Research map / Probabilistic and scientific modeling / Causal discovery

Detecting and Identifying Selection Structure in Sequential Data

ICML 2024 · Probabilistic and scientific modeling › Causal discovery · approach: Causal approaches to fairness · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that selection structure (selection pairs, direct relations, confounded pairs) in sequential data is identifiable from observational data alone, and proposes an O(N^2) constraint-based algorithm to detect it, validated on synthetic data and real music.

Problem

Selection (preferential inclusion of data points based on latent objectives) is usually treated as bias to correct, and autoregressive models in sequential data may overlook or confuse selection with direct causal relations or latent confounding. The paper asks whether selection structure can be ide

Approach

Identifies selection pairs, direct relations and confounded pairs in sequential data using pairwise conditional independence tests exploiting the collider created by conditioning on latent selection variables, under structural conditions that exclude degenerate cases; no parametric assumptions or interventions are needed.

Key results

• On synthetic data with 8-20 observed variables, the proposed algorithm recovers selection pairs and other dependencies under the structural conditions, while the base model shows lower precision for first-order direct relations due to misclassified confounded pairs.
• The algorithm handles a graph with roughly 600 observed variables within one hour on CPU with 12 GB memory.
• On real music, the identified selection pairs (e.g., X1 and X5 in Hotel California) are qualitatively aligned with the Andalusian cadence and the Panis Angelicus lyrics, with no quantitative metric reported.

Measured effects reported in the paper

Surprising findings

• Latent confounders and latent selection variables are distinguishable from observational data alone, given the stated structural conditions.

Limitations

• Requires structural conditions (degenerate-case exclusions) that may not hold in practice; the authors suggest condition checkers that flag ambiguity.
• Real-world validation is qualitative, with alignment judged against musical knowledge rather than ground truth.
• Current work is limited to understanding without leveraging the discovered structure for downstream tasks such as foundation models.

Experimental setup

Datasets: synthetic linear Gaussian SCM data Hotel California guitar soundtrack clips Panis Angelicus string section soundtrack clips

Benchmarks: synthetic graphs with 8-20 observed variables synthetic graphs with 50-1000 variables (scalability)

Metrics: precision recall running time

Baselines: base model without structural conditions

Builds on: FCI PC algorithm d-separation causal-learn

Closest papers by meaning

• Latent Variable Causal Discovery under Selection Bias (ICML 2025 · similarity 0.86)
• Bivariate Causal Discovery using Bayesian Model Selection (ICML 2024 · similarity 0.86)
• Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation (NeurIPS 2026 · similarity 0.86)
• Identifying Structural Biases from Causal Mechanism Shifts (NeurIPS 2026 · similarity 0.85)
• When Selection Meets Intervention: Additional Complexities in Causal Discovery (ICLR 2025 · similarity 0.84)
• Local Causal Structure Learning in the Presence of Latent Variables (ICML 2024 · similarity 0.84)