Causal discovery—the process of uncovering cause-and-effect relationships from observational data—is a fundamental challenge in machine learning and statistics. A major hurdle in this field is the presence of latent variables, which are unmeasured factors that secretly influence the observed system.
To make the mathematics of causal discovery tractable, most existing algorithms rely on a restrictive premise known as the "pure children" assumption. This assumes that every latent variable has at least one observed child whose only parent is that specific latent variable. In the real world, this is rarely true. For example, a company's observed stock price is influenced by its intrinsic value (a latent variable), but also by the broader macroeconomic environment (another latent ancestor). Because the stock price has multiple parents, it isn't a "pure child," causing standard discovery methods to break down.
This paper introduces a novel concept called homologous surrogates to move beyond this restrictive assumption, providing a theoretical framework and algorithm (CDHS) capable of recovering causal graphs in complex, real-world scenarios.
The Core Innovation: Homologous Surrogates
The authors propose a more relaxed and realistic framework based on the Linear Non-Gaussian Acyclic Model (LiNGAM). Their key innovation is the introduction of the Homologous Surrogate.
Unlike a pure child, a homologous surrogate is an observed variable that is a child of a latent variable but is allowed to have other parents, provided those additional parents are ancestors of the latent variable.
This shift from "pure children" to "homologous surrogates" acknowledges that observed proxies in fields like economics, biology, and the social sciences are inherently messy and interconnected. By allowing observed variables to have more flexible parent structures, the applicability of causal discovery is significantly broadened.
Methodology: Peeling Back the Causal Layers
To reconstruct the causal graph, the authors utilize the non-Gaussian nature of the data. By assuming the data distribution is non-Gaussian, the algorithm can use high-order cumulants to identify causal directions that would be mathematically indistinguishable in a standard Gaussian (Normal) distribution.
The recovery process is split into two distinct stages:
- Partial Recovery: Under a weak assumption (where each latent variable has at least one homologous surrogate), the algorithm uses pseudo-residuals and independence tests to identify the ancestors of each variable.
- Full Recovery: Under a stronger assumption regarding "generalized" surrogates, the algorithm employs high-order cumulants to estimate exact mixing coefficients and direct parental relationships.
Operationally, the algorithm works "top-down." It identifies root variables first and then mathematically "removes" their effects from the data via residuals. This progressive identification simplifies the remaining graph step-by-step, preventing the complexity of the computational search space from exploding.
Experimental Validation
The authors validated their CDHS algorithm against both synthetic and real-world datasets:
- Synthetic Data: The team generated distinct types of causal Directed Acyclic Graphs (DAGs) specifically designed to violate the "pure children" assumption. In these complex cases, the proposed method achieved an impressive F1-score of 0.94 and a Correct-Ordering Rate of 0.97. In contrast, established baseline methods like GIN and PO-LiNGAM failed entirely, falling below a 0.20 F1-score. Furthermore, the algorithm demonstrated vastly superior computational scaling compared to PO-LiNGAM, which suffers from exponential time complexity.
- Real-World Data (Holzinger and Swineford, 1939): Applied to a classic psychometric dataset containing mental ability test scores for children, the algorithm successfully extracted meaningful latent structures. It correctly identified the "Textual" latent factor and logically grouped "Visual" and "Speed" factors, proving its efficacy on human-centric data.
Limitations and Future Directions
While a significant step forward, the framework does have limitations. Currently, the algorithm cannot handle "deep" latent structures—scenarios where a latent variable has no observed children at all and only influences other latent variables.
Additionally, because the method relies on high-order cumulants, it is sensitive to outliers and generally requires larger sample sizes to function optimally (performance dips were noted when sample sizes dropped to 5,000). Finally, the model strictly assumes the causal graph is a DAG (no feedback loops) and that the data distribution is stationary. Extending this homologous surrogate framework to cyclic or time-varying systems remains an exciting open challenge for future research.
Summary updated