Research map / Probabilistic and scientific modeling / Causal discovery
Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning
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.
Introduces Causal Concept Graph Models, concept-based architectures whose inference follows a learned sparse DAG among concepts and tasks, making the decision process causally transparent by design and enabling interventions and counterfactuals.
Problem
Deep neural networks are causally opaque: their hidden causal structure of decisions cannot be inspected, so users cannot answer interventional or counterfactual questions about model reasoning, which hinders verification in high-stakes settings.
Approach
Duplicate the layer of interpretable variables into explaining copies and explained variables, and learn a sparse, acyclic adjacency matrix among concepts and tasks (with an acyclicity penalty and do-operator training), so that each prediction is a cause-effect chain of concepts that can be unfolded, intervened on, and queried counterfactually—unlike CBMs, which assume independent concepts that di
Key results
- Causal CGM label accuracy is within about one point of black box, CBM and CEM on Checkmark, dSprites and CelebA (e.g., CelebA 78.23 vs 79.55 black box), and reaches 93.32 on CIFAR10 vs 94.85 black box.
- Ground-truth interventions on Causal CGM improve non-intervened nodes by about 15 percentage points after 7 interventions on CelebA, while CBM and CEM need interventions on all concepts to achieve similar gains.
- Blocking via do-interventions yields a residual concept causal effect of 0% for Causal CGM across all datasets, versus about 98-106% for CBM and CEM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a Causal CGM rather than a black-box model | classification models · Checkmark, dSprites, CelebA, CIFAR-10 | label accuracy | Mixed |
| Use a ground-truth causal graph fixed in the model instead of learning the graph end-to-end | Causal CGM · Checkmark, dSprites, CelebA | label accuracy | Mixed |
| Apply ground-truth interventions to mispredicted ancestors | Causal CGM · CelebA | accuracy of non-intervened nodes: ~15 percentage points after only 7 ground-truth interventions | Helped |
| Use Causal CGM rather than CBM or CEM for ground-truth interventions on concepts | concept-based models · concept nodes | concept accuracy: up to ~25 percentage points | Helped |
| Block all causal paths between related variables using do-interventions | Causal CGM · Checkmark, dSprites, CelebA, CIFAR-10 | Residual Concept Causal Effect: Table 2: 97.99 {\scriptstyle \pm 5.64} -> 0.00 {\scriptstyle \pm 0.00} (Checkmark CBM vs Causal CGM); 100.00 {\scriptsty | Helped |
| Apply the same blocking procedure with CBM | CBM · causally related variables | average causal effect after blocking: 3 percentage points | No clear effect |
| Apply blocking using CEM | CEM · causally related variables | Residual Concept Causal Effect: >100% in some cases (Table 2: 102.58 {\scriptstyle \pm 12.95}, 106.00 {\scriptstyle \pm 0.50}) | Mixed |
Surprising findings
- CEM can increase rather than decrease the causal effect after blocking, with residual effect above 100% on CelebA.
Limitations
- Quality of the learned causal graph depends on dataset quality and annotations; missing or noisy labels may yield suboptimal graphs.
- Cyclic variables require special unfolding techniques not addressed; acyclicity is enforced via penalty and edge removal at inference.
- The learned graph reflects the model's inference rather than the data-generating process, so it is not suitable for understanding the dataset distribution.
Experimental setup
Datasets: Checkmark dSprites CelebA CIFAR10
Benchmarks: Checkmark dSprites CelebA CIFAR10
Metrics: label accuracy task accuracy after ground-truth interventions residual concept causal effect probability of necessity and sufficiency
Baselines: Black box CBM CEM Causal CGM+CD
Builds on: Concept Bottleneck Models Concept Embedding Models Deep Concept Reasoning Structural Causal Models NOTEARS-style acyclicity constraint
Closest papers by meaning
- From Black-box to Causal-box: Towards Building More Interpretable Models (NeurIPS 2025 · similarity 0.90)
- Causally Reliable Concept Bottleneck Models (NeurIPS 2025 · similarity 0.89)
- The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability? (NeurIPS 2025 · similarity 0.88)
- Causality can systematically address the monsters under the bench(marks) (NeurIPS 2026 · similarity 0.88)
- Rethinking Causal Relationships Learning in Graph Neural Networks (AAAI 2024 · similarity 0.88)
- Since Faithfulness Fails: The Performance Limits of Neural Causal Discovery (ICML 2025 · similarity 0.87)
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.