Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
Knowledge-­‐empowered	
  Probabilis3c	
  Graphical	
  
Models	
  for	
  Physical-­‐Cyber-­‐Social	
  Systems	
  
Pramod	
 ...
2	
  
Multimodal Manifestation of Real-World Events: Power Grid Scenario
Image	
  Credit:	
  Twi%er,	
  hUp://bit.ly/1SsE9...
3	
  
Multimodal Manifestations of Real-World Events: Asthma Scenario
Image	
  Credit:	
  hUp://www.rtmagazine.com/2015/10...
Multimodal Manifestation of Real-World Events: Traffic Scenario
4	
  
Traffic	
  related	
  events	
  manifest	
  in	
  phys...
Processing Multimodal Manifestations of Real-World Events
5	
  
“Informa3on	
  is	
  a	
  source	
  of	
  learning.	
  But...
Processing Multimodal Manifestations in PCS Systems
6	
  
The	
  Tao	
  of	
  Boyd:	
  How	
  to	
  Master	
  the	
  OODA	...
7	
  
Thesis Statement
Observa3ons	
   from	
   diverse	
   modali3es	
   can	
   provide	
   complementary,	
   corrobora...
8	
  
“Graphical	
  models	
  are	
  a	
  marriage	
  between	
  probability	
  theory	
  and	
  graph	
  theory.	
  They	...
9	
  
Example of Declarative Domain Knowledge
road	
  ice	
  
Causes	
  
accident	
  
Linked	
  Open	
  Data	
  
(Declara+...
Processing Multimodal Manifestations in PCS Systems
10	
  
PCS	
  Event	
  	
  
Extrac3on	
  
PCS	
  Event	
  	
  
Underst...
Processing Multimodal Manifestations in PCS Systems
11	
  
PCS	
  Event	
  	
  
Extrac3on	
  
PCS	
  Event	
  	
  
Underst...
•  Why?	
  
–  Explain/Interpret	
   average	
   speed	
   and	
   link	
   travel	
   +me	
  
varia+ons	
   using	
   eve...
•  How?	
  
o  Step	
  1:	
  Extract	
  textual	
  events	
  from	
  tweets	
  stream	
  
o  Step	
  2:	
  Build	
  sta+s+...
•  How?	
  
o  Step	
  1:	
  Extract	
  textual	
  events	
  from	
  tweets	
  stream	
  
o  Step	
  2:	
  Build	
  sta+s+...
Processing Multimodal Manifestations in PCS Systems
15	
  
PCS	
  Event	
  	
  
Extrac3on	
  
PCS	
  Event	
  	
  
Underst...
16	
  
People Reporting Various Events in a City on Twitter
Public	
  Safety	
  
Urban	
  planning	
  
Gov.	
  &	
  agency...
17	
  
Extracting City Events from Twitter: Proposed Solution
[ACM-­‐TIST-­‐15]	
  Extrac+ng	
  City	
  Traffic	
  Events	
 ...
18	
  
Label	
  image	
  sequence	
  of	
  Jus+n	
  Bieber’s	
  day	
  J	
  	
  
Sleeping	
   Driving	
   Exercising
Driv...
19	
  
The	
  global	
  normaliza+on	
  and	
  the	
  discrimina+ve	
  nature	
  of	
  the	
  model	
  dis+nguishes	
  
CR...
20	
  
0.6	
  miles	
  
Max-­‐lat	
  
Min-­‐lat	
  
Min-­‐long	
  
Max-­‐long	
  
0.38	
  miles	
  
37.7545166015625, -122...
21	
  
•  City	
  Event	
  Annota+on	
  
–  Automated	
  crea+on	
  of	
  training	
  data	
  	
  
–  Annota+on	
  task	
 ...
Evaluation: City Event Annotation
22	
  
Baseline	
  Annota+on	
  Model	
  [RiUer	
  et	
  al.	
  2012]	
   Our	
  Annota+...
Complementary	
  Events	
  
Textual Events from Tweets vs. 511.org: Complementary
23	
  
traffic	
   incident;	
  road-­‐con...
Textual Events from Tweets vs. 511.org: Corroborative
Corrobora+ve	
  Events	
  
24	
  
fog	
   visibility-­‐air-­‐quality...
Timeliness	
  
Textual Events from Tweets vs. 511.org: Timeliness
25	
  
concert	
   concert	
  
Extracting Textual Events from Tweets for Data from May-14 to May-15
1Event	
  Extrac+on	
  Tool	
  on	
  Open	
  Science	...
•  How?	
  
o  Step	
  1:	
  Extract	
  textual	
  events	
  from	
  tweets	
  stream	
  
o  Step	
  2:	
  Build	
  sta+s+...
Image	
  credit:	
  hUp://traffic.511.org/index	
  	
  
Mul+ple	
  events	
  	
  
Varying	
  influence	
  	
  
Event	
  inter...
•  Temporal	
  landmarks	
  :	
  peak	
  hour	
  vs.	
  off-­‐peak	
  traffic	
  
vs.	
  weekend	
  traffic	
  
•  Effect	
  of	...
Modeling City Traffic Dynamics: A Closer Look
Image	
  credits:	
  hUp://bit.ly/1N1wu5g,	
  hUp://bit.ly/1O8d9gn,	
  hUp:/...
Modeling City Traffic Dynamics: Nature of the Problem
Hidden	
  States	
   Observed	
  Evidence	
  1.	
  There	
  are	
  b...
Modeling the Problem as Linear Dynamical System (LDS)
1.	
  There	
  are	
  both	
  
hidden	
  states	
  and	
  
observed	...
Probabilistic Reasoning Over Time: Discrete Variables
Russell,	
  Stuart,	
  and	
  Peter	
  Norvig.	
  "Ar+ficial	
  intel...
Probabilistic Reasoning Over Time: Continuous Variables
v1	
  
s1	
  
…	
  
…	
  
v2	
  
s1	
  
vT	
  
sT	
  
Linear	
  Dy...
Hourly Link Speed Dynamics Over all Mondays between Aug-14 to Jan-15
x-­‐axis:	
  observa3on	
  number	
  for	
  each	
  h...
36	
  
Switching Linear Dynamical Systems
v1	
  
s1	
  
…	
  
…	
  
v2	
  
s1	
  
vT	
  
sT	
  
h1	
   h2	
   hT	
  …	
   ...
Modeling City Traffic Dynamics: Choosing a Suitable Model
"All	
  models	
  are	
  wrong,	
  but	
  some	
  are	
  useful....
38	
  
Learning Context Specific LDS Models
7	
  ×	
  24	
  
LDS(1,1),	
  LDS(1,2)	
  	
  	
  ,….,	
  LDS(1,24)	
  
LDS(7,...
Learning Normalcy for Each Link, Day of Week, and Hour of Day
Log-­‐likelihood	
  	
  
	
  score	
  
39	
  
Five-­‐number	...
40	
  
Tagging Anomalies using Context Specific LDS Models
Compute	
  Log	
  Likelihood	
  for	
  	
  
each	
  hour	
  of	...
•  How?	
  
o  Step	
  1:	
  Extract	
  textual	
  events	
  from	
  tweets	
  stream	
  
o  Step	
  2:	
  Build	
  sta+s+...
 
•  Anomaly	
  in	
  link	
  data	
  during	
  +me	
  period	
  [ast,aet],	
  is	
  
explained	
  by	
  an	
  event	
  if...
•  Data	
  collected	
  from	
  San	
  Francisco	
  Bay	
  Area	
  between	
  
May	
  2014	
  to	
  May	
  2015	
  
–  511...
Multimodal Data Integration: Evaluation
44	
  
•  Examined	
  the	
  theore3cal	
  nature	
  of	
  the	
  problem	
  of	
  
modeling	
  traffic	
  dynamics	
  to	
  system...
Processing Multimodal Manifestations in PCS Systems
46	
  
PCS	
  Event	
  	
  
Extrac3on	
  
PCS	
  Event	
  	
  
Underst...
•  Contributed	
  to	
  a	
  language	
  to	
  represent	
  tasks	
  
–  Using	
  Seman+c	
  Web	
  based	
  representa+on...
Revisiting the Thesis Statement
48	
  
PCS	
  Event	
  	
  
Extrac3on	
  
PCS	
  Event	
  	
  
Understanding	
  
PCS	
  Ac...
49	
  
Conclusion
•  Observa+ons	
  from	
  people	
  can	
  provide	
  complementary,	
  
corrobora3ve,	
  and	
  3mely	
...
50	
  
Probabilistic Graphical Models, Declarative Knowledge, and PCS Systems
Declara+ve	
  
Knowledge	
  
Data	
  
Textua...
51	
  
Personalized Digital Health for Asthma Management in Children
Sensordrone
(Carbon monoxide,
temperature, humidity)
...
52	
  
PhD @ Kno.e.sis
Awards	
  and	
  Recogni3on	
  
2016	
  Outstanding	
  Graduate	
  Student	
  Award	
  in	
  the	
 ...
[AAAI-­‐16]	
  Pramod	
  Anantharam,	
  Krishnaprasad	
  Thirunarayan,	
  Surendra	
  Marupudi,	
  Amit	
  Sheth,	
  Tanvi...
54	
  
Dr.	
  Payam	
  Barnaghi	
   Dr.	
  Biplav	
  Srivastava	
  
Dr.	
  Cory	
  Henson	
   Dr.	
  Shalini	
  Forbis,	
 ...
Thank you J
55	
  
kHealth	
  Team	
  
Dr.	
  Tanvi	
  Banerjee	
  
Surendra	
  Marupudi	
  
Vaikunth	
  Sridharan	
  	
 ...
Upcoming SlideShare
Loading in …5
×

Knowledge-empowered Probabilistic Graphical Models for Physical-Cyber-Social Systems

2,420 views

Published on

There is a rapid intertwining of sensors and mobile devices into the fabric of our lives. This has resulted in unprecedented growth in the number of observations from the physical and social worlds reported in the cyber world. Sensing and computational components embedded in the physical world is termed as Cyber-Physical System (CPS). Current science of CPS is yet to effectively integrate citizen observations in CPS analysis. We demonstrate the role of citizen observations in CPS and propose a novel approach to perform a holistic analysis of machine and citizen sensor observations. Specifically, we demonstrate the complementary, corroborative, and timely aspects of citizen sensor observations compared to machine sensor observations in Physical-Cyber-Social (PCS) Systems.

Physical processes are inherently complex and embody uncertainties. They manifest as machine and citizen sensor observations in PCS Systems. We propose a generic framework to move from observations to decision-making and actions in PCS systems consisting of: (a) PCS event extraction, (b) PCS event understanding, and (c) PCS action recommendation. We demonstrate the role of Probabilistic Graphical Models (PGMs) as a unified framework to deal with uncertainty, complexity, and dynamism that help translate observations into actions. Data driven approaches alone are not guaranteed to be able to synthesize PGMs reflecting real-world dependencies accurately. To overcome this limitation, we propose to empower PGMs using the declarative domain knowledge. Specifically, we propose four techniques: (a) automatic creation of massive training data for Conditional Random Fields (CRFs) using domain knowledge of entities used in PCS event extraction, (b) Bayesian Network structure refinement using causal knowledge from Concept Net used in PCS event understanding, (c) knowledge-driven piecewise linear approximation of nonlinear time series dynamics using Linear Dynamical Systems (LDS) used in PCS event understanding, and the (d) transforming knowledge of goals and actions into a Markov Decision Process (MDP) model used in PCS action recommendation.

We evaluate the benefits of the proposed techniques on real-world applications involving traffic analytics and Internet of Things (IoT).

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Knowledge-empowered Probabilistic Graphical Models for Physical-Cyber-Social Systems

  1. 1. Knowledge-­‐empowered  Probabilis3c  Graphical   Models  for  Physical-­‐Cyber-­‐Social  Systems   Pramod  Anantharam   PhD  Disserta+on  Defense   April  14,  2016   The  Ohio  Center  of  Excellence  in  Knowledge-­‐enabled  Compu+ng  (Kno.e.sis),   Wright  State  University     Commi%ee:  Dr.  Payam  Barnaghi  (University  of  Surrey),  Dr.  Shalini  Forbis  (BoonshoP  School   of  Medicine),  Dr.  Cory  Henson  (Bosch  Research),  Dr.  Biplav  Srivastava  (IBM  Research),     Prof.  Shaojun  Wang  (Wright  State  University/Alibaba)   Advisors:  Prof.  Amit  Sheth,  Prof.  Krishnaprasad  Thirunarayan    
  2. 2. 2   Multimodal Manifestation of Real-World Events: Power Grid Scenario Image  Credit:  Twi%er,  hUp://bit.ly/1SsE924     1Six  Degrees:  The  Science  of  Connected  Age,  Duncan  WaUs   2One  of  four  main  reasons  of  failure.  Inves+ga+on  report  by  The  U.S.-­‐Canada  Power  System  Outage  Task  Force         August  14,  2003  Blackout  in  the  Midwest  U.S.   "failed  to  manage  adequately  tree  growth  in   its  transmission  right-­‐of-­‐way.”  2     August  10,  1996  Blackout  in  the  West  U.S.   “  …  inadequate  understanding  of  the   interdependencies  present  in  the  system.”  1     Power  Grid  related  events  manifest  in  physical,  cyber,  and  social  (PCS)  modali+es  
  3. 3. 3   Multimodal Manifestations of Real-World Events: Asthma Scenario Image  Credit:  hUp://www.rtmagazine.com/2015/10/brown-­‐univ-­‐fight-­‐childhood-­‐asthma-­‐au+sm-­‐obesity/     NODE Sensor (exhaled Nitric Oxide) Fitbit ChargeHR (Activity, sleep quality) Sensordrone (Carbon monoxide, temperature, humidity) Pollen level Temperature & Humidity Air  Quality   Prevalence of Asthma Personal   Level  Signals   Popula+on   Level  Signals   Asthma  related  events  manifest  in  physical,  cyber,  and  social  (PCS)  modali+es  
  4. 4. Multimodal Manifestation of Real-World Events: Traffic Scenario 4   Traffic  related  events  manifest  in  physical,  cyber,  and  social  (PCS)  modali+es   Amit  Sheth,  Pramod  Anantharam,  Cory  Henson,  'Physical-­‐Cyber-­‐Social  Compu+ng:  An  Early  21st  Century  Approach,'  IEEE  Intelligent  Systems,  vol.  28,  no.  1,  pp.   78-­‐82,  Jan.-­‐Feb.,  2013.  hUp://doi.ieeecomputersociety.org/10.1109/MIS.2013.20    
  5. 5. Processing Multimodal Manifestations of Real-World Events 5   “Informa3on  is  a  source  of  learning.  But  unless  it  is  organized,  processed,  and  available  to   the  right  people  in  a  format  for  decision  making,  it  is  a  burden,  not  a  benefit.”                        —  William  Pollard,  (1828  –  1893)   “…the  OODA  Loop  is  an  explicit  representa+on  of  the  process  that  human  beings  and   organiza+ons  use  to  learn,  grow,  and  thrive  in  a  rapidly  changing  environment  —  be  it   in  war,  business,  or  life.”1   1The  Tao  of  Boyd:  How  to  Master  the  OODA  Loop:  hUp://www.artofmanliness.com/2014/09/15/ooda-­‐loop/     Observe   Orient   Decide   Act   John  Boyd’s  Observe,  Orient,  Decide,  and  Act  (OODA)  Loop  for  organizing,  processing,   and  decision  making:   Feedback   Feed   Forward   Feed   Forward   Feed   Forward  
  6. 6. Processing Multimodal Manifestations in PCS Systems 6   The  Tao  of  Boyd:  How  to  Master  the  OODA  Loop:  hUp://www.artofmanliness.com/2014/09/15/ooda-­‐loop/     Observe   Orient   Decide   Act   Feedback   Feed   Forward   Feed   Forward   Feed   Forward   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   Observe  –  Collect  as  much  informa+on  as  possible  from  the  environment   Orient  –  Assimilate  all  the  informa+on  to  understand  the  environment   Decide  –  Determine  the  course  of  ac+on  based  on  an  objec+ve   Act  –  Follow  through  the  course  of  ac+on  
  7. 7. 7   Thesis Statement Observa3ons   from   diverse   modali3es   can   provide   complementary,   corrobora3ve,   and   3mely  informa+on  about  events  in  Physical-­‐Cyber-­‐Social  systems.  Probabilis3c  Graphical   Models  with  the  help  of  declara3ve  domain  knowledge  provide  an  effec+ve  mechanism  to:   (a)  uncover  and  interpret  mul3modal  event  manifesta3ons  in  textual  and  numerical  data,   (b)   explore   event   interac3ons   and   dynamics,   and   (c)   formalize   op3mal   ac3on   recommenda3on  in  Physical-­‐Cyber-­‐Social  systems.  
  8. 8. 8   “Graphical  models  are  a  marriage  between  probability  theory  and  graph  theory.  They   provide  a  natural  tool  for  dealing  with  two  problems  that  occur  throughout  applied   mathema3cs  and  engineering  -­‐-­‐  uncertainty  and  complexity  …”                                      -­‐  Michael  Jordan,  UC  Berkley,  1998.   What are Probabilistic Graphical Models (PGMs)? Alex  wants  to  model  the  reasons  for  asthma  a%acks.   Random  Variables:  AUack  (A),  Medica+on  (M),  Steps  (S),  Pollen  (P)   Joint  Probability  distribu3on:  p(A,  M,  S,  P)   Parameters:  For  four  binary  variables,  there  are  24  =  16  probability  assignments1     p(A,  M,  S,  P)  =  p(A  |  M,  S,  P)  p(M,  S,  P)                =  p  (A  |  M,  S,  P)  p(M  |  S,  P)  p(S,  P)                =  p  (A  |  M,  S,  P)  p(M  |  S,  P)  p(S  |  P)  P(P)                  =  p  (A  |  M,  P)  p(M)  p(S  |  P)  p(P),  because,   #  of  parameters  =  22  +  1  +  2  +  1  =  8  probability  assignments   (A ! S),(M ! S),(M ! P) A   M P   S Structure:         Parameters:   (8  probability  assignments)       1hUp://www.freemars.org/jeff/2exp100/powers.htm     p  (A  |  M,  P)     p(M)   p(S  |  P)     p(P)  
  9. 9. 9   Example of Declarative Domain Knowledge road  ice   Causes   accident   Linked  Open  Data   (Declara+ve  Knowledge  from  ConceptNet  5)   Delay   go  to  baseball  game   traffic  jam   traffic  accident   traffic  jam   Ac+veEvent   ScheduledEvent   Causes   traffic  jam   Causes   traffic  jam   CapableOf   slow  traffic   CapableOf   occur  twice  each  day   Causes   is_a   bad  weather   CapableOf     slow  traffic   TimeOfDay   go  to  concert   HasSubevent   car  crash   accident   RelatedTo   car  crash   BadWeather   Causes   Causes   is_a   is_a   is_a   is_a   is_a   is_a   is_a  
  10. 10. Processing Multimodal Manifestations in PCS Systems 10   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   •  What  are  the  events  of  interest?     •  How  do  they  manifest  in  observa+onal  data?     •  How  can  we  extract  events  from  observa+onal  data?     •  What  is  the  role  of  declara+ve  knowledge  in  event  extrac+on?   •  How  do  events  influence  one  another?   •  How  do  we  infer  the  interac3ons  from  observa3onal  data  across   mul3ple  modali3es  (numerical  and  textual  data)?     •  What  is  the  role  of  declara+ve  knowledge  in  event  understanding?   •  How  can  we  represent  tasks  and  ac+ons?   •  How  can  we  u+lize  declara+ve  knowledge  to  recommend  ac+ons?     •  How  can  we  formalize  the  no+on  of  op+mal  ac+on?   [ACM-­‐TIST-­‐15]     [ITS-­‐13]     [AAAI-­‐16]  [SDM-­‐13]       [IEEE-­‐Int.-­‐Sys.-­‐13]     [IBM-­‐Tech.-­‐Rep.-­‐14]     [Bosch-­‐Internship-­‐14]  
  11. 11. Processing Multimodal Manifestations in PCS Systems 11   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   •  What  are  the  events  of  interest?     •  How  do  they  manifest  in  observa+onal  data?     •  How  can  we  extract  events  from  observa+onal  data?     •  What  is  the  role  of  declara+ve  knowledge  in  event  extrac+on?   •  How  do  events  influence  one  another?   •  How  do  we  infer  the  interac3ons  from  observa3onal  data  across   mul3ple  modali3es  (numerical  and  textual  data)?     •  What  is  the  role  of  declara+ve  knowledge  in  event  understanding?   •  How  do  we  u+lize  our  understanding  to  recommend  ac+ons?     •  How  can  we  recommend  best  possible  ac+on?     •  What  is  the  role  of  declara+ve  knowledge  and  PGMs  in  ac+on   recommenda+on?   [AAAI-­‐16]  [SDM-­‐13]       [IEEE-­‐Int.-­‐Sys.-­‐13]     [ASG-­‐14]       [AAAI-­‐16]  Understanding  City  Traffic  Dynamics  U+lizing  Sensor  and  Textual  Observa+ons.  The  Thir+eth   AAAI  Conference  on  Ar+ficial  Intelligence,  2016   [SDM-­‐13]  Traffic  Analy+cs  using  Probabilis+c  Graphical  Models  Enhanced  with  Knowledge  Bases,  2nd   Interna+onal  Workshop  on  Analy+cs  for  Cyber-­‐Physical  Systems  (ACS-­‐2013)  at  SIAM  Interna+onal   Conference  on  Data  Mining  (SDM13),  2013   [IEEE-­‐Int.-­‐Sys.-­‐13]  Physical-­‐Cyber-­‐Social  Compu+ng:  An  Early  21st  Century  Approach,  IEEE  Intelligent   Systems,  2013   [ACM-­‐TIST-­‐15]     [ITS-­‐13]    
  12. 12. •  Why?   –  Explain/Interpret   average   speed   and   link   travel   +me   varia+ons   using   events   provided   by   city   authori+es   and   traffic  events  shared  on  TwiUer   –  Prior  work:  Predict  conges+on  based  on  historical  sensor   data   •  What?   –  Combine   •  511.org  data  about  Bay  Area  Road  Network  Traffic     –  E.g.,  Average  speed  and  link  travel  +me  data  stream  (Sensor  data)   –  E.g.,  (Happened  or  planned)  event  reports  (Textual  data)   •  Tweets  that  report  traffic  related  events  (Textual  data)   Multimodal Data Integration: Traffic Scenario 12  
  13. 13. •  How?   o  Step  1:  Extract  textual  events  from  tweets  stream   o  Step  2:  Build  sta+s+cal  models  of  normalcy,  and  thereby   anomaly,  for  sensor  +me  series  data   o  Step  3:  Correlate  mul3modal  streams,  using  spa+o-­‐ temporal  informa+on,  to  explain  “anomalies”  in  sensor   +me  series  data  with  textual  events   Multimodal Data Integration: Traffic Scenario 13  
  14. 14. •  How?   o  Step  1:  Extract  textual  events  from  tweets  stream   o  Step  2:  Build  sta+s+cal  models  of  normalcy,  and  thereby   anomaly,  for  sensor  +me  series  data   o  Step  3:  Correlate  mul3modal  streams,  using  spa+o-­‐ temporal  informa+on,  to  explain  “anomalies”  in  sensor   +me  series  data  with  textual  events   Multimodal Data Integration: Traffic Scenario 14  
  15. 15. Processing Multimodal Manifestations in PCS Systems 15   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   •  What  are  the  events  of  interest?     •  How  do  they  manifest  in  observa+onal  data?     •  How  can  we  extract  events  from  observa3onal  data?     •  What  is  the  role  of  declara+ve  knowledge  in  event  extrac+on?   •  How  do  events  influence  one  another?   •  How  do  we  infer  the  interac3ons  from  observa3onal  data  across   mul3ple  modali3es  (numerical  and  textual  data)?     •  What  is  the  role  of  declara+ve  knowledge  and  PGMs  in  event   understanding?   •  How  can  we  represent  tasks  and  ac+ons?   •  How  can  we  u+lize  declara+ve  knowledge  to  recommend  ac+ons?     •  How  can  we  formalize  the  no+on  of  op+mal  ac+on?   [ACM-­‐TIST-­‐15]     [ITS-­‐13]     [AAAI-­‐15]  [SDM-­‐13]       [IEEE-­‐Int.-­‐Sys.-­‐13]     [ACM-­‐TIST-­‐15]  Extrac+ng  City  Traffic  Events  from  Social  Streams.  ACM  Transac+ons  on  Intelligent   Systems  and  Technology  Journal  2015.   [ITS-­‐13]  City  No+fica+ons  as  a  Data  Source  for  Traffic  Management,  20th  ITS  World  Congress  2013.         [IBM-­‐Tech.-­‐Rep.-­‐14]     [Bosch-­‐Internship-­‐14]  
  16. 16. 16   People Reporting Various Events in a City on Twitter Public  Safety   Urban  planning   Gov.  &  agency     admin.   Energy  &  water   Environmental   Transporta3on   Social  Programs   Healthcare   Educa+on  
  17. 17. 17   Extracting City Events from Twitter: Proposed Solution [ACM-­‐TIST-­‐15]  Extrac+ng  City  Traffic  Events  from  Social  Streams.  ACM  Transac+ons  on  Intelligent  Systems  and  Technology  Journal  2015.   Event  Extrac+on  Tool  on  Open  Science  Founda+on:  hUps://osf.io/b4q2t/wiki/home/    
  18. 18. 18   Label  image  sequence  of  Jus+n  Bieber’s  day  J     Sleeping   Driving   Exercising Driving   Sleeping   Singing   This  image  of  concert  was   Important  in  labeling  the  next  image   Edwin  Chen’s  blog  on  CRF:  hUp://blog.echen.me/2012/01/03/introduc+on-­‐to-­‐condi+onal-­‐random-­‐fields/     Image  Credit:  hUp://bit.ly/1Th8CgL,  hUp://bit.ly/1Nzk5DR,  hUp://bit.ly/1VBbx7e,  hUp://bit.ly/1QkmBhb,  hUp://bit.ly/1SsyYzd,   hUp://bit.ly/1Nzl7j7     City Event Annotation: Conditional Random Fields (CRFs) – Intuition
  19. 19. 19   The  global  normaliza+on  and  the  discrimina+ve  nature  of  the  model  dis+nguishes   CRFs  from  other  models  allowing  it  to  capture  long  distance  dependencies     City Event Annotation: Conditional Random Fields (CRFs) – Formalism Last  O  night  O  I  O  was  O  in  O  CA...  O  (@  O  Half  B-­‐LOCATION  Moon  I-­‐LOCATION  Bay  B-­‐ LOCATION  Brewing  I-­‐LOCATION  Company  O  w/  O  8  O  others)  O  hUp://t.co/w0eGEJjApY  O     {B-­‐LOCATION,  I-­‐LOCATION,  B-­‐EVENT,  I-­‐EVENT,  O}  Tagset  =  
  20. 20. 20   0.6  miles   Max-­‐lat   Min-­‐lat   Min-­‐long   Max-­‐long   0.38  miles   37.7545166015625, -122.40966796875   37.7490234375, -122.40966796875   37.7545166015625,  -122.420654296875   37.7490234375, -122.420654296875   4   37.74933, -122.4106711   Hierarchical  spa+al  structure  of  geohash  for     represen+ng  loca+ons  with  variable  precision.   Here,  the  loca+on  string  is  5H34   0   1   2   3   4   5   6   7   8   9   B   C   D   E   F   G   H   I   J   K   L   0   1   7   2   3   4   5   6   8   9   0   1   2   3   4   5   6   7   0   1   2   3   4   5   6   7   8   Geohashing  wiki:  hUp://wiki.xkcd.com/geohashing/   Image  Credit:  Google  Maps     City Event Extraction: Spatio-Temporal-Thematic Aggregation
  21. 21. 21   •  City  Event  Annota+on   –  Automated  crea+on  of  training  data     –  Annota+on  task  (our  CRF  model  vs.  baseline  CRF  model)   •  City  Event  Extrac+on   –  Use  aggrega+on  algorithm  for  event  extrac+on   –  Extracted  events  vs.  ground  truth   •  Dataset  (Aug  –  Nov  2013)   –  Over  8  million  tweets  from  San  Francisco  Bay  Area  (extracted   1042  events)   –  311  ac+ve  events  and  170  scheduled  events  from  511.org   (ground  truth)   Evaluation: Extracting City Events from Twitter
  22. 22. Evaluation: City Event Annotation 22   Baseline  Annota+on  Model  [RiUer  et  al.  2012]   Our  Annota+on  Model   •  Baseline  CRF  model  (trained  on  a  huge  manually  created  data)  works  well  on  generic   tasks   •  Our  CRF  model  trained  on  automa+cally  generated  training  data  performs  on  par   with  the  baseline   •  Our  CRF  model  does  beUer  on  the  event  extrac+on  task  due  to  the  availability  of   event  related  knowledge     [RiUer  et  al.  2012]  Alan  RiUer,  Mausam,  Oren  Etzioni,  and  Sam  Clark  2012.  Open  domain  event  extrac+on  from  TwiUer.  In  Proceedings  of   the  18th  ACM  SIGKDD  Interna+onal  Conference  on  Knowledge  Discovery  and  Data  Mining.  ACM,  New  York,  NY,  1104–1112.  
  23. 23. Complementary  Events   Textual Events from Tweets vs. 511.org: Complementary 23   traffic   incident;  road-­‐construc+on  
  24. 24. Textual Events from Tweets vs. 511.org: Corroborative Corrobora+ve  Events   24   fog   visibility-­‐air-­‐quality;  fog  
  25. 25. Timeliness   Textual Events from Tweets vs. 511.org: Timeliness 25   concert   concert  
  26. 26. Extracting Textual Events from Tweets for Data from May-14 to May-15 1Event  Extrac+on  Tool  on  Open  Science  Founda+on:  hUps://osf.io/b4q2t/wiki/home/     NER  –  Named  En+ty  Recogni+on   OSM  –  Open  Street  Maps   39,208  traffic  related  incidents  extracted  from  over  20  million  tweets1   26   [ACM-­‐TIST-­‐15]  Extrac+ng  City  Traffic  Events  from  Social  Streams.  ACM  Transac+ons  on  Intelligent  Systems  and  Technology  Journal  2015.  
  27. 27. •  How?   o  Step  1:  Extract  textual  events  from  tweets  stream   o  Step  2:  Build  sta+s+cal  models  of  normalcy,  and  thereby   anomaly,  for  sensor  +me  series  data   o  Step  3:  Correlate  mul3modal  streams,  using  spa+o-­‐ temporal  informa+on,  to  explain  “anomalies”  in  sensor   +me  series  data  with  textual  events   Multimodal Data Integration: Traffic Scenario 27  
  28. 28. Image  credit:  hUp://traffic.511.org/index     Mul+ple  events     Varying  influence     Event  interac+ons   Time  of  Day  (approx.  1  observa+on/minute)  Speed  in  km/h   Building Normalcy Models of Traffic Dynamics*: Challenges *Traffic  Dynamics  here  refers  to  speed  and  travel  +me  varia+ons  observed  in  sensor  data   28  
  29. 29. •  Temporal  landmarks  :  peak  hour  vs.  off-­‐peak  traffic   vs.  weekend  traffic   •  Effect  of  loca+on   •  Scheduled  events  such  as  road  construc+on,  baseball   game,  or  music  concert   •  Unexpected   events   such   as   accidents,   heavy   rains,   fog   •  Random  varia+ons  (viz.,  stochas+city)  such  as  people   visi+ng  downtown  by  mere  coincidence     Possible Causes of Nonlinearity in Traffic Dynamics 29  
  30. 30. Modeling City Traffic Dynamics: A Closer Look Image  credits:  hUp://bit.ly/1N1wu5g,  hUp://bit.ly/1O8d9gn,  hUp://bit.ly/1N8L5•,  hUp://bit.ly/1HLDYui         Events   People   Influx   Vehicle   Influx   Vehicle   Speed   Hidden  State   Observed  Evidence   30   link1   link2   link3   road1  =  [link1,link2,link3]  
  31. 31. Modeling City Traffic Dynamics: Nature of the Problem Hidden  States   Observed  Evidence  1.  There  are  both  hidden  states  and  observed  evidence   2.  Current  observed  evidence  indica3ve  of    the  current  hidden  state   3.  Current  hidden  states  depends  on  the  previous  hidden  states   T  is  a  discrete  3me  step  in  the   3me  series  data  being  modeled   31   Events   People   Influx   Vehicle   Influx   Events   (T)   People   Influx   (T)   Vehicle   Influx  (T)   Events   (T)   People   Influx   (T)   Vehicle   Influx   (T)   Events   (T-­‐1)   People   Influx   (T-­‐1)   Vehicle   Influx   (T-­‐1)   Vehicle   Speed   Vehicle   Speed   (T)  
  32. 32. Modeling the Problem as Linear Dynamical System (LDS) 1.  There  are  both   hidden  states  and   observed  evidence   2.  Current  observed   evidence  indica3ve  of  the   current  hidden  state   3.  Current  hidden   state  depends  on   the  previous   hidden  state   v1   s1   …   …   v2   s1   vT   sT   v1   s1   …   …   v2   s1   vT   sT   v1   s1   …   …   v2   s1   vT   sT   For  simplicity  of  explana+on,  we   consider  vehicle  influx  as  a   hidden  variable  and  the  observed   speed  as  evidence     variable   Vehicle  influx  at  a  certain  point  in   +me  t  would  influence  speed  of   vehicles  at  the  same  +me  t   Vehicle  influx  at  a  certain  point  in   +me  t  depends  only  on  the   vehicle  influx  at  +me  t-­‐1   32  
  33. 33. Probabilistic Reasoning Over Time: Discrete Variables Russell,  Stuart,  and  Peter  Norvig.  "Ar+ficial  intelligence:  a  modern  approach."  (1995).   Image  credits:  hUp://bit.ly/1Q9qmvk,  hUp://bit.ly/1lm9BAs,  hUp://bit.ly/1LXqOFd     Evidence  (U)   States  (R)   State  transi+on  model  is  given  by     With  First-­‐Order  Markov  assump3on,   the  transi+on  model  is     Transi3on  model   Observa3on  model   Observa+on  model  with  sensor  Markov   assump3on  is  given  by   P(Rt  |  R0:t-­‐1)   P(Rt  |  Rt-­‐1)   P(Ut  |  R0:t,U0:t-­‐1)  =  P(Ut  |  Rt)     Specifying  t  transi+on  and  observa+on  models   is  imprac+cal.  So,  another  assump+on:   sta3onary  process   Rt-­‐1        P(Rt)   t                    0.7   f                    0.3   Rt        P(Ut)   t                    0.9   f                    0.1   33  
  34. 34. Probabilistic Reasoning Over Time: Continuous Variables v1   s1   …   …   v2   s1   vT   sT   Linear  Dynamical  System  (LDS):  Replacing   discrete  valued  state  and  observa+on  nodes   (previous  slide)  with  conHnuous  valued  states   and  observa+ons,  we  get  an  LDS  model   The  transi3on  model  is  specified  by  At   and  the  observa3on  model  is  specified  by   Bt  along  with  associated  Gaussian  noise   The  joint  distribu+on  over  all  the  hidden   and  observed  variables  is  shown  along   with  the  condi+onal  distribu+ons   Barber,  David.  Bayesian  reasoning  and  machine  learning.  Cambridge  University  Press,  2012.   34  
  35. 35. Hourly Link Speed Dynamics Over all Mondays between Aug-14 to Jan-15 x-­‐axis:  observa3on  number  for  each  hour  of  day   y-­‐axis:  average  speed  of  vehicles  in  km/h     35  
  36. 36. 36   Switching Linear Dynamical Systems v1   s1   …   …   v2   s1   vT   sT   h1   h2   hT  …   Switching  Linear  Dynamical  System  (SLDS):  A   discrete  switch  variable  at  each  +me  t  describes   the  appropriate  LDS  to  be  used.  SLDS  can  capture   jumps  between  mul3ple  linear  dynamics.     v1   s1   …   …   v2   s1   vT   sT   h1   h2   hT  …   Restricted  Switching  Linear  Dynamical  System   (RSLDS):  Restric+ng  the  switch  variable  transi+ons  in   SLDS,  we  proposed  RSLDS  [AAAI-­‐16]  which  captures   the  switching  behavior  based  on  hour  of  the  day  and   day  of  the  week.   The  transi3on  model  is  specified  by  At(ht)  and  the   observa3on  model  is  specified  by  Bt(ht)   [AAAI-­‐16]  Understanding  City  Traffic  Dynamics  U+lizing  Sensor  and  Textual  Observa+ons.  The  Thir+eth   AAAI  Conference  on  Ar+ficial  Intelligence,  2016  
  37. 37. Modeling City Traffic Dynamics: Choosing a Suitable Model "All  models  are  wrong,  but  some  are  useful.”  -­‐  George  Box   •  Differen+ate  various  traffic  dynamics   –  Gaussian  mixture  model  does  not  discriminate  between   increasing  speed  vs.  decreasing  speed  dynamics   •  Account  for  unobserved  factors   –  Autoregressive  models  cannot  capture  unobserved  factors   •  E.g.,  “Unobservable”  traffic  volume  dictates  event  manifesta+ons   in  link  speed  and  travel  +me  varia+ons   –  Linear  Dynamical  System  introduces  latent  state-­‐based   model   •  E.g.,  Traffic  volume,  road  lane  closures,  and  weather  condi+ons     •  Emission/Transi+on  matrix  and  Gaussian  noise  captures   stochas+city   37  
  38. 38. 38   Learning Context Specific LDS Models 7  ×  24   LDS(1,1),  LDS(1,2)      ,….,  LDS(1,24)   LDS(7,1),  LDS(7,2)      ,….,  LDS(7,24)   .   .   .   di   hj   Mon. Tue. Wed. Thu. Fri. Sat. Sun. Mon. Tue. Wed. Thu. Fri. Sat. Sun.Speed/travel-­‐+me  +me     series  data  from  a  link   Time  series  data  for   each  hour  of  day  (1-­‐24)   for  each  day  of  week   (Monday  –  Sunday)   Mean  +me  series   computed  for  each  day   of  week  and  hour  of  day   along  with  the  medoid   168  LDS  models  for   each  link;  Total  models   learned  =  425,712  i.e.,   (2,534  links  ×  168   models  per  link)       Step  1:  Index  data  for  each   link  for  day  of  week  and  hour   of  day  u+lizing  the  traffic   domain  knowledge  for  piece-­‐ wise  linear  approxima+on   Step  2:  Find  the  “typical”   dynamics  by  compu+ng  the   mean  and  choosing  the   medoid  for  each  hour  of  day   and  day  of  week   Step  3:  Learn  LDS  parameters   for  the  medoid  for  each  hour   of  day  (24  hours)  and  each  day   of  week  (7  days)  resul+ng  in   24  ×  7  =  168  models  for  each   link  
  39. 39. Learning Normalcy for Each Link, Day of Week, and Hour of Day Log-­‐likelihood      score   39   Five-­‐number  summary  of  log-­‐likelihood  scores  for  a  link,  day  of  week,  hour  of  day  
  40. 40. 40   Tagging Anomalies using Context Specific LDS Models Compute  Log  Likelihood  for     each  hour  of  observed  data   (di,hj)   LDS(hj,di)   7  ×  24   Lik(1,1),  Lik(1,2)      ,….,  Lik(1,24)   Lik(7,1),  Lik(7,2)      ,….,  Lik(7,24)   .   .   .   Train ?   Yes  (Training  phase)   Tag  Anomalous  hours  using  the   Log  Likelihood  Range   No   (di,hj)   (min.  likelihood)   Anomalies   L  =   Par33on  based  on  (di,hj)   Speed  and  travel-­‐+me  +me     Observa+ons  from  a  link   Log  likelihood  min.  and     max.  values  obtained  from     five  number  summary   Par33on  based  on  (di,hj)   7  ×  24   LDS(1,1),  LDS(1,2)      ,….,  LDS(1,24)   LDS(7,1),  LDS(7,2)      ,….,  LDS(7,24)   .   .   .   di   hj   (Input)   (Output)  
  41. 41. •  How?   o  Step  1:  Extract  textual  events  from  tweets  stream   o  Step  2:  Build  sta+s+cal  models  of  normalcy,  and  thereby   anomaly,  for  sensor  +me  series  data   o  Step  3:  Correlate  mul3modal  streams,  using  spa+o-­‐ temporal  informa+on,  to  explain  “anomalies”  in  sensor   +me  series  data  with  textual  events   Multimodal Data Integration: Traffic Scenario 41  
  42. 42.   •  Anomaly  in  link  data  during  +me  period  [ast,aet],  is   explained  by  an  event  if  the  event  occurs  within   0.5km  radius  and  during  [ast-­‐1,  aet+1].   •  CAVEAT:  An  anomaly  may  not  be  explained  because   of  missing  data.     Explaining Anomalies in Sensor Data using Textual Events 42   Anomalies   ⟨et,  el,  est,  eet,  ei⟩   Explained_by     Link  sensor  data   City  tweets   ⟨ast,  aet⟩   Δte  =  est  ~  eet   Δta  =  (ast  –  1)  ~  (aet  +  1)   Explains   (if  there  is  an  overlap     between  Δte  and  Δta)   PCS  Event     Extrac3on  
  43. 43. •  Data  collected  from  San  Francisco  Bay  Area  between   May  2014  to  May  2015   –  511.org:  (1)  1,638  traffic  incident  reports  (2)  1.4  billion   speed  and  travel  +me  observa+ons   –  TwiUer  Data:  39,208  traffic  related  incidents  extracted   from  over  20  million  tweets   •  Learning  normalcy  model  for  one  link  takes  40   minutes1  (~  2  months  for  processing  2,534  links)   •  Scalable  implementa+on  on  Apache  Spark2  resulted   in  learning  normalcy  models  for  2,534  links  within  24   hours   Real-World Dataset and Scalability Issues 43   12.66  GHz,  Intel  Core  2  Duo  with  8  GB  main  memory  machine   2Cluster  used  for  evalua+on  had  865  cores  and  17TB  main  memory  
  44. 44. Multimodal Data Integration: Evaluation 44  
  45. 45. •  Examined  the  theore3cal  nature  of  the  problem  of   modeling  traffic  dynamics  to  systema+cally   recommend  Linear  Dynamical  Systems  (LDS)   •  Formalized  nonlinear  traffic  dynamics  using   piecewise  linear  approxima+on  derived  from  traffic   domain  knowledge   •  Created  normalcy  models  based  on  log-­‐likelihood   scores  for  spo‡ng  traffic  anomalies  in  sensor  data   •  Evaluated  our  approach  over  a  real-­‐world  dataset   collected  from  511.org  and  TwiUer  for  over  a  year   (May-­‐2014  to  May  2015)  with  promising  results   45   Multimodal Data Integration: Conclusion
  46. 46. Processing Multimodal Manifestations in PCS Systems 46   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   •  What  are  the  events  of  interest?     •  How  do  they  manifest  in  observa+onal  data?     •  How  can  we  extract  events  from  observa3onal  data?     •  What  is  the  role  of  declara+ve  knowledge  and  PGMs  in  event   extrac+on?   •  How  do  events  influence  one  another?   •  How  do  we  infer  the  interac3ons  from  observa3onal  data  across   mul3ple  modali3es  (numerical  and  textual  data)?     •  What  is  the  role  of  declara+ve  knowledge  and  PGMs  in  event   understanding?   •  How  can  we  represent  tasks  and  ac+ons?   •  How  can  we  u+lize  declara+ve  knowledge  to  recommend  ac+ons?     •  How  can  we  formalize  the  no+on  of  op+mal  ac+on?   [ATMSB-­‐15]  [ATS-­‐13]       [SAH-­‐13]     [IBM-­‐Tech.-­‐Rep.-­‐14]     [Bosch-­‐Internship-­‐14]         [IBM-­‐Tech.-­‐Rep.-­‐14] Dynamic  Update  of  Public  Transport  Schedules  in  Ci+es  Lacking  Traffic  Instrumenta+on,  IBM   Research  Technical  Report  2014.   [Bosch-­‐Internship-­‐14]  Task  Assistance  within  IoTS  Network,  Bosch  Summer  Internship  Work,  2014.   [ACM-­‐TIST-­‐15]     [ITS-­‐13]    
  47. 47. •  Contributed  to  a  language  to  represent  tasks   –  Using  Seman+c  Web  based  representa+on  for   •  Reusing  knowledge  on  the  web   •  Integra+on  of  knowledge  in  distributed  environments        (like  the  web  and  UhU1  /  IoTS  network)   •  Developed  algorithms  to  recommend  tasks     –  Formulated  the  problem  of  recommending  op+mal  ac+on   toward  a  goal2  by  handling  task  failure  in  a  robust  manner   •  Developed  a  framework  to  evaluate  task   recommenda+on   –  Using  a  simulator  for  world  states  and  user  ac+ons   47   Do-It-Yourself (DIY) Task Recommendation: Bosch Internship, 2014 1Bosch  IoT  middleware   2  Op+mal  ac+on  is  formulated  as  a  Markov  Decision  Process  with  transi+on  and  cost  matrices  ini+alized  using  declara+ve  knowledge  of  tasks    
  48. 48. Revisiting the Thesis Statement 48   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   [ACM-­‐TIST-­‐15]     [ITS-­‐13]     [AAAI-­‐16]  [SDM-­‐13]       [IEEE-­‐Int.-­‐Sys.-­‐13]     [IBM-­‐Tech.-­‐Rep.-­‐14]     [Bosch-­‐Internship-­‐14]   U3lize  declara3ve  knowledge   of  loca3ons  and  events  to   train  sequence  labeling  models   for  annota3on  and  event   extrac3on   U3lize  declara3ve  knowledge   of  ac3ons  to  formulate  the   problem  of  op3mal  ac3on   recommenda3on  as  a   sequen3al  decision  problem       U3lize  textual  events  to   explain  varia3ons  in  sensor   data  modeled  using  context   (link,  loca3on,  3me)  specific   probabilis3c  3me  series   models     Observa3ons   from   diverse   modali3es   can   provide   complementary,   corrobora3ve,   and   3mely  informa+on  about  events  in  Physical-­‐Cyber-­‐Social  systems.  Probabilis3c  Graphical   Models  with  the  help  of  declara3ve  domain  knowledge  provide  an  effec+ve  mechanism  to:   (a)  uncover  and  interpret  mul3modal  event  manifesta3ons  in  textual  and  numerical  data,   (b)   explore   event   interac3ons   and   dynamics,   and   (c)   formalize   op3mal   ac3on   recommenda3on  in  Physical-­‐Cyber-­‐Social  systems.  
  49. 49. 49   Conclusion •  Observa+ons  from  people  can  provide  complementary,   corrobora3ve,  and  3mely  informa+on  in  PCS  systems.   •  We  demonstrated  that  probabilis+c  graphical  models   (PGMs)  are  a  natural  fit  to  deal  with  PCS  challenges.   •  We  found  that  declara3ve  domain  knowledge  can   complement  PGMs  in   –  Automa+c  crea+on  of  large  training  data  for  training  sequence   labeling  models   –  Knowledge-­‐driven  piecewise  linear  approxima+on  of  nonlinear   +me  series  dynamics  using  Linear  Dynamical  Systems  (LDS)   –  Bayesian  Network  structure  refinement  using  ConceptNet5   –  Transforming  knowledge  of  goals  and  ac+ons  into  a  Markov   Decision  Process  (MDP)  formalism  
  50. 50. 50   Probabilistic Graphical Models, Declarative Knowledge, and PCS Systems Declara+ve   Knowledge   Data   Textual   Numerical   Parameters   Annotate   Parameters   Structure   PGMs  (e.g.,  CRF,  BN,  LDS,  MDP)   PCS  Applica+ons  (e.g.,  SmartCity,  SmartHealth,  DIY  Task  Recommenda+on)   Commonsense     Knowledge   Domain  Ontologies     and  Open  Data   Mul+modal  Data   Top-­‐down   Bokom-­‐up   PCS  Event     Extrac3on   PCS  Event     Understanding   PCS  Ac3on     Recommenda3on   [ACM-­‐TIST-­‐15]     [AAAI-­‐16]     [ACM-­‐TIST-­‐15]     [Bosch-­‐Internship-­‐14]   [SDM-­‐13]   CRF  –  Condi+onal  Random  Field   BN  –  Bayesian  Network   LDS  –  Linear  Dynamical  Systems   MDP  –  Markov  Decision  Process   Structure   [SDM-­‐13]   [AAAI-­‐16]     [Bosch-­‐Internship-­‐14]  
  51. 51. 51   Personalized Digital Health for Asthma Management in Children Sensordrone (Carbon monoxide, temperature, humidity) Sensor Platforms Android Device (w/ kHealth App) Node Sensor (exhaled Nitric Oxide) Fitbit ChargeHR (Activity, sleep quality) Pollen level Air  Quality   Temperature & Humidity kHealth  for  asthma  project  page:  hUp://wiki.knoesis.org/index.php/Asthma     kHealth  project  page:  hUp://knoesis.org/projects/khealth    
  52. 52. 52   PhD @ Kno.e.sis Awards  and  Recogni3on   2016  Outstanding  Graduate  Student  Award  in  the  PhD  in  Computer  Science  and   Engineering  Program.   2015  Selected  to  par+cipate  in  the  NSF-­‐funded  Data  Science  Workshop  at  University  of   Washington,  SeaUle,  Aug  5–7.   2014  Offered  the  Eric  &  Wendy  Schmidt  Data  Science  for  Social  Good  Fellowship.   2013  A  short  ar+cle  on  my  research  appeared  in  Wright  State  University  newsroom.   2013  Invited  to  aUend  Dagstuhl  Seminar  on  Physical-­‐Cyber-­‐Social  Compu+ng.   2012  Best  research  showcase  award  for  my  internship  work  at  IBM  Research,  India.   Professional  Experience   •  2014  Internship  at  Bosch  Research  and  Technology  Center   •  2013  Visi+ng  Doctoral  Student  at  University  of  Surrey   •  2011,  2012  Internships  at  IBM  Research     Published  in  ACM  TIST  Journal,  AAAI,  ACM  Web   Science,  and  IEEE  Computer   Program  Commikee  (PC)  member  of   conferences  such  as  WWW-­‐16,  WWW-­‐15,   WWW-­‐14,  ISWC-­‐15,  ISWC-­‐14,  ISWC-­‐13,  ESWC-­‐16,   IJCAI-­‐13       Tutorials     •  Data  Processing  and  Seman+cs  for  Advanced  Internet  of  Things  (IoT)  Applica+ons:  modeling,  annota+on,  integra+on,   and  percep+on,  Tutorial  Presenta+on  at  The  3rd  Interna+onal  Conference  on  Web  Intelligence,  Mining  and   Seman+cs  (WIMS  '13),  Madrid,  Spain.   •  Trust  Networks:  Interpersonal,  Sensor,  and  Social,  Tutorial  Presenta+on  at  Interna+onal  Conference  on  Collabora+ve   Technologies  and  Systems  (CTS  2011),  Philadelphia,  Pennsylvania,  USA.   Proposals   NSF:  Contributed  to  mul+ple,  out   of  which,  one  collabora+ve   proposal  was  funded  ($1.9  million).   NIH:  Lead  one  proposal  which  is   recommended  for  funding  ($900K).   EU  FP7:  Contributed  to  CityPulse,  a   mul+-­‐ins+tu+on  IoT  based  Smart   City  project  (€2.5  million).   Patents   •  US20150006644  A1:  Assessing  Impact  of  Events  on  Public  Transporta+on  Network   •  US20140372364  A1:  A  System  and  Method  for  U+lity-­‐Based  Evolu+on  in  a  Constrained  Ontology  
  53. 53. [AAAI-­‐16]  Pramod  Anantharam,  Krishnaprasad  Thirunarayan,  Surendra  Marupudi,  Amit  Sheth,  Tanvi  Banerjee.  (2016)   Understanding  City  Traffic  Dynamics  U+lizing  Sensor  and  Textual  Observa+ons.  at  The  Thir+eth  AAAI  Conference  on   Ar+ficial  Intelligence  (AAAI-­‐16),  February  12-­‐-­‐17,  Phoenix,  Arizona,  USA  (accepted)   [ACM-­‐TIST-­‐15]  Pramod  Anantharam,  Payam  Barnaghi,  Krishnaprasad  Thirunarayan,  and  Amit  Sheth.  2015.   Extrac+ng  City  Traffic  Events  from  Social  Streams.  ACM  Trans.  Intell.  Syst.  Technol.  6,  4,  Ar+cle  43  (July  2015),  27  pages.   DOI=10.1145/2717317  hUp://doi.acm.org/10.1145/2717317       [IBM-­‐Tech.-­‐Rep.-­‐14]  Pramod  Anantharam,  Biplav  Srivastava,  Raj  Gupta.   Dynamic  Update  of  Public  Transport  Schedules  in  Ci+es  Lacking  Traffic  Instrumenta+on,  IBM  Research  Technical  Report   2014.   [ITS-­‐13]  Pramod  Anantharam  and  Biplav  Srivastava,  City  No+fica+ons  as  a  Data  Source  for  Traffic  Management,  In   Proceedings  of  the  20th  ITS  World  Congress  2013,  October  14-­‐18,  2013,  Tokyo,  Japan.   [SDM-­‐13]  Pramod  Anantharam,  Krishnaprasad  Thirunarayan,  and  Amit  Sheth,   Traffic  Analy+cs  using  Probabilis+c  Graphical  Models  Enhanced  with  Knowledge  Bases,  2nd  Interna+onal  Workshop  on   Analy+cs  for  Cyber-­‐Physical  Systems  (ACS-­‐2013)  at  SIAM  Interna+onal  Conference  on  Data  Mining  (SDM13),  Texas,  USA,   May  2-­‐4,  2013.   [ACM-­‐WebScience-­‐12]  Pramod  Anantharam,  Krishnaprasad  Thirunarayan,  and  Amit  Sheth,   Topical  Anomaly  Detec+on  from  TwiUer  Stream,  Research  Note:  In  the  Proceedings  of  ACM  Web  Science  2012,  Evanston,   Illinois,  pp.  23-­‐26,  June  22-­‐24,  2012.   [IEEE-­‐Int.-­‐Sys.-­‐13]  Amit  Sheth,  Pramod  Anantharam,  Cory  Henson,   Physical-­‐Cyber-­‐Social  Compu+ng:  An  Early  21st  Century  Approach,  IEEE  Intelligent  Systems,  vol.  28,  no.  1,  pp.  78-­‐82,  Jan.-­‐ Feb.,  2013.      hUp://doi.ieeecomputersociety.org/10.1109/MIS.2013.20     [FCGS-­‐13]  Krishnaprasad  Thirunarayan,  Pramod  Anantharam,  Cory  Henson,  and  Amit  Sheth,   Compara+ve  Trust  Management  with  Applica+ons:  Bayesian  Approaches  Emphasis,  In  the  Journal  of  Future  Genera+on   Computer  Systems  (FGCS),  Elsevier,  25  pages,  May  2013,  hUp://dx.doi.org/10.1016/j.future.2013.05.006     [Bosch-­‐Internship-­‐14]  Task  Assistance  within  IoTS  Network,  Bosch  Internship  Work,  Summer  2014.   53   Selected Publications
  54. 54. 54   Dr.  Payam  Barnaghi   Dr.  Biplav  Srivastava   Dr.  Cory  Henson   Dr.  Shalini  Forbis,  MD,  MPH   Prof.  Amit  Sheth   (Advisor)   Prof.  Krishnaprasad     Thirunarayan   (Advisor)   Prof.  Shaojun  Wang   Acknowledgements
  55. 55. Thank you J 55   kHealth  Team   Dr.  Tanvi  Banerjee   Surendra  Marupudi   Vaikunth  Sridharan     Dan  Vanuch   Sujan  Perera   And  all  my  colleagues  and  friends…   Vahid  Taslimi   Kno.e.sis,  Data  Mining  Lab  

×