Modeling the evolution of temporal sequences is of great interest in many areas.
Dynamic Bayesian Networks (DBN).
Unrolling a BN in time. (Bold assumption)
Inadequate if the relationships change through time.
Alternative: Incorporating the temporal variation of network structure and parameter.
Problem
Data scarcity
\[\small{\mathbf{s}_t=[\mathbf{X}[t], G[t], \mathbf{\Theta}[t]]}\]
\[\small{\begin{aligned} & p\left(\mathbf{s}_t \mid \mathbf{s}_{t-1}\right)=p(G[t] \mid G[t-1]) p(\boldsymbol{\Theta}[t] \mid \boldsymbol{\Theta}[t-1], G[t]) \\ & \times p(\mathbf{X}[t] \mid \mathbf{X}[t-1], \boldsymbol{\Theta}[t], G[t]) \end{aligned}}\]
\[\small{p\left(\mathbf{s}_t \mid \mathbf{o}_{1: t}\right) \approx \sum_{i=1}^{N_s} w_t^i \delta\left(\mathbf{s}_t-\mathbf{s}_t^i\right)}\]
\[\small{\mathbf{s}_t^i \sim q\left(\mathbf{s}_t \mid \mathbf{s}_{t-1}^i, \mathbf{o}_t\right)}\]
\[w_t^i \propto w_{t-1}^i \frac{p\left(\mathbf{o}_t \mid \mathbf{s}_t^i\right) p\left(\mathbf{s}_t^i \mid \mathbf{s}_{t-1}^i\right)}{q\left(\mathbf{s}_t^i \mid \mathbf{s}_{t-1}^i, \mathbf{o}_t\right)}\]
Estimated network distribution at time 150 with random structure initialization at time 101. (a) Average posterior probability of true network structure G[150] = 15. (b) KL divergence for p(X[150] | .). Median, minimal, maximal trials are shown.
The time needed to converge to true network status from different initial network structures.
\[precision = \frac{tp}{tp+ fp}\]