Discrete score matching improves structure discovery from count data
Discrete Score Matching Enables Causal Discovery from Count Data
Machine Learning
Summary
Understanding cause-and-effect relationships in data that count things can be tricky because standard methods rely on smooth changes in data, which counts often lack. The authors develop a new way to measure relationships by looking at scores that capture how connections curve or change, even when the data are counts. They introduce a method called DISCO that uses these new scores to accurately find cause-and-effect structures in data with counts, like how one event influences another. Their approach works well across different types of count data and can handle large networks efficiently.
What this means in practice
- •For data engineers: Recover causal structures from count-based event logs or transaction data using the new CCS and OCS criteria.
- •For bioinformatics teams: Analyze gene expression or cell count data to infer causal regulatory networks with the discrete score matching approach.
Authors
Euijong Song, Hyewon Park, Gunwoong Park
Abstract
Count data pose a challenge for score-matching-based causal discovery: derivatives are unavailable, and simply replacing them with finite differences does not generally suffice for causal discovery. We generalize SCORE's constant-curvature criterion (Rolland et al., 2022) by conditioning on the node's value, yielding the conditional curvature score (CCS) for ordering. We also extend curvature-based parent recovery through the off-diagonal curvature score (OCS), enabling directed acyclic graph (DAG) recovery with both scores constructed from score functions for continuous data and concrete scores for counts. In the bivariate setting, zero CCS exactly characterizes a semiparametric generalized linear model (GLM) conditional form in which the conditional family need not be specified in advance, unlike in classical GLMs. For bivariate semiparametric GLM DAGs under our regularity condition, canonical-parameter nonlinearity is necessary and sufficient for identifiability. In multivariate DAGs, this nonlinearity enables DAG recovery through CCS and OCS. Our framework identifies a new class of semiparametric GLM DAGs that strictly contains the nonlinear Gaussian ANM class identified by SCORE. We introduce DISCO (DIscrete SCOre), a count-DAG recovery algorithm that estimates CCS and OCS using discrete diffusion. Experiments demonstrate accurate DAG recovery across Poisson, negative binomial, binomial, and mixed-family settings, as well as scalability to 1,000-node DAGs on a single GPU.