Bayesian Network Data Fusion Visualization

1
Bayesian Network Data Fusion Visualization ISDS
2007 Conference October 12, 2007 Charles
Hodanics charles.hodanics_at_jhuapl.edu
2
Objectives

• Use Bayesian Networks for detection
• Fuse data from multiple networks
• Visualize Bayesian Network outputs
• Create a configurable visualization tool
• Better assist public health users

3
Background

• What is a Bayesian Network?
• A probabilistic graphical model
• That represents dependencies and relationships
• Where the network structure and conditional
  probabilities capture an experts view of a
  system
• They have been applied to the public health
  domain for research purposes
• They have not been used directly by the end users
  of public health systems

4
Example Network
Expert established relationship and probability
model
Figure from PHIN 2007 Conference - Zaruhi
Mnatsakanyan
5
Issue

• Bayesian Network technology is becoming more
  accepted in public health
• Visualization becomes a critical component of the
  overall system design
• The tools developed
• Utilize computer-assisted analysis on Bayesian
  Networks in the public health domain
• Provide a concise view of the data for better
  decision support
• Shorten the decision making phase allowing rapid
  dissemination of information to public health
• Can be applied to other locations and other
  domains

6
Result DFV

• The Data Fusion Visualization (DFV)
• Provides an intuitive graphical interface
• Represents Bayesian Network Data Fusion
• Supports users in three ways
• Provides a drill-down interpretation
• Provides an intuitive representation of a
  Bayesian Network
• Abstracts the visualization from the underlying
  model making it capable of masking
  inter-operating Bayesian Networks

7
Method Probability

• The DFV provides several screens
• Each with clear charts and tools for user
  interaction
• The first screen includes a time series graph
• Shows the probability of a public health event in
  a selected jurisdiction for each day
• Is annotated with a color scheme highlighting
  potential events
• A potential event is the Bayesian Network
  detection of epidemiological significance

8
Probability Time Series
The DFV allows the user to further investigate a
detected event.
The user can zoom and double click resulting in
the second screen.
Probability
Days
9
Method Hierarchical

• The second screen presents a node-based graph
• This visualization shows a hierarchical
  drill-down with alert annotations
• Each node has the potential to be investigated
  further
• Each element is specified in an XML file that
  defines the node, the datasource, and an entry
  point for further investigation

10
Example XML Config

• ltdisplaygt
• ltregion name"NCR id"NCR"
• bbnNode"influenza"
  nodeState"true" path"c\"gt
• ltstate name"Maryland" id"MD
• path"c\" bbnNode"influenza
  nodeState"true"gt
• ltcounty name"Montgomery"
  id"MONTGOMERY"
• bbnNode"influenza"
  nodeState"true newScreen"true
• templateFile"reference.xml
  path"c\" /gt
• lt/stategt
• lt/regiongt
• lt/displaygt

The XML can be configured to meet any domains
needs.
11
Node-Based Graph
Example screen, higher level nodes are not
implemented.
The leaf nodes in the graph are entry nodes into
deeper data visualizations as specified in the
xml configuration.
12
Method Decision Logic

• County-level node graphs visualize decision
  support logic implemented in the Bayesian Network
• County-level node graphs can also be grouped by
  different criteria as the domain sees fit

13
County Node Graph
At the county level, the user can drill down to
actual data records that are contributing to the
event detection.
Intermediate nodes represent the decision making
process implemented in the Bayesian Network.
This can be configured to any design in xml.
Nodes provide visual summaries of more specific
details about the data distribution.
14
Method Specifics

• The set of data can now be researched and further
  narrowed based on specific dates of inquiry

Date ranges can be narrowed or expanded. This
lets the user see records that led to the event.
Sensitive Data Removed
Sensitive Data Removed
15
Conclusion

• With the various graphs in this visualization
• A user is capable of walking through an event of
  interest
• Starting at a higher level perspective
• Down to actual data occurrences
• Does not require Bayesian Network experience
• Users can accept or decline the systems decision
  in a timely manner
• High-level alert walkthroughs to low-level
  details will be very useful in research

16
Future Steps

• Deployment to public health areas
• The capability to switch underlying Bayesian
  Networks
• Apply different data fusion techniques
• Switch algorithms for event notification
• Introduce simulated data and observe the higher
  level reactions

17
Acknowledgements

• JHU Applied Physics Lab Colleagues
• Raj Ashar, MS.
• Zaruhi Mnatsakanyan, PhD.
• Open Source Community
• JUNG
• JFreeChart