Title: Relational Factor Graphs
1Relational Factor Graphs
- Lin Liao
- Joint work with Dieter Fox
2A Running Example
Collective classification of a persons
significant places
3Features to Consider
- Local features
- Temporal time of day, day of week, duration
- Geographic near restaurants, near stores
- Pair-wise features
- Transitions which place follows which place
- Global features
- Aggregates number of homes or workplaces
4Which Graphical Model?
- Option 1 Bayesian networks and Probabilistic
Relational Models - But the pair-wise relations may introduce cycles
Place 2
Place 1
Place 3
Place 4
5Which Graphical Model?
- Option 2 Markov networks and Relational Markov
Networks - But aggregations can introduce huge cliques and
lose independence relations.
Number of homes
Place 2
Place 1
Place 3
Place 4
6Motivation
- We want a relational probabilistic model that is
- Suitable to represent both undirected relations
(e.g., pair-wise features) and directed relations
(e.g., deterministic aggregation) - Able to address some of the computational issues
at the template level
7Outline
- Representation
- Factor graphs Kschischang et al. 2001, Frey
2003 - Relational factor graphs
- Inference
- Belief propagation
- Inference templates
- Summation template based on FFT
- Experiments
8Factor Graph
- Undirected factor graph Kschischang et al. 2001
- Bipartite graph that includes both variable nodes
(x1,,xN) and factor nodes (f1,,fM) - Joint distribution of variables is proportional
to the product of factor functions
x1
x3
f2
f1
f3
x4
x2
9Factor Graph
- Directed factor graph Frey 2003
- Allow some edges to be directed so as to unify
Bayesian networks and Markov networks - A valid graph should have no directed cycles
x1
x3
f2
f1
f3
x4
x2
10Markov Network to Factor Graph
Markov network
Factor graph
Factors represent the potential functions
11Bayesian Network to Factor Graph
Bayesian network
Factor graph
Factors represent the conditional probability
table
12Unify MN and BN
Aggregate features
Number of homes
Aggregation factor
Place labels
Local features
13Relational Factor Graph
- A set of factor templates that can be used to
instantiate (directed) factor graphs given data - Representation template
- Use SQL (similar to RMN)
- Guarantee no directed cycles
- Inference template
- Optimization within a factor (discussed later)
14Place Labeling Schema
15Place Labeling Transition Features
Pair-wise factor
Label1
Label2
Label3
16Place Labeling Aggregate Features
Aggregate feature
Num of homes
Bool variables
Home?
Home?
Home?
Label1
Label2
Label3
17Outline
- Representation
- Factor graphs Kschischang et al. 2001, Frey
2003 - Relational factor graphs
- Inference
- Belief propagation
- Inference templates
- Summation template based on FFT
- Experiments
18Inference in Factor Graph
- Belief propagation two types of messages
- Message from variable x to factor f
- Message from factor f to variable x
nx factors adjacent to x nf variables adjacent
to f
19Inference Templates
- Simplest case specify the function f(nf) and use
the above formula to compute message f -gt x - Problem complexity is exponential in the number
of factor arguments. This can be very expensive
for aggregation factors - Inference templates allow users to specify
optimized algorithms at the template level - Be in general form and easy to be shared
- Support template level complexity analysis
20Summation Templates
xout
..
xin1
xin2
xin7
xin8
21Summation Forward Message
- Compute the distribution of the sum of
independent variables xin1, . , xin8
xout
..
xin1
xin2
xin7
xin8
22Summation Forward Message
- Convolution tree each node can be computed using
FFT total complexity O(nlog2n)
23Summation Backward Message
- Message from xout defines a prior distribution of
the sum. For each value of xin2, compute the
distribution of sum and weighted by the prior
xout
..
xin1
xin2
xin7
xin8
24Summation Backward Message
- If we reuse the results cached for the forward
message, complexity becomes O(nlogn)
25Summation Templates
- By using convolution tree, FFT, and caching, the
average complexity of passing a message through
summation factor is O(nlogn), instead of
exponential.
26Learning
- Estimate the weights for probabilistic factors
(local features, pair-wise features, and
aggregate features) - Optimize the weights to maximize the conditional
likelihood of the labeled training data - The same algorithm as RMN
27Experiments
- Two data sets
- Single data set one persons GPS data for 4
months - Multiple data set one-week GPS data from 5
subjects - Six candidate labels Home, Work, Shopping,
Dining, Friend, Others - Get the geographic knowledge from Microsoft
MapPoint Web Service
28How Much Aggregates Help
- Test on multiple data set leave-one-subject-cro
ssvalidation - Test on single data set crossvalidation (train
on 1 month, test on 3 months)
Error rate Multiple Single
No aggregate 28 9
With aggregate 18 6
29How Efficient the Optimized BP
30Summary
- Relational factor graph is
- SQL (directed) factor graph
- It is
- Suitable to represent both undirected relations
and directed relations - Convenient to use no directed cycles
- Able to address computation issues at the
template level