Factor graphs

1 Classic factor graphs

A long history of multiple invention, probably. A recognisable modern form appears in (Frey et al. 1997; Frey 2003; Kschischang, Frey, and Loeliger 2001). This classic version is explained lots of places, but for various reasons I needed the FFG version first, so my explanation of that is longer.

The representation that we need to “make all the inference steps easy” ends up being a bipartite graph with two types of nodes: variables and factors. Given a factorization of a function $g(X_1,X_2,\dots ,X_n)$, $$g(X_1,X_2,\dots ,X_n)=\prod _{j=1}^m{f}_j\left(S_j\right)$$ where $S_j\subseteq \{X_1,X_2,\dots ,X_n\}$, the corresponding factor graph $G=(X,F,E)$ consists of variable vertices $X=\{X_1,X_2,\dots ,X_n\}$, factor vertices $F=\{f_1,f_2,\dots ,f_m\}$, and edges $E$. The edges depend on the factorization as follows: there is an undirected edge between factor vertex $f_j$ and variable vertex $X_k$ if $X_k\in S_j$.

Quibble: We often assume the function to be factored is a probability density function; more generally we wish to factorise measures.

Once the factor graph is written out, computations on the graph are essentially the same for every node. By contrast, in a classic I-graph DAG there are several different rules to remember for leaf nodes and branches, for colliders and forks etc.

I may return to classic factor graphs.

For now, a pretty good introduction is included in Ortiz, Evans and Davison’s tutorial on Gaussian belief propagation, or Minka’s Minka (2005).

2 Forney-style factor graphs (FFGs)

A tweaked formalism. Citations tell me this was introduced in a recondite article (Forney 2001) which I did not remotely understand because it was way too far into coding theory for me. It is explained and exploited much better for someone of my background in subsequent articles and used in several computational toolkits (Korl 2005; Cox, van de Laar, and de Vries 2019; H.-A. Loeliger et al. 2007; H.-A. Loeliger 2004; van de Laar et al. 2018; Akbayrak, Bocharov, and de Vries 2021).

Relative to classic factor graphs, H.-A. Loeliger (2004) advocates for FFGs for the following advantages

• suited for hierarchical modeling (“boxes within boxes”)
• compatible with standard block diagrams
• simplest formulation of the summary-product message update rule
• natural setting for Forney’s results on Fourier transforms and duality.

Mao and Kschischang (2005) argues:

Forney graphs possess a strikingly elegant duality property: by a local dualization operation, a Forney graph for a linear code may be transformed into another graph, called the dual Forney graph, which represents the dual code.

The explanation in Cox, van de Laar, and de Vries (2019) gives the flavour of how Forney-style graphs work:

A Forney-style factor graph (FFG) offers a graphical representation of a factorized probabilistic model. In an FFG, edges represent variables and nodes specify relations between variables. As a simple example, consider a generative model (joint probability distribution) over variables $x_1,\dots ,x_5$ that factors as $$f(x_1,\dots ,x_5)=f_a\left(x_1\right)f_b(x_1,x_2)f_c(x_2,x_3,x_4)f_d(x_4,x_5),$$ where $f_{\bullet }(\cdot )$ denotes a probability density function. This factorized model can be represented graphically as an FFG, as shown in Fig. 1. Note that although an FFG is principally an undirected graph, in the case of generative models we specify a direction for the edges to indicate the “generative direction”. The edge direction simply anchors the direction of messages flowing on the graph (we speak of forward and backward messages that flow with or against the edge direction, respectively). In other words, the edge directionality is purely a notational issue and has no computational consequences.…

The FFG representation of a probabilistic model helps to automate probabilistic inference tasks. As an example, consider we observe $x_5={\hat{x}}_5$ and are interested in calculating the marginal posterior probability distribution of $x_2$ given this observation.

In the FFG context, observing the realization of a variable leads to the introduction of an extra factor in the model which “clamps” the variable to its observed value. In our example where $x_5$ is observed at value ${\hat{x}}_5$, we extend the generative model to $f(x_1,\dots ,x_5)\cdot \delta (x_5-{\hat{x}}_5).$ Following the notation introduced in Reller (2013), we denote such “clamping” factors in the FFG by solid black nodes. The FFG of the extended model is illustrated in Fig. 2…

Another place clamping arises is that, since a variable can only appear in two factors, if we want a variable to appear in more than two, we add extra factors in, each of which constrains the variables touching it to be equal to one another. This seems weird; but it kinda-sorta corresponds to be what we would do in a classic factor graph anyway, in the sense that we would add an extra message passing step in for each extra factor, so in a sense this weird contrivance could be simply described as writing out our plan in full.

Computing the marginal posterior distribution of $x_2$ under the observation $x_5={\hat{x}}_5$ involves integrating the extended model over all variables except $x_2,$ and renormalising: $$f(x_2\mid x_5={\hat{x}}_5)\propto \int \dots \int f(x_1,\dots ,x_5)\cdot \delta (x_5-{\hat{x}}_5)\mathrm{d}{x}_1\mathrm{d}{x}_3\mathrm{d}{x}_4\mathrm{d}{x}_5$$ $$=\stackrel{2}{\overbrace{\int \underset{1}{\underbrace{f_a\left(x_1\right)}}{f}_b(x_1,x_2)\mathrm{d}{x}_1}}\stackrel{(4)}{\overbrace{\iint f_c(x_2,x_3,x_4)\underset{3}{\underbrace{(\int f_d(x_4,x_5)\cdot \delta (x_5-{\hat{x}}_5)\mathrm{d}{x}_5)}}\mathrm{d}{x}_3\mathrm{d}{x}_4}}.$$ The nested integrals result from substituting the [original] factorization and rearranging the integrals according to the distributive law. Rearranging large integrals of this type as a product of nested sub-integrals can be automated by exploiting the FFG representation of the corresponding model. The sub-integrals indicated by circled numbers correspond to integrals over parts of the model (indicated by dashed boxes in Fig. 2), and their solutions can be interpreted as messages flowing on the FFG. Therefore, this procedure is known as message passing (or summary propagation). The messages are ordered (“scheduled”) in such a way that there are only backward dependencies, i.e., each message can be calculated from preceding messages in the schedule. Crucially, these schedules can be generated automatically, for example by performing a depth-first search on the FFG.

What local means in the context of graphical models was not intuitive to me at first. Local means that: in this high dimensional integral, only some dimensions are involved in each sub-step; the graph tells us which integrals can be moved outside the calculation by showing how it factorises. This is nothing to do with locality over the domain of the functions. This needs a graphical illustration.

3 Plated

Obermeyer et al. (2019):

A plated factor graph is a labelled bipartite graph $(V,F,E,P)$ whose vertices are labelled by the plates on which they are replicated $P:V\cup F\to \mathcal{P}(B)$, where $B$ is a set of plates. We require that each factor is in each of the plates of its variables: $\mathrm{\forall }(v,f)\in E,P(v)\subseteq P(f).$

What now?

The key operation for plated factor graphs will be “unrolling” to standard factor graphs. First define the following plate notation for either a factor or variable $z$ : ${\mathcal{M}}_z(b)=\{1,\dots ,M(b)\}$ if $b\in P(z)$ and $\{1\}$ otherwise. This is the set of indices that index into the replicated variable or factor. Now define a function to unroll a plate, $$(V^{\prime },F^{\prime },E^{\prime },P^{\prime })=\mathrm{unroll}((V,F,E,P),M,b)$$ where $v_i$ indicates an unrolled index of $v$ and, $$\begin{array}{rl}{V}^{\prime }& =\{v_i\mid v\in V,i\in {\mathcal{M}}_v(b)\}\\ F^{\prime }& =\{f_i\mid f\in F,i\in {\mathcal{M}}_f(b)\}\\ E^{\prime }& =\{(v_i,f_j)\mid (v,f)\in E,i\in {\mathcal{M}}_v(b),j\in {\mathcal{M}}_f(b)\\ & (i=j)\vee b\notin (P(v)\cap P(f))\}\\ P^{\prime }(z)& =P(z)\mathrm{\setminus }\{b\}\end{array}$$

An illustration might make that clearer:

Consider a plated factor graph with two variables $X,Y,$ three factors $F,G,H,$ and two nested plates Figure 3. Assuming sizes $I=2$ and $J=3,$ this plated factor graph unrolls to a factor graph, Figure 4.

Or:

Consider a plated factor graph with two variables, one factor, and two overlapping non-nested plates, denoting a Restricted Boltzmann Machine (RBM) Figure 5. Assuming sizes $I=2$ and $J=2$, this plated factor graph unrolls to a factor graph Figure 5.

I think that gives the vibe.

4 Generic messages

Question: Introductory texts discuss sum-product message passing, which is essentially about solving integrals. It might seem like I want to pass some other kind of update, e.g. maximum likelihood? Does this factor graph still help us? Yes, we can calculate max-product messages then. What else can we calculate locally? I am not sure how wide the class of things for which this holds true is. Something about which algebras are defined?

5 Fourier transforms in

What does it benefit us to take a Fourier transform on a graph? What does that even mean? (Kschischang, Frey, and Loeliger 2001; Forney 2001; Mao and Kschischang 2005)

6 In Bayesian brain models

See de Vries and Friston (2017) and van de Laar et al. (2018) for a connection to predictive coding.

7 Causal inference in

How does do-calculus work in factor graphs?

8 Tooling

9 References