Target Centric Causal Structure Learning

604 concepts250 papers1 proposal

Traditional causal discovery algorithms attempt to map an entire system by reconstructing a global Directed Acyclic Graph (DAG). While a complete causal map is valuable, finding it is an NP-hard problem that quickly becomes computationally intractable as the number of variables grows. As Lu Zhang et al. note in Achieving Non-Discrimination in Data Release, learning a full global graph fails to scale to high-dimensional datasets.

Beyond computational complexity, global discovery suffers from statistical error propagation. In a massive network, executing thousands of interconnected conditional independence tests inevitably introduces false positives. A single mistaken edge orientation in a remote part of the graph can cascade, corrupting the structural accuracy of the entire network.

Target-centric causal structure learning bypasses these limitations. If the ultimate goal is to understand or intervene on a specific target variable—such as a disease outcome, a decision node, or a treatment effect—reconstructing the entire universe is unnecessary. Instead, researchers can isolate and map only the immediate local neighborhood of the target variable.

The Markov Blanket as a Statistical Shield

The core mechanism behind local causal discovery is the Markov Blanket (or Markov boundary). For any given target variable, its Markov blanket is a localized substructure composed of its direct parents, its direct children, and the other parents of those children (spouses).

The fundamental intuition behind this approach relies on the Causal Markov condition: conditioning on a variable's Markov blanket renders that target variable statistically independent of the rest of the network. The blanket acts as a statistical shield, blocking external paths and isolating the target from distant variables.

       [Parents]
          │
          ▼
   ►► [Target] ◄◄  (Isolated by the Blanket)
          │
          ▼
       [Children] ◄── [Spouses]

By focusing exclusively on this boundary, algorithms can drastically simplify the search space. As Jean-Philippe Pellet et al. demonstrate in Finding Latent Causes in Causal Networks: an Efficient Approach Based on Markov Blankets, restricting d-separation tests to the local Markov blanket reduces the computational complexity of structural search from exponential to quadratic.

Furthermore, this localized blocking protects the target from external noise. Tian Gao et al. show in Local Causal Discovery of Direct Causes and Effects that using the Markov blankets of neighboring nodes to block external paths prevents distant, irrelevant variables from confounding local statistical tests. This minimizes edge orientation errors and ensures that local structure learning remains independent of the overall system size.

Orienting Edges with Local Constraints

A common challenge in local causal discovery is determining edge directionality. Without a global view of the graph, how can an algorithm distinguish a direct cause (parent) from a direct effect (child)?

Local discovery resolves this by exploiting structural and distributional invariants within the immediate neighborhood:

  • Local Structural Constraints: Algorithms can identify directionality by analyzing specific local connectivity patterns. For example, Zhuangyan Fang et al. demonstrate in A local method for identifying causal relations under Markov equivalence that evaluating chordless paths and maximal cliques within a variable's immediate neighborhood provides sufficient constraints to determine causal directionality without full graph reconstruction. Similarly, the authors of Towards Efficient Local Causal Structure Learning utilize localized "N-structures" to orient edges during the initial discovery phase, bypassing the need to map the full blankets of adjacent nodes.
  • Non-Gaussian and Cyclic Invariants: When systems contain feedback loops or non-Gaussian distributions, local statistical properties become highly informative. Haoyue Dai et al. show in Local Causal Discovery with Linear non-Gaussian Cyclic Models that Independent Subspace Analysis (ISA) can successfully isolate local causal neighbors and determine their strengths, even in massive systems containing cyclic relationships and hidden confounders.

Managing Hidden Confounders Locally

Real-world datasets are rarely fully observed; unmeasured latent confounders frequently threaten the validity of causal discovery. Global methods often fail to converge or require exhaustive graph enumeration when latent variables are introduced. Local discovery methods handle this by executing targeted, sequential searches to find valid adjustment sets.

In Local Causal Structure Learning in the Presence of Latent Variables, Feng Xie et al. introduce a sequential search strategy that first identifies a target's Markov blanket and then selectively evaluates neighboring variables from a managed waitlist. This iterative process achieves the exact same theoretical consistency as exhaustive global ancestral graph searches while requiring a fraction of the conditional independence tests.

Similarly, Daniel Malinsky et al. demonstrate in Estimating Causal Effects with Ancestral Graph Markov Models that local neighborhood searches can isolate valid adjustment sets in high-dimensional systems, preserving theoretical guarantees without the computational burden of global graph enumeration.

Efficiency Through Task-Oriented Inference

Ultimately, structural learning serves downstream tasks, such as predicting the outcomes of external manipulations or estimating heterogeneous treatment effects. Local causal discovery aligns perfectly with this pragmatic goal by focusing only on the features that matter.

For instance, in causal classification and uplift modeling, Jiuyong Li et al. show in A general framework for causal classification that identifying just the direct causal parents of an outcome variable provides a sufficient basis for predicting how an intervention will alter behavior.

This task-driven philosophy is formalized in representation learning by Zirui Yan et al. in Reward-oriented Causal Representation Learning. The authors demonstrate that aiming for full identifiability of a global latent graph introduces unnecessary data overhead. Prioritizing the recovery of only those local causal relationships necessary for a specific downstream reward optimizes sample efficiency and yields faster, more robust decision-making.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 270 papers
  2. Version 1 · 2026-10-04 · 319 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_local_causal_discovery_c7a3d05e,
  title = {local causal discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/local-causal-discovery/rdr_c7a3d05e2ce0483f87002a0f2e2edfd6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_local_causal_discovery_c7a3d05e_v2,
  title = {local causal discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/local-causal-discovery/rdr_c7a3d05e2ce0483f87002a0f2e2edfd6/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake