Causally Reliable Concept Bottleneck Models

Concept-based models are an emerging paradigm in deep learning that constrains the inference process to operate through human-interpretable variables, facilitating explainability and human interaction. However, these architectures, on par with popular opaque neural models, fail to account for the true causal mechanisms underlying the target phenomena represented in the data. This hampers their ability to support causal reasoning tasks, limits out-of-distribution generalization, and hinders the implementation of fairness constraints. To overcome these issues, we propose Causally reliable Concept Bottleneck Models (C$^2$BMs), a class of concept-based architectures that enforce reasoning through a bottleneck of concepts structured according to a model of the real-world causal mechanisms. We also introduce a pipeline to automatically learn this structure from observational data and unstructured background knowledge (e.g., scientific literature). Experimental evidence suggests that C$^2$BMs are more interpretable, causally reliable, and improve responsiveness to interventions w.r.t. standard opaque and concept-based models, while maintaining their accuracy.

💡 Research Summary

The paper addresses a fundamental shortcoming of existing Concept Bottleneck Models (CBMs): they treat concepts as merely statistically correlated intermediates, ignoring the true causal relationships that generate the data. To remedy this, the authors introduce Causally reliable Concept Bottleneck Models (C²BMs), a new class of architectures that embed a causal graph of concepts—derived from real‑world mechanisms—directly into the model’s reasoning pipeline.

A C²BM consists of two main components. First, a neural encoder g(·) maps raw inputs X to a set of latent exogenous variables U, which capture unobserved factors. Second, a parametric Structural Causal Model (SCM) ⟨V, U, F_Θ, P(U|X)⟩ operates on the concept set V. The SCM’s directed acyclic graph (DAG) G encodes causal links among concepts; each structural equation f_i ∈ F_Θ is a weighted linear combination of its parent concepts, but the weights θ_i are not fixed. Instead, a hypernetwork r_i(·) receives the corresponding exogenous variable U_i (or the whole U) and dynamically generates θ_i for each input, allowing the linear form to approximate nonlinear causal mechanisms. This design yields an interpretable yet expressive model that can answer both forward (prediction) and interventional queries.

Because C²BMs require a labeled concept dataset and a causal graph—resources often unavailable—the authors propose a fully automated pipeline to construct them from (i) an unlabeled dataset D_x and (ii) an unstructured knowledge repository K (e.g., scientific literature). The pipeline has three stages:

1. Concept discovery and labeling – Large Language Models (LLMs) with Retrieval‑Augmented Generation (RAG) are prompted to list concepts relevant to the target task. The returned candidates are filtered for brevity, distinctiveness, and presence in the data. Labels are then generated automatically by projecting both images (or other raw inputs) and concept names into a shared CLIP embedding space, assigning the most compatible concept to each sample.

2. Causal graph discovery – From the labeled data, the Greedy Equivalence Search (GES) algorithm produces a Completed Partially Directed Acyclic Graph (CPDAG), representing the Markov equivalence class of plausible causal structures. For each undirected edge, the pipeline queries the LLM‑RAG ten times with prompts such as “Does A cause B?” and adopts the majority‑vote direction, thereby injecting domain knowledge to resolve ambiguities and discard spurious links.

3. Learning structural equations – Given the finalized DAG and concept annotations, separate hypernetworks r_i(·) are trained to predict the linear coefficients of each structural equation from the corresponding exogenous variable U_i. The loss combines concept prediction error, downstream target prediction error, and a regularizer encouraging consistency with the causal graph. The authors prove in the appendix that, regardless of the underlying DAG, C²BM is a universal approximator.

Empirical evaluation spans several domains, including medical diagnosis (lung cancer), object recognition, and synthetic benchmarks. The authors compare C²BM against (a) standard deep neural networks, (b) vanilla CBMs, (c) stochastic CBMs and Concept Graph Models (which capture associations but not causality), and (d) DiConStruct, a post‑hoc method that infers causal links from a trained DNN. Four metrics are reported:

• Causal consistency – measured by Structural Hamming Distance between the learned graph and a ground‑truth expert graph.
• Interventional efficiency – the number of concept interventions required to achieve a target accuracy improvement.
• Robustness to spurious correlations – performance under distribution shift (e.g., new hospital data).
• Fairness – reduction of bias with respect to sensitive attributes (gender, ethnicity).

C²BM consistently outperforms all baselines. Its learned graphs align far more closely with expert knowledge, and it achieves comparable or higher downstream accuracy while requiring far fewer interventions (often only two concepts) to correct predictions. Under distribution shift, performance degradation is limited to under 5 %, whereas baselines drop by 10–20 %. Fairness experiments show a marked decrease in demographic disparity, demonstrating that causal structuring can be leveraged to enforce ethical constraints.

The paper’s contributions are threefold: (1) a principled integration of causal modeling into concept‑based architectures, moving beyond the simplistic bipartite assumption; (2) an end‑to‑end automated pipeline that extracts both concepts and their causal relations from raw data and unstructured knowledge, dramatically lowering the need for expert annotation; and (3) a hypernetwork‑driven parametrization that preserves interpretability while capturing complex, input‑dependent causal effects.

Limitations include the reliance on linear structural equations (mitigated but not eliminated by hypernetworks) and the dependence on the quality of LLM‑RAG queries, which may vary across domains. Future work could explore richer nonlinear SCMs (e.g., neural ODEs), continual updating of the causal graph in streaming settings, and the use of domain‑specific LLMs to improve knowledge extraction.

In summary, C²BMs represent a significant step toward truly explainable AI systems that are not only transparent but also causally trustworthy, enabling reliable intervention, robust generalization, and principled fairness enforcement.