Research map / Probabilistic and scientific modeling / Causal discovery
Structured Learning of Compositional Sequential Interventions
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.
Proposes a compositional sequential intervention model (factorization of recursive intervention effects with exponential-decay warping) with identification results and a VAE-based learner (CSI-VAE) that predicts outcomes for unseen intervention sequences, outperforming black-box sequence models on synthetic and Spotify-based semi-synthetic data.
Problem
Predicting future behavior under sequences of categorical, sparse interventions, especially novel combinations never observed together, where black-box sequence models generalize poorly.
Approach
Model the effect of sequential interventions as a multilinear factorization in which unit-level latent coefficients are warped multiplicatively by per-intervention temporal effect functions (time-bounded or exponential-decay), with identification conditions derived for this structure and learned via amortized variational inference (GRU encoders).
Key results
- On the fully-synthetic simulator, CSI-VAE-1 achieves RMSE about 36.5 at T+1 and about 41 at T+2 to T+5, versus 93.7 or lower for best black-box GRU-2 and 88 to 230 for other black-box baselines.
- On the semi-synthetic Spotify simulator, CSI-VAE-1 achieves about 68.2 at T+1 and 79.6 at T+5, while the best black-box baseline (GRU-0) gets about 215 at T+1 and 117 at T+5.
- Ablations that drop the product (compositional) form (CSI-VAE-2) or replace the intervention-effect equation with a black-box (CSI-VAE-3) degrade performance substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the compositional CSI-VAE-1 structure instead of GRU-2 | Sequential intervention prediction · Fully synthetic, T+1 to T+5 | RMSE: 36.53 vs 93.73 (T+1); 41.32 vs 132.28 (T+5) | Helped |
| Use the compositional CSI-VAE-1 structure instead of GRU-2 | Sequential intervention prediction · Semi-synthetic Spotify, T+1 to T+5 | RMSE: 68.23 vs 154.18 (T+1); 79.63 vs 127.58 (T+5) | Helped |
| Add the product-form composition rather than relax it to a black-box MLP (ablation) | CSI-VAE · Fully synthetic, T+1 to T+5 | RMSE: 36.53 vs 97.80 (T+1); 41.32 vs 135.03 (T+5) | Helped |
| Add the product-form composition rather than relax it to a black-box MLP (ablation) | CSI-VAE · Semi-synthetic Spotify, T+1 to T+5 | RMSE: 68.23 vs 253.85 (T+1); 79.63 vs 302.83 (T+5) | Helped |
| Use the structured intervention-effect equation instead of a black-box sequence function (ablation) | CSI-VAE · Fully synthetic, T+1 to T+5 | RMSE: 36.53 vs 138.78 (T+1); 41.32 vs 125.55 (T+5) | Mixed |
| Use the structured intervention-effect equation instead of a black-box sequence function (ablation) | CSI-VAE · Semi-synthetic Spotify, T+1 to T+5 | RMSE: 68.23 vs 757.94 (T+1); 79.63 vs 634.72 (T+5) | Helped |
Surprising findings
- Black-box models with full intervention history (GRU-2, Transformer) do not reliably beat simpler black-box variants, and in several cases GRU-0 (no interventions) matches or beats intervention-aware baselines on the semi-synthetic data.
Limitations
- Assumes no unmeasured confounding among past actions and states; sequential ignorability is assumed.
- Identification requires a burn-in period and sufficient units per intervention level, with no free identification.
- Evaluation uses only simulators, not real-world intervention data.
Experimental setup
Datasets: Spotify user skip interaction data (semi-synthetic simulator) fully-synthetic simulator
Benchmarks: fully-synthetic simulator semi-synthetic Spotify simulator
Metrics: root mean squared error mean squared error two-sample t-test p-values
Baselines: GRU-0 GRU-1 GRU-2 LSTM Transformer CSI-VAE-2 CSI-VAE-3
Builds on: causal matrix factorization variational autoencoder conformal prediction blip models
Closest papers by meaning
- Deriving Causal Order from Single-Variable Interventions: Guarantees & Algorithm (ICLR 2025 · similarity 0.85)
- Learning Mixtures of Unknown Causal Interventions (NeurIPS 2024 · similarity 0.84)
- Learning Joint Interventional Effects from Single-Variable Interventions in Additive Models (ICML 2025 · similarity 0.84)
- Local Causal Structure Learning in the Presence of Latent Variables (ICML 2024 · similarity 0.84)
- Identifying General Mechanism Shifts in Linear Causal Representations (NeurIPS 2024 · similarity 0.84)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.