A novel hybrid approach for positive-valued DAG learning
| Authors | Yao Zhao |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2604.08935 |
| Download | |
| Categories | stat.ML, cs.LG |
Abstract
Causal discovery from observational data remains a fundamental challenge in machine learning and statistics, particularly when variables represent inherently positive quantities such as gene expression levels, asset prices, company revenues, or population counts, which often follow multiplicative rather than additive dynamics. We propose the Hybrid Moment-Ratio Scoring (H-MRS) algorithm, a novel method for learning directed acyclic graphs (DAGs) from positive-valued data by combining moment-based scoring with log-scale regression. The key idea is that for positive-valued variables, the moment ratio \frac{\mathbb{E}[X_j^2]}{\mathbb{E}[(\mathbb{E}[X_j \mid S])^2]} provides an effective criterion for causal ordering, where denotes candidate parent sets. H-MRS integrates log-scale Ridge regression for moment-ratio estimation with a greedy ordering procedure based on raw-scale moment ratios, followed by Elastic Net-based parent selection to recover the final DAG structure. Experiments on synthetic log-linear data demonstrate competitive precision and recall. The proposed method is computationally efficient and naturally respects positivity constraints, making it suitable for applications in genomics and economics. These results suggest that combining log-scale modeling with raw-scale moment ratios provides a practical framework for causal discovery in positive-valued domains.
Engineering Breakdown
Plain English
This paper addresses causal discovery from observational data where variables are inherently positive-valued (like gene expression, stock prices, or revenue), which typically follow multiplicative rather than additive dynamics. The authors introduce the Hybrid Moment-Ratio Scoring (H-MRS) algorithm that combines moment-based scoring with log-scale regression to learn directed acyclic graphs (DAGs) from such data. The key insight is that the moment ratio E[X_j²]/E[(E[X_j|S])²] provides an effective criterion for determining causal ordering in positive-valued systems. This approach is specifically designed to handle the unique statistical properties of positive-valued variables, which existing DAG learning methods struggle with because they assume linear or additive relationships.
Core Technical Contribution
The core novelty is recognizing that positive-valued variables require fundamentally different scoring mechanisms for causal discovery than standard additive models. Rather than using traditional regression-based scores (like BIC or GES), H-MRS leverages moment ratios that naturally capture the multiplicative structure inherent in positive-valued data. The hybrid approach combines moment-based scoring—which provides theoretical guarantees for positive distributions—with log-scale regression to exploit the log-linear relationships that emerge when you transform positive variables. This is a departure from conventional DAG learning literature, which has primarily focused on linear Gaussian assumptions or additive noise models that are inappropriate for inherently positive quantities.
How It Works
The algorithm operates on positive-valued observational data by first computing empirical moment ratios for candidate parent sets S. For each variable X_j and potential parent configuration S, it calculates the moment ratio E[X_j²]/E[(E[X_j|S])²], which measures the reduction in variance (on the original scale) when conditioning on S. Simultaneously, it performs log-scale regression by transforming variables as log(X_j) and fitting linear models, since multiplicative relationships become additive under log transformation. The scoring function combines evidence from both the moment-ratio channel and log-regression channel, using a weighting scheme that balances their contributions. The algorithm then searches over possible DAG structures using standard constraint-based or score-based search procedures (like greedy equivalence search), but with the novel H-MRS score replacing traditional scoring functions. Finally, it returns an estimated causal DAG that explains the observed positive-valued data, along with a ranking of causal relationships by strength.
Production Impact
For teams building causal inference systems on real-world positive-valued data (genomics, finance, epidemiology), this approach eliminates the need for ad-hoc variable transformations or approximations that violate model assumptions. In a production pipeline, you would replace your existing DAG learning module with H-MRS when your domain variables are strictly positive and you suspect multiplicative dynamics—this is particularly valuable in biotech (gene expression prediction), fintech (asset price modeling), and economics (production functions). The method requires computing moment statistics and log-regression fits, which scales linearly with the number of variables and quadratically with the search over parent sets, so computational cost is reasonable even for moderate-sized networks (10-100 variables). The main trade-off is that H-MRS assumes variables are positive and follow particular distributional properties; if your data violates these assumptions or contains zeros/negative values, you'll need preprocessing or alternative methods, which adds implementation complexity.
Limitations and When Not to Use This
The approach fundamentally requires strictly positive-valued data; any presence of zeros or negative values breaks the moment-ratio calculation and log-transformation. The paper appears incomplete (abstract is cut off), so critical details on sample complexity, convergence guarantees, and empirical validation are missing—we cannot assess whether H-MRS outperforms baselines on real datasets or under what sample sizes it becomes reliable. The method inherits the general challenge of causal discovery: identifiability is limited by the Markov equivalence class, and without additional constraints (causal sufficiency, no hidden confounders, no selection bias), the discovered DAG may not be unique or correct. The moment-ratio criterion, while theoretically motivated for positive distributions, may be sensitive to outliers or heavy-tailed distributions common in real financial and biological data, and robustness properties are not discussed.
Research Context
This work extends the causal discovery literature (building on foundational DAG learning methods like GES, PC, and FCI) by introducing domain-specific scoring functions tailored to positive-valued variables, which have received limited attention in the classical DAG learning community. It bridges ideas from multiplicative models in econometrics and log-linear models in statistics with modern graphical causal inference. The paper likely benchmarks against standard DAG learning baselines (which assume linearity or additivity) on synthetic positive-valued datasets and potentially real datasets from genomics or finance. Future work would naturally explore how H-MRS scales to high-dimensional settings, how it handles mixed data types (some positive, some continuous), and whether the moment-ratio score can be extended to latent variable models or cyclic structures.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::