Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria...

28
Anomaly detec-on under mul-ple criteria Kevin S. Xu, 3M February 20, 2013 So?ware, Electronic, and Mechanical Systems Laboratory

Transcript of Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria...

Page 1: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Anomaly  detec-on  under  mul-ple  criteria  

Kevin  S.  Xu,  3M  February  20,  2013  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 2: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Outline  •  Overview  of  3M  Computa-onal  Intelligence  laboratory  

•  Technical  focus:  anomaly  detec-on  under  mul-ple  criteria  – Tradi-onal  anomaly  detec-on  – Challenges  introduced  by  mul-ple  criteria  – Experiment  on  pedestrian  trajectories  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 3: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Computa-onal  Intelligence  laboratory  •  Mission:  build  and  grow  3M  businesses  through  innova-ve  computa-onal  algorithms  that  learn  from  data  

•  Lab  members:  –  Brian  Stankiewicz,  PhD,  UCLA,  Cogni-ve  Science  –  Eric  Lobner,  PhD  candidate,  Minnesota,  Computer  Science  

–  Jennifer  Schumacher,  PhD,  Minnesota,  Neuroscience  –  Ravi  Sivalingam,  PhD,  Minnesota,  Electrical  Engineering  –  Guru  Somasundaram,  PhD,  Minnesota,  Computer  Science  –  Kevin  Xu,  PhD,  Michigan,  Electrical  Engineering:  Systems  –  Anthony  Sabelli,  PhD,  Cornell,  Applied  Math  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 4: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Visual  AYen-on  Model  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Probability  of  geZng  aYen-on  during  first  3  to  5  seconds  

Page 5: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Comparison  with  eye  tracking  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

3M  VAM  predic-on  Eye  tracking  data  

B.  J.  Stankiewicz,  N.  J.  Anderson,  and  R.  J.  Moore  (2011).  Using  performance  efficiency  for  tes-ng  and  op-miza-on  of  visual  aYen-on  models.  Proceedings  of  SPIE  7867,  Image  Quality  and  System  Performance  VIII,  pp.  78670Y.  

hYps://vas.3m.com/  hYp://www.youtube.com/3MVisualAYen-on  

Page 6: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Traffic  sign  management  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 7: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Other  technologies  •  We  work  on  a  variety  of  problems  involving  –  Classifica-on  –  Anomaly  detec-on  –  Time  series  models  –  Computer  vision  –  Inverse  problems  –  Ubiquitous  sensing  –  …  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 8: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

ANOMALY  DETECTION  UNDER  MULTIPLE  CRITERIA  

Joint  work  with  K.-­‐J.  Hsiao,  J.  Calder,  and  A.  O.  Hero  III  (University  of  Michigan)  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 9: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Anomaly  detec-on  

•  Anomaly  detec:on:  automa-cally  detec-ng  significant  devia-ons  from  nominal  behavior  

•  Example:  which  one  of  these  groups  of  pedestrian  trajectories  is  anomalous?  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Anomalous  trajectories   Nominal  trajectories  

Page 10: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Tradi-onal  anomaly  detec-on  •  Many  approaches:  – Nearest  neighbor-­‐based  –  Clustering-­‐based  –  Sta-s-cal  modeling  – …  

•  Many  applica-on  seZngs:  –  Fraud  detec-on  – Medical  informa-cs  – Detec-ng  device  failures  or  malfunc-ons  – …  

•  We  focus  on  unsupervised  anomaly  detec-on  – Unlabeled  training  set  of  mostly  nominal  data  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 11: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Nearest-­‐neighbor  anomaly  detec-on  •  Typically  a  variant  of  the  following  algorithm:  – Training  phase:  •  Obtain  a  training  set  of  mostly  nominal  data  samples  •  For  each  training  sample,  compute  dissimilarity  with  k  nearest  neighboring  samples  

– Test  phase:  •  For  each  test  sample,  compute  dissimilarity  with  k  nearest  training  samples  •  If  dissimilarity  exceeds  some  threshold,  declare  test  sample  to  be  anomalous  

•  Requires  user  to  pick  a  single  dissimilarity  measure  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 12: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Mul--­‐criteria  anomaly  detec-on  •  Complex  data  sets  may  require  mul:ple  dissimilarity  measures  corresponding  to  mul-ple  criteria  

•  Example:  pedestrian  trajectories  

•  2  possible  criteria:  –  Dissimilarity  in  shapes  of  trajectories  –  Dissimilarity  in  walking  speeds  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 13: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

First  aYempt:  convex  combina-ons  •  First  aYempt  at  mul--­‐criteria  anomaly  detec-on:  – Scalariza:on:  take  convex  combina-on  of  dissimilarity  measures  

– How  do  we  choose  weight          in  unsupervised  seZng?  

– Sweep  over  en-re  range                                  and  perform  tradi-onal  (single-­‐criteria)  anomaly  detec-on  for  each  choice  of  weight  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 14: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Scalariza-on  and  Pareto  fronts  •  Alterna-ve  approach:  examine  Pareto  fronts  •  A  mul--­‐criteria  op-miza-on  problem:  – Given          items  and            func-ons                                ,  select  the  item          that  minimizes  

– Typically  cannot  simultaneously  op-mize  all  func-ons  (criteria)  è  no  single  op:mizer  

– An  item            is  Pareto-­‐op:mal  if  no  other  item  is  superior  in  every  criterion  •  No          such  that                                                for  all  criteria  •  Pareto  front:  set  of  all  Pareto-­‐op-mal  items  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 15: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Scalariza-on  and  Pareto  fronts  •  Proper-es  of  Pareto  front:  – Contains  all  op-mizers  found  by  scalariza-on  (taking  convex  combina-ons)  

– Contains  other  items  that  cannot  be  found  by  scalariza-on  

– Scalariza-on  only  iden-fies  items  on  the  convex  por-on  of  the    Pareto  front  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 16: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Proper-es  of  Pareto  fronts  •       :  Pareto  front  (set  of  Pareto-­‐op-mal  points)  •       :  op-mal  points  iden-fied  by  scalariza-on  •  How  large  is                  ?  – Assume  i.i.d.  samples                                    with  density        that  is  zero  outside  of  a  bounded  set    

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 17: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Proper-es  of  Pareto  fronts  •       :  Pareto  front  (set  of  Pareto-­‐op-mal  points)  •       :  op-mal  points  iden-fied  by  scalariza-on  •  How  large  is                  ?  – Assume  i.i.d.  samples                                    with  density        that  is  zero  outside  of  a  bounded  set    

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

If                              is  non-­‐convex  and  sa-sfies  condi-ons  of  Thm.  1  then  for  large        ,  scalariza-on  fails  to  iden:fy  on  the  order  of                              points  

Page 18: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Proper-es  of  Pareto  fronts  •       :  Pareto  front  (set  of  Pareto-­‐op-mal  points)  •       :  op-mal  points  iden-fied  by  scalariza-on  •  How  large  is                  ?  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Even  if          is  convex,  the  Pareto  front  can  s-ll  be  non-­‐convex.  For  large        ,  scalariza-on  fails  to  iden:fy  on  the  order  of                  points  

Page 19: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Pareto  depth  analysis  (PDA)  •  So  far  we  have  looked  at  the  first  Pareto  front  •  We  can  compute  deeper  Pareto  fronts  –  Remove  all  points  from  first  front  –  Find  Pareto  front  on  remaining  points  –  Remove  all  points  from  second  front  –  Find  Pareto  front  on  remaining  points  

– …  •  Pareto  depth  analysis  (PDA)    first  proposed  by  Hero  and  Fleury  (2004)  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

A.  O.  Hero  III  and  G.  Fleury  (2004).  Pareto-­‐op-mal  methods  for  gene  ranking.  The  Journal  of  VLSI  Signal  Processing  38(3):259–275.  

Page 20: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Connec-on  to  anomaly  detec-on  •  Q:  What  do  Pareto  fronts  have  to  do  with  mul--­‐criteria  anomaly  detec-on?  

•  A:  We  can  use  the  Pareto  front  depth  as  “combined”  dissimilarity  measure  –  Transform  dissimilari-es  between  samples  into  dyads  in  K-­‐dimensional  space  

–  Compute  Pareto  fronts  on  dyads  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 21: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Distribu-ons  of  Pareto  front  depths  •  Pareto  front  depths  of  nominal  samples  are  shallower  than  those  of  anomalous  samples  

•  Compute  anomaly  score  for  each  test  sample  – Anomaly  score  =  average  depth  of  dyads  corresponding  to  test  sample  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

40  training  samples  and  2  test  samples  

Dyads  of  nominal  test  sample  o  

Dyads  of  anomalous  test  sample  Δ  

Page 22: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Experiment:  pedestrian  trajectories  •  500  training  trajectories,  200  test  trajectories  •  Trajectories  have  differing  lengths  •  Use  2  criteria:  – Dissimilari-es  in  shapes  of  trajectories  – Dissimilari-es  in  walking  speeds  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 23: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Comparison  of  results  •  We  compare  PDA  to  tradi-onal  

anomaly  detec-on  with  100  uniformly  spaced  weights  

•  PDA  performs  slightly  beYer  than  best  weight  

•  PDA  performs  much  beKer  than  median  weight  

•  Best  weight  is  unknown  in  prac-ce  –  Median  weight  is  a  beYer  representa-on  of  tradi-onal  anomaly  detec-on  performance  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 24: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

More  results  •  Pareto  fronts  of  dyads  are  highly  non-­‐convex  

•  Recall:  scalariza-on  can  only  iden-fy  convex  por-on  of  Pareto  front  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

0 0.01 0.02 0.03 0.04 0.050

0.01

0.02

0.03

0.04

0.05

0.06

Walking speed dissimilarity

Shape d

issi

mila

rity

Anomalous  trajectories   Nominal  trajectories  

Page 25: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Summary  •  3M  Computa-onal  Intelligence  lab  works  on  many  problems  that  involve  learning  from  data  (including  anomaly  detec-on)  

•  Anomaly  detec-on  with  mul-ple  criteria  can  be  performed  using  Pareto  depth  analysis  (PDA)  –  BeYer  performance  than  taking  convex  combina-ons  of  the  mul-ple  criteria  

•  Future  work  –  Can  we  create  a  faster  PDA  algorithm  by  approxima-ng  the  Pareto  fronts?  

–  Can  we  modify  the  PDA  algorithm  so  it  can  be  efficiently  updated  as  more  training  data  is  received?  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

K.-­‐J.  Hsiao,  K.  S.  Xu,  J.  Calder,  and  A.  O.  Hero  III  (2012).  Mul--­‐criteria  anomaly  detec-on  using  Pareto  depth  analysis.  In  Advances  in  Neural  Informa-on  Processing  Systems  25,  pp.  854-­‐862.  

Page 26: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

ADDITIONAL  SLIDES  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Page 27: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Simula-on  experiment  •  300  training  samples,  100  test  samples  •  Nominal  distribu-on:  Uniform  on  the  hyper  cube    

•  Anomalous  distribu-on:  Differs  in  one  anomalous  dimension  (uniform  on                      )  

•  4  criteria:  squared  differences  in  each  dimension  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory  

Nominal  region    

Anomalous  region  

Anomalous  region  

Page 28: Xu anomaly UMN 20130220 final.pptx (Read-Only) · Anomaly(detec-on(under(mul-ple(criteria Kevin(S.(Xu,(3M February(20,(2013(So?ware,(Electronic,(and(Mechanical(Systems(Laboratory

Advantages  and  disadvantages  of  PDA  •  Advantages  – U-lizes  Pareto  fronts,  which  are  superior  to  scalariza-on  for  mul--­‐criteria  op-miza-on  

–  Scales  linearly  in  the  number  of  criteria  •  Sweeping  over  linear  combina-ons  for  scalariza-on  using  a  grid  search  is  exponen-al  in  

•  Disadvantages  –  Compu-ng  all  Pareto  fronts  in  training  phase  requires                          comparisons  and                          floa-ng-­‐point  opera-ons  (worst-­‐case)  

– No  known  efficient  method  to  update  anomaly  detector  as  more  training  data  is  received  

So?ware,  Electronic,  and  Mechanical  Systems  Laboratory