Graphical Models with Continuous Variables
- Everything from discrete probabilities holds for continuous distributions
- However cannot use tables, have to use probability distribution functions
Gaussian Bayesian Networks
- All variables are continuous and modelled by Gaussian Densities
- Often good approximation for many real-world distributions
- Root nodes use univariate distributions
- Conditional Probability Distributions (CPDs) are commonly represented with linear Gaussian models
- Variance is independent of parents
- Matrix notation:
\[ P(X|u)=\mathcal{N}(\beta_0+\beta^\top u,\sigma^2_X)\]
- Learn parameters \(\mu,\sigma^2\) from data
- Once we have parameters, we can sample from distribution
- Multivariate Gaussian Distribution:
- Can identify independence assumptions directly from the Gaussian distribution parameters
- X_i, X_j are independent iff \(\Sigma_{i,j}=0\)
- \(X_i\perp X_j|X-{X_i,X_j}\) iff \(\Sigma^{-1}_{i,j}=0\)
Linear to Gaussian
- Linear Gaussian Network defines a joint multivariate Gaussian Dist:
- Y is linear Gaussian with parents \(X_1,...,X_k\)
- \(P(Y|x)=N(\beta_0+\beta^\top x,\sigma^2)\)
- \(X_1,..,X_k\) are jointly Gaussian with \(N(\mu,\Sigma\)
- Then Distribution \(Y\) is normal, \(P(Y)=N(\mu_Y,\sigma^2_Y)\)
- Using parents:
- \(\mu_y = \beta_0+\beta^\top\mu\)
- \(\sigma_y^2 = \sigma^2+\beta^\top\Sigma\beta\)
- Using parents:
- Joint dist over \(\{X,Y\}\) is normal with \(Cov(X_i,Y)=\Sigma^k_{k=1}B_k\Sigma_{i,j}\)
Gaussian to Linear
- A joint multivariate Gauss distribution defines a linear Gaussian network
- Given a set of variables \(\{X,Y\}\) in the form of a joint normal distribution
- \(p(Y|X)=N(\beta_0+\beta^\top X,\sigma^2)\)
- With parameters
Inference Operations
- Need to use canonical form to simplify inference operations over normal distribution
- Allows inference in closed form
- If factors have two different scopes, we extend scopes to make them match by adding 0 entries to \(K\) and \(h\)
Canonical Form
- Canonical form \(C(X;K,h,g)\) is defined as:
\[ C(X;K,h,g) = \exp\left(-\frac{1}{2}X^\top KX+h^\top X+g\right)\]
- With Variables:
Join
- Canonical join operation:
\[ C(K_1,h_1,g_1)*C(K_2,h_2,g_2)=C(K_1+K_2,h_1+h_2,g_1+g_2)\]
Marginalisation
- Consider marginalising \(Y\) for the canonical form, resulting in \(C'\):
Reduction
- Consider setting evidence \(Y=Y\), resulting in \(C'\):
Linear Model
- For linear Gaussian model \(Y\sim N(\beta_0+\beta^\top X,\sigma^2)\):
Variable Elimination
- Use canonical form operations and CPDs with evidence set.