Concept Bottleneck Models

The defining architectural constraint of a CBM. The model first predicts a set of predefined, human-specified concepts from the raw input. The final prediction is then made exclusively from these concept scores, not the raw features. This creates an information bottleneck that forces the model to learn a reasoning process aligned with domain expertise. For fraud detection, concepts might include 'transaction velocity,' 'merchant category mismatch,' or 'device location anomaly.'

A critical advantage over black-box models. Because the final prediction depends solely on concept scores, a human expert can manually correct mispredicted concepts at inference time and observe the impact on the final output. This allows for real-time debugging and override. For example, an investigator can set the 'known device' concept to 'true' and see if a transaction is still flagged as fraudulent, directly testing the model's reliance on that signal.

CBMs are typically implemented in two forms:

• Sequential CBM: The concept predictor and label predictor are trained jointly end-to-end. The final layer receives concept probabilities.
• Independent CBM: The concept predictor is trained first, and the label predictor is trained separately on the ground-truth concepts. This allows the label predictor to learn from perfect concept labels, but may suffer from distribution shift when fed predicted concepts at test time.

A known failure mode where the model encodes information in the concept layer that is not semantically aligned with the intended concept, effectively bypassing the bottleneck. For instance, a model might encode the raw transaction amount within the 'high-risk country' concept score. Mitigation strategies include concept regularization, adversarial training on the concept layer, and using binary rather than continuous concept representations to limit information capacity.

A practical extension that allows the model to use both interpretable concepts and a residual, uninterpretable embedding for its final prediction. This trades off some interpretability for increased accuracy. The residual pathway captures patterns not covered by the predefined concept set, while the concept pathway remains fully auditable. This is often a pragmatic choice for complex fraud detection tasks where perfect concept coverage is infeasible.

The primary bottleneck in deploying CBMs is the manual curation of the concept set. Each training instance must be annotated with concept labels, which is expensive. Emerging techniques include using large language models to automatically propose and label concepts from raw transaction descriptions, or leveraging unsupervised concept discovery methods that find interpretable directions in a pre-trained embedding space without explicit supervision.

Regulatory-mandated explanations required by statutes like the Fair Credit Reporting Act (FCRA) and Equal Credit Opportunity Act (ECOA). These codes must disclose the principal factors driving an adverse model decision.

• Fraud Context: When a transaction is blocked, reason codes must justify the action to both regulators and customers.
• CBM Alignment: Concept Bottleneck Models naturally produce concept-level reason codes (e.g., 'Atypical merchant category for this account') that are more auditable and human-comprehensible than raw feature attributions from SHAP or LIME.