  to  get  a  sense  of  what  
to  learn  from  posi-ve  examples   posi-ve  examples  are  like,    future  
what  the  anomalies  look  like;   posi-ve  examples  likely  to  be  
future  anomalies  may  look  nothing   similar  to  ones  in  training  set.  
like  any  of  the  anomalous  
examples  we’ve  seen  so  far.  
Andrew  Ng  
Anomaly  detec-on   vs.   Supervised  learning  
• Fraud  detec-on   • Email  spam  classifica-on  
   
• Manufacturing  (e.g.  aircra9   • Weather  predic-on  (sunny/
engines)   rainy/etc).  
   
• Monitoring  machines  in  a  data   • Cancer  classifica-on  
center  

Andrew  Ng  
Anomaly  
detec-on  
Choosing  what  
features  to  use  
Machine  Learning  
Non-­‐gaussian  features  
Error  analysis  for  anomaly  detec-on  
Want    large  for  normal  examples        .  
   small  for  anomalous  examples        .  
Most  common  problem:  
   is  comparable  (say,  both  large)  for  normal  
 and  anomalous  examples  
Monitoring  computers  in  a  data  center  
Choose  features  that  might  take  on  unusually  large  or  
small  values  in  the  event  of  an  anomaly.  
 =  memory  use  of  computer  
 =  number  of  disk  accesses/sec  
 =  CPU  load  
 =  network  traffic  
 
Anomaly  
detec-on  
Mul-variate  
Gaussian  distribu-on  
Machine  Learning  
Mo-va-ng  example:  Monitoring  machines  in  a  data  center  
(Memory  Use)  

(CPU  Load)  

(CPU  Load)  
(Memory  Use)  
Andrew  Ng  
Mul-variate  Gaussian  (Normal)  distribu-on  
                         .  Don’t  model            etc.  separately.  
Model                      all  in  one  go.  
Parameters:        (covariance  matrix)  

Andrew  Ng  
Mul-variate  Gaussian  (Normal)  examples  

Andrew  Ng  
Mul-variate  Gaussian  (Normal)  examples  

Andrew  Ng  
Mul-variate  Gaussian  (Normal)  examples  

Andrew  Ng  
Mul-variate  Gaussian  (Normal)  examples  

Andrew  Ng  
Mul-variate  Gaussian  (Normal)  examples  

Andrew  Ng  
Mul-variate  Gaussian  (Normal)  examples  

Andrew  Ng  
Anomaly  
detec-on  
Anomaly  detec-on  using  
the  mul-variate  
Gaussian  distribu-on  
Machine  Learning  
Mul-variate  Gaussian  (Normal)  distribu-on  
Parameters    

Parameter  fifng:  
Given  training  set  

Andrew  Ng  
Anomaly  detec-on  with  the  mul-variate  Gaussian  
1.  Fit  model                    by  sefng  

2.  Given  a  new  example        ,  compute  


 
 
 
       Flag  an  anomaly  if    
Andrew  Ng  
Rela-onship  to  original  model  
Original  model:    

Corresponds  to  mul-variate  Gaussian  


 
 
 
where  
Andrew  Ng  
Original  model   vs.   Mul-variate  Gaussian  

Manually  create  features  to   Automa-cally  captures  


capture  anomalies  where                             correla-ons  between  features  
take  unusual  combina-ons  of  
values.  

Computa-onally  cheaper   Computa-onally  more  expensive  


(alterna-vely,  scales  beger  to  
large  
       )  
OK  even  if          (training  set  size)  is   Must  have                          ,  or  else          is  
small   non-­‐inver-ble.  
Andrew  Ng  

You might also like