A novel hybrid approach for positive-valued DAG learning
A new algorithm infers cause-and-effect relationships from strictly positive data — like gene counts or asset prices — more naturally than standard methods.
The Thesis
Most causal-discovery algorithms assume that variables can take any value and that effects add together linearly. Many real-world quantities — gene expression levels, stock prices, revenues — are always positive and multiply rather than add. This paper proposes H-MRS (Hybrid Moment-Ratio Scoring), an algorithm that exploits a mathematical property of positive numbers to rank variables into a causal order, then prunes that ordering into a directed acyclic graph (DAG — a network of arrows showing who causes whom, with no feedback loops). The core insight is that a specific ratio of statistical moments (summary numbers derived from a variable's distribution) reliably identifies which variable is the cause and which is the effect when both must be positive. Results so far are limited to synthetic data, so real-world performance remains unproven.
Catalyst
Causal discovery for non-Gaussian, positive-valued data has lagged behind the explosion of methods for normally distributed variables. Genomics and single-cell RNA sequencing have created enormous datasets of strictly positive counts that don't fit standard assumptions, creating demand for tailored methods. Simultaneously, the theoretical toolkit for exploiting distributional shape as a causal fingerprint has matured enough to support moment-based approaches.
What's New
Earlier causal-discovery systems such as PC (a constraint-based search algorithm) and NOTEARS (which frames DAG learning as a continuous optimization problem) were designed for data that can be negative and follow additive noise models. Log-linear or multiplicative structure breaks their core assumptions, often producing wrong causal orderings. H-MRS instead works directly with moment ratios on the raw (positive) scale while fitting regression models on the log scale, which is the natural scale for multiplicative processes — a combination previous methods did not use.
The Counter
The paper tests H-MRS only on synthetic data it generated itself using the same log-linear model the algorithm assumes — this is the most favorable possible evaluation environment. Real gene-expression or financial data rarely obeys clean log-linear structure; noise is correlated, distributions have heavy tails, and feedback loops exist that DAGs cannot represent. The moment-ratio criterion works in theory when the data-generating process matches the model, but the paper provides no evidence it degrades gracefully when that assumption is violated. Well-established methods like NOTEARS and its variants have already been extended to handle non-Gaussian noise, and the claimed advantage over them on realistic positive-valued data is asserted but not demonstrated. Without benchmarks on real genomic or financial datasets, the practical value over simply log-transforming the data and applying a standard causal algorithm is unclear. The greedy ordering procedure is also known to accumulate errors early in the search, a weakness the paper acknowledges but does not fully address.
Longs
- ILMN (Illumina) — genomics data provider whose customers would use causal tools like this
- PACIFIC BIOSCIENCES (PACB) — long-read sequencing generates positive-count expression data
- NTRA (Natera) — genomic analytics firm that could integrate causal discovery in pipelines
- ARKG (ARK Genomic Revolution ETF) — broad genomics-sector exposure
Shorts
- Vendors of standard Gaussian-assumption causal tools (e.g., TETRAD-based commercial packages) — their core modeling assumption is violated on positive data
- Consulting firms offering causal inference services built on additive structural equation models — a domain-specific alternative undercuts their methodology for biotech clients
Enablers (Picks & Shovels)
- scikit-learn (open-source Python ML library used for Ridge and Elastic Net regression components)
- NumPy/SciPy ecosystem — moment computation infrastructure
- causal-learn (open-source causal discovery library that provides baselines and integration points)
- Single-cell RNA-seq data repositories such as GEO and CELLxGENE — testing ground for positive-valued biological data
Private Watchlist
- Causality Link — causal AI for financial forecasting
- Radical Ventures portfolio companies in causal ML
- Cellarity — causal modeling of cell biology for drug discovery
Resources
The Paper
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 $S$ 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.