Statistical inference · Embeddings & representations / Open question
Revealing Hidden Interactions
A learned interaction graph can confuse direct dependence with unobserved causes. Hidden-spin models make this ambiguity mathematically concrete.
Seed catalog · Showing the bundled record.
Problem statement
Consider synchronous kinetic Ising networks on n visible and k hidden spins, whose next spins are conditionally independent with logistic fields h_i plus the sum of J_ij x_j, maximum in-degree d and bounded couplings/fields. Observe all visible spins through known independent bit-flip noise; hidden spins are never observed. Characterize identifiable coupling equivalence classes from visible trajectories, up to hidden-spin relabeling/sign symmetries, and the extra identifiability gained by interventions fixing visible spins. Give finite-trajectory bounds with k, d and intervention costs explicit.
Assumptions and scope
- Fix the latent-state count, graph/emission families, observation process and allowed symmetries, including hidden-spin relabeling or sign changes where applicable.
- Require substantive observability and nondegeneracy conditions; arbitrary hidden-network recovery from a visible marginal is not identifiable.
- State whether interventions or multiple environments are available and charge them. Variational fit or prediction of visible data is not equivalent to recovery of the underlying couplings.
What would settle it?
A necessary/sufficient identification criterion or a sharply separated restricted class, with a consistent finite-data estimator and explicit observational-equivalence examples.
Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.
Status and evidence
Fully observed Ising structure/dynamics already admit algorithms. Latent kinetic Ising models with point-process emissions and variational EM already exist, including empirical recovery studies. The residual target is an observation-aware identifiability and finite-sample guarantee for the specified latent family. This is an editorial, source-motivated research target. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.
Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.
Sources and further reading
- Ghio and Saad — Latent Kinetic Ising Models of Neural Spike Trains ↗related algorithm · source checked 2026-11-19
Primary abstract checked: asymmetric kinetic Ising latent states with point-process emissions, variational EM and synthetic/misspecified-data recovery experiments. No finite-sample global identifiability theorem was checked.
- Gaitonde, Moitra and Mossel — Learning Ising Models from Evolutions ↗partial theory · source checked 2026-11-19
Publisher abstract checked: efficient recovery from actual continuous-time state evolution without observing failed update attempts, for a broad reversible single-site dynamics class. Temporal dependence alone is not an unresolved barrier.
- Bresler — Efficiently Learning Ising Models on Arbitrary Graphs ↗partial theory · source checked 2026-11-19
Primary abstract checked: efficient bounded-degree graph recovery from independent samples, including low-temperature and nonuniform cases under parameter-identifiability assumptions. Slow physical sampling is separate from learning with independent samples.