Modeling Count Data over Time Using Dynamic Bayesian Networks Jonathan Hutchins Advisors: Professor Ihler and Professor Smyth
Optical People Counter at a Building Entrance Loop Sensors on Southern California Freeways Sensor Measurements Reflect Dynamic Human Activity
Outline Introduction, problem description Probabilistic model Single sensor results Multiple sensor modeling Future Work
Modeling Count Data In a Poisson distribution: mean = variance = λ p(count|λ) count
mean people count variance Simulated Data 15 weeks, 336 time slots
mean people count variance Building Data
mean people count variance Freeway Data
One Week of Freeway Observations
One Week of Freeway Data
Detecting Unusual Events: Baseline Method Ideal model car count Baseline model car count Unsupervised learning faces a “chicken and egg” dilemma
Persistent Events Notion of Persistence missing from Baseline model
Quantifying Event Popularity Ideal model Baseline model
My contribution Adaptive event detection with time-varying Poisson processes A. Ihler, J. Hutchins, and P. Smyth Proceedings of the 12th ACM SIGKDD Conference (KDD-06), August Baseline method, Data sets, Ran experiments Validation Learning to detect events with Markov-modulated Poisson processes A. Ihler, J. Hutchins, and P. Smyth ACM Transactions on Knowledge Discovery from Data, Dec 2007 Extended the model to include a second event type (low activity) Poisson Assumption Testing Modeling Count Data From Multiple Sensors: A Building Occupancy Model J. Hutchins, A. Ihler, and P. Smyth IEEE CAMSAP 2007,Computational Advances in Multi-Sensor Adaptive Processing, December 2007.
"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 mathematics and engineering -- uncertainty and complexity” Michael Jordan 1998 Graphical Models
Nodes variables Directed Graphical Models observed Observed Count hidden EventRate Parameter
Directed Graphical Models Nodes variables Edges direct dependencies A B C
Graphical Models: Modularity Observed Count t Observed Count t-2 Observed Count t-1 Observed Count t+2 Observed Count t+1
Graphical Models: Modularity hidden observed Poisson Rate λ(t) Normal Count t-1 Observed Count t Observed Count t-1 Observed Count t+1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1
Graphical Models: Modularity hidden observed Poisson Rate λ(t) Normal Count t-1 Observed Count t Observed Count t-1 Observed Count t+1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1
Graphical Models: Modularity Event t Event t-1 Event t+1 hidden observed Poisson Rate λ(t) Normal Count t-1 Observed Count t Observed Count t-1 Observed Count t+1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1
Graphical Models: Modularity Event t Event t-1 Event t+1 hidden observed Poisson Rate λ(t) Normal Count t-1 Observed Count t Observed Count t-1 Observed Count t+1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1 Event State Transition Matrix
Event t Event t-1 Event t+1 Event State Transition Matrix Observed Count t Observed Count t-1 Observed Count t+1 Event Count t Event Count t-1 Event Count t+1 hidden observed Poisson Rate λ(t) Normal Count t-1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1
Event t Event t-1 Event t+1 Event State Transition Matrix Observed Count t Observed Count t-1 Observed Count t+1 Event Count t Event Count t-1 Event Count t+1 hidden observed Poisson Rate λ(t) Normal Count t-1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1 β α η ηη
Event t Event t-1 Event t+1 Event State Transition Matrix Observed Count t Observed Count t-1 Observed Count t+1 Event Count t Event Count t-1 Event Count t+1 hidden observed Poisson Rate λ(t) Normal Count t-1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1 Markov Modulated Poisson Process (MMPP) model e.g., see Heffes and Lucantoni (1994), Scott (1998)
Approximate Inference
Gibbs Sampling * ** * * ** * ** * * * * * * * *
* x y ** * * ** *
Block Sampling
Gibbs Sampling Event t Event t-1 Event t+1 Event State Transition Matrix Observed Count t Observed Count t-1 Observed Count t+1 Event Count t Event Count t-1 Event Count t+1 Poisson Rate λ(t) Normal Count t-1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1
Gibbs Sampling Event t Event t-1 Event t+1 Event State Transition Matrix Observed Count t Observed Count t-1 Observed Count t+1 Event Count t Event Count t-1 Event Count t+1 Poisson Rate λ(t) Normal Count t-1 Day, Time t-1 Normal Count t-1 Day, Time t Normal Count t-1 Day, Time t+1 Poisson Rate λ(t) Poisson Rate λ(t) Event State Transition Matrix Event State Transition Matrix For the ternary valued event variable with chain length of 64,000 Brute force complexity ~
Gibbs Sampling Event t Event t-1 Event t+1 A AA Poisson Rate λ(t) Day, Time t-1 Observed Count t-1 Normal Count t-1 Event Count t-1 Poisson Rate λ(t) Day, Time t-1 Observed Count t-1 Normal Count t-1 Event Count t-1 Poisson Rate λ(t) Day, Time t-1 Observed Count t-1 Normal Count t-1 Event Count t-1
Chicken/Egg Delima car count
Event Popularity car count
Notion of Persistence missing from Baseline model Persistent Event
Detecting Real Events: Baseball Games Total Number Of Predicted Events Graphical Model Detection of the 76 known events Baseline Model Detection of the 76 known events %86.8% %81.6% %72.4% %60.5% Remember: the model training is completely unsupervised, no ground truth is given to the model
Multi-sensor Occupancy Model Modeling Count Data From Multiple Sensors: A Building Occupancy Model J. Hutchins, A. Ihler, and P. Smyth IEEE CAMSAP 2007,Computational Advances in Multi-Sensor Adaptive Processing, December 2007
Where are the People? Building LevelCity Level
Optical People Counter at a Building Entrance Loop Sensors on Southern California Freeways Sensor Measurements Reflect Dynamic Human Activity
Application: Multi-sensor Occupancy Model CalIt2 Building, UC Irvine campus
Building Occupancy, Raw Measurements Occ t = Occ t-1 + inCounts t-1,t – outCounts t-1,t
Building Occupancy: Raw Measurements Noisy sensors make raw measurements of little value Over-counting Under-counting
Adding Noise Model Event t Event t-1 Event State Transition Matrix Event Count t Event Count t-1 Poisson Rate λ(t) Normal Count t-1 Day, Time t-1 Normal Count t-1 Day, Time t Observed Count t Observed Count t-1 True Count t-1 True Count t
Probabilistic Occupancy Model In(Entrance) Sensors Out(Exit) Sensors Occupancy In(Entrance) Sensors Out(Exit) Sensors Constraint Time Occupancy Time tTime t+1
24 hour constraint 47 Constraint Occupancy Building Occupancy Geometric Distribution, p=0.5
Gibbs Sampling | Forward-Backward | Complexity Learning and Inference In(Entrance) Sensors Out(Exit) Sensors Occupancy In(Entrance) Sensors Out(Exit) Sensors Occupancy
Typical Days Thursday Friday Saturday Building Occupancy
Missing Data Building Occupancy time
Corrupted Data Building Occupancy Thursday Friday
Future Work Freeway Traffic On and Off ramps 2300 sensors 6 months of measurements
Sensor Failure Extension
Spatial Correlation
Four Off-Ramps
Publications Modeling Count Data From Multiple Sensors: A Building Occupancy Model J. Hutchins, A. Ihler, and P. Smyth IEEE CAMSAP 2007,Computational Advances in Multi-Sensor Adaptive Processing, December Learning to detect events with Markov-modulated Poisson processes A. Ihler, J. Hutchins, and P. Smyth ACM Transactions on Knowledge Discovery from Data, Dec 2007 Adaptive event detection with time-varying Poisson processes A. Ihler, J. Hutchins, and P. Smyth Proceedings of the 12th ACM SIGKDD Conference (KDD-06), August Prediction and ranking algorithms for event-based network data J. O Madadhain, J. Hutchins, P. Smyth ACM SIGKDD Explorations: Special Issue on Link Mining, 7(2), 23-30, December 2005