# Collider-Aware DAG Variational Network

- ID: 654
- Canonical URL: https://synthcore.org/idea/654/collider-aware-dag-variational-network
- API JSON: https://synthcore.org/api/idea/654.json
- API Markdown: https://synthcore.org/api/idea/654.md
- Verification status: unverified
- Source: [arXiv:2607.09645](https://arxiv.org/abs/2607.09645)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: graph-nn, vae, world-model, training
- Math tags: probability, graph-theory, stochastic-processes
- Ratings: usefulness 6/10; difficulty 5/10; novelty 5/10

## Idea description

Replace independent uncertainty heads in a branching neural network with a structured variational posterior whose non-root node distributions condition on jointly sampled latent states of all parents. This allows collider evidence to explain away upstream uncertainty: evidence at a child can alter the posterior over several parent branches instead of leaving their uncertainties artificially independent. The approach can be implemented as a stochastic DAG network and trained with an evidence lower bound.

## Mathematical statement

The paper defines a DAG-DGP by assigning an independent function module to each non-root node and evaluating it recursively: $F_{w}^{(i)}=f_{w}(\{F_{p}^{(i)}\}_{p\in\mathrm{Pa}(w)})$, where $w$ is a node, $i$ indexes a data case, $\mathrm{Pa}(w)$ is the set of parents, $F_w^{(i)}$ is the node latent, and $f_w$ is the node function. The paper states that the joint prior factorizes into one Gaussian conditional per non-root node given its parents. Adapt this factorization to a neural variational model: $q_\phi(\mathbf z\mid\mathbf x)=\prod_{r\in\mathcal R}q_{\phi,r}(z_r\mid x)\prod_{w\in\mathcal U}q_{\phi,w}(z_w\mid\{z_p:p\in\mathrm{Pa}(w)\},x)$, where $\mathcal R$ is the root set, $\mathcal U$ is the non-root set, and each conditional is a diagonal Gaussian whose parameters are output by a neural node module. Train with $\mathcal L(\theta,\phi)=\mathbb E_{q_\phi}[\log p_\theta(y\mid z_{\mathrm{out}},x)]-\sum_{r\in\mathcal R}\mathrm{KL}(q_{\phi,r}\|p_r)-\sum_{w\in\mathcal U}\mathbb E_{q_\phi(z_{\mathrm{Pa}(w)})}[\mathrm{KL}(q_{\phi,w}(\cdot\mid z_{\mathrm{Pa}(w)},x)\|p_{\theta,w}(\cdot\mid z_{\mathrm{Pa}(w)}))]$. The original recursion supplies the computation graph, while the adapted conditional factorization supplies dependence-preserving uncertainty propagation.

## Key formulas

- $$F_{w}^{(i)}=f_{w}\!\bigl(\{F_{p}^{(i)}\}_{p\in\mathrm{Pa}(w)}\bigr),\qquad w\in\mathcal{U},\;i\in[n].$$
- $$q_\phi(\mathbf z\mid\mathbf x)=\prod_{r\in\mathcal R}q_{\phi,r}(z_r\mid x)\prod_{w\in\mathcal U}q_{\phi,w}\!\left(z_w\mid\{z_p:p\in\mathrm{Pa}(w)\},x\right).$$
- $$q_{\phi,w}(z_w\mid z_{\mathrm{Pa}(w)},x)=\mathcal N\!\left(z_w;\mu_{\phi,w}(z_{\mathrm{Pa}(w)},x),\operatorname{diag}\exp(2s_{\phi,w}(z_{\mathrm{Pa}(w)},x))\right).$$
- $$\mathcal L(\theta,\phi)=\mathbb E_{q_\phi}\![\log p_\theta(y\mid z_{\mathrm{out}},x)]-\sum_{r\in\mathcal R}\mathrm{KL}(q_{\phi,r}\|p_r)-\sum_{w\in\mathcal U}\mathbb E_{q_\phi(z_{\mathrm{Pa}(w)})}\!\left[\mathrm{KL}\left(q_{\phi,w}(\cdot\mid z_{\mathrm{Pa}(w)},x)\|p_{\theta,w}(\cdot\mid z_{\mathrm{Pa}(w)})\right)\right].$$

## Implementation notes

(1) Integrate this into a branching MLP, graph neural network, or multi-fidelity surrogate. Represent every latent node $w$ by a tensor $z_w$; for a node with multiple parents, concatenate projected parent tensors before computing its posterior mean and log standard deviation. Keep the posterior parameters conditional on sampled parent values rather than creating one unconditional Gaussian per node. Use a decoder or prediction head from one or more output nodes. (2) Pseudocode: sample root latents $z_r=\mu_r(x)+\exp(s_r(x))\odot\epsilon_r$ for every $r\in\mathcal R$, with $\epsilon_r\sim\mathcal N(0,I)$; traverse nodes in topological order; compute $h_w=\mathrm{MLP}_w([x,\{z_p:p\in\mathrm{Pa}(w)\}])$; set $(\mu_w,s_w)=\mathrm{split}(\mathrm{MLP}^{post}_w(h_w))$; sample $z_w=\mu_w+\exp(s_w)\odot\epsilon_w$; evaluate the likelihood and the diagonal-Gaussian KL terms in the ELBO above. (3) The DAG recursion and factorization are fixed by the paper's mathematics. Estimate neural parameters, prior scales, and KL weights empirically; use one Monte Carlo sample during training and 20--100 samples for uncertainty evaluation. To test explaining away, add an observed child likelihood and measure whether posterior covariance between its two parents changes relative to the prior. (4) First experiment: compare this model against an otherwise identical branching variational autoencoder with independent node posteriors on a synthetic collider $z_1,z_2\to y$, followed by a small multi-fidelity regression task. Use identical parameter counts, optimizer settings, and random seeds. Success means lower negative log-likelihood and better calibrated 90% predictive intervals, with posterior parent correlation changing appropriately when the collider observation is revealed. Also monitor ELBO variance and gradient norm; the structured model should avoid overconfident independent-branch predictions without a large training-time increase.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.