Markov_network

Markov network

A Markov network, or Markov random field, is a model of the (full) joint probability distribution of a set $\mathcal{X}$ of random variables having the Markov property. A Markov network is similar to a Bayesian network in its representation of dependencies. It can represent certain dependencies that a Bayesian network cannot (such as cyclic dependencies); on the other hand, it can't represent certain dependencies that a Bayesian network can (such as induced dependencies). The prototypical Markov network is the Ising model; indeed, the Markov network was introduced as the general setting for the Ising model^[1].

1 Formal Definition
- 1.1 Joint Distribution Function
- 1.2 Markov property
2 Inference
3 Conditional Random Fields
4 See also
5 References

Formal Definition

Formally, a Markov network consists of:

an undirected graph G = (V,E), where each vertex v ∈V represents a random variable in $\mathcal{X}$ and each edge {u,v} ∈ E represents a dependency between the random variables u and v,
a set of functions $f k$ (also called factors or clique factors and sometimes features), where each $f k$ has the domain of some clique (or subclique) k in G. Each $f k$ is a mapping from possible joint assignments (to the elements of k) to non-negative real numbers.

Joint Distribution Function

The joint distribution (or Gibbs measure) represented by a Markov network is given by:

$P(X=x) = \frac{1}{Z} \prod_{k} f_k (x_{ \{ k \}})$

where $x {k}$ is the state of the random variables in the kth clique, and the product runs over all the cliques in the graph. Here, $Z$ is the partition function, so that

$Z = \sum_{x \isin \mathcal{X}} \prod_{k} f_k(x_{ \{ k \} })$ .

In practice, a Markov network is often conveniently expressed as a log-linear model, given by

$f_k=\exp \left(w_k \phi_k (x_{ \{ k \}}) \right)$

so that

$P(X=x) = \frac{1}{Z} \exp \left( \sum_{k} w_k \phi_k (x_{ \{ k \}}) \right)$

with partition function

$Z = \sum_{x \isin \mathcal{X}} \exp \left(\sum_{k} w_k\phi_k(x_{ \{ k \} })\right)$ .

In this context, the $w k$ 's are weights and the $φ k$ 's are potential functions from the clique $k$ to the reals. These functions are also sometimes called Gibbs potentials; the term potential comes from physics, where these can be literally interpreted as the potential energy between nearest neighbors.

Log-linear models are especially convenient for their interpretation. A log-linear model can provide a much more compact representation for many distributions, especially when variables have large domains. They are convenient too because their negative log likelihoods are convex. Unfortunately, though the likelihood of a log-linear Markov network is convex, evaluating the likelihood or gradient of the likelihood of a model requires inference in the model, which is in general computationally infeasible.

Markov property

Markov networks have the Markov property: the probability that a vertex u of the graph is in the state $x u$ depends only on the nearest neighbors of the vertex u; the vertex is conditionally independent of all of the other verticies in the graph. One writes this as

$P(X_u=x_u|X_v, v\ne u) = P(X_u=x_u|X_v, v\isin N_u)$

The set $N u$ of the nearest neighbors of u is sometimes referred to as the Markov blanket of u.

Conversely, one may define a Markov network to be a collection of random variables having the Markov property; in this case, it may be proven that such a network can always be written in terms of the Gibbs measure, as shown above. The potential thus obtained is unique, up to an additive constant; this potential is called the canonical potential. Precisely, there is at least one state of the variables X which has maximum probability; this state is referred to as the ground state. The canonical potential is defined so that the potential of the ground state is zero.

Inference

As in a Bayesian network, one may calculate the conditional distribution of a set of nodes $V' = {v 1,..., v i}$ given values to another set of nodes $W' = {w 1,..., w j}$ in the Markov network by summing over all possible assignments to $u \notin V',W'$ ; this is called exact inference. However, exact inference is a #P-complete problem, and thus computationally intractable in the general case. Approximation techniques such as Markov chain Monte Carlo and loopy belief propagation are often more feasible in practice. Some particular subclasses of MRFs, such as trees, have polynomial-time inference algorithms; discovering such subclasses is an active research topic. There are also subclasses of MRFs which permit efficient MAP, or most likely assignment, inference; examples of these include associative networks.

Conditional Random Fields

One notable variant of a Markov network is a conditional random field, in which each random variable may also be conditioned upon a set of global observations $o$ . In this model, each function $φ k$ is a mapping from all assignments to both the clique k and the observations $o$ to the nonnegative real numbers. This form of the Markov network may be more appropriate for producing discriminative classifiers, which do not model the distribution over the observations.

References

^ Ross Kindermann and J. Laurie Snell, Markov Random Fields and Their Applications (1980) American Mathematical Society, ISBN 0-8218-5001-6

Contents