Eliciting Bayesian network structure from experts
PhD Mid-candidature Review
August 10, 2012
Student: Peter Serwylo
Supervisors: Grace Rumantir and Frada Burstein
Some background
on the thesis...
The problem
Each mass gathering requires a
Risk assessment
Generate patient scenarios
Where each scenario is
an attendee requiring first aid
Simple (roulette wheel) scenario generation
Better scenario generation...
Instead of:
Pr( Asthma )
Condition on the characteristics of the event:
Pr( Asthma | Seated, Indoors, Alcohol, etc... )
And ignore those things which don't matter:
Pr( Asthma | Indoors )
As it happens...
Bayesian Networks
are pretty good for this
Constructing a Bayesian Network
Using data and algorithms:
Analyse the data sets for relationships.
Using experts and their knowledge:
Elicit knowledge from experts about relationships.
This part of the project presents
A technique for expert elicitation
of Bayesian Networks
This technique
is to
existing techniques
as
Surveys
are to
interviews
The analogy (in sweeping generalisations)
- Surveys are rigid
- Surveys require more questions
- Interviews are flexible
- Interviews require less questions
- Surveys require less coordination
- Surveys require less commitment from the participant
- Surveys don't require transcribing
(if done online)
- Interviews require more coordination
- Interviews require more commitment from the participant
- Interviews require transcribing
- Surveys reach
more people
- Interviews reach
less people
How can the
Survey analogy be used to
elicit Bayesian Networks?
Consider this way of conceptualising a Bayesian Network:
A (sparse) matrix of relationships
Should we ask the expert to fill in this matrix?
We can
Constrain the possibilities
Variables cannot influence themselves...
...and injuries cannot influence event characteristics...
...and we assume each patient only has one injury...
...and we usually know the event characteristics...
All of a sudden it is not so sparse
Specifying the presence of relationships via the matrix
doesn't allow experts to:
- Specify their confidence in the existence of the relationship
- Justify the relationship to others
- Review other experts opinions about the relationship
- Add new variables which were not originally specified
So we replace the matrix with
More understandable questions
Delphi-fying
verb
A word I just made up for this talk...
When is a Delphi round complete?
All participants have finished.
or when
A certain time has elapsed.
After each round,
Remove unused variables
and
Provide feedback,
then do it again
Removing variables which nobody selected
Provide feedback by
Identifying the odd ones out
Provide them with results from other participants.
Ask them to clarify or correct their stance.
When is the whole thing complete?
The parents of each variable cease to change.
or when
A certain number of rounds has completed.