Concept models improve predictions while keeping user control intact
Residual Stochastic Concept Bottleneck Models
Machine Learning
Summary
Some computer models explain their reasoning using human-understandable concepts, letting users adjust predictions by changing these concepts. However, if the concepts don’t capture everything important, the model’s accuracy drops. The authors introduce a new kind of model that adds an extra part to handle missing information while still linking it to the known concepts, so user changes affect all related parts. This approach improves accuracy without losing the ability to let users guide the model’s decisions.
What this means in practice
- •For machine learning engineers: Build predictive models that maintain interpretability even when not all relevant concepts are known at training time.
- •For medical data scientists: Develop diagnostic tools allowing clinicians to intervene on known medical concepts while capturing unknown factors for better prediction.
Authors
Stephen Marcou, Mateo Espinosa Zarlenga, Sonia Laguna, Moritz Vandenhirtz, Julia E. Vogt, Mateja Jamnik
Abstract
Concept-based models (CMs) learn human-interpretable intermediate representations that enable users to understand the model's reasoning and steer its predictions through test-time concept interventions. However, in realistic settings, their performance can be limited by concept incompleteness, where the training concepts fail to capture all task-relevant information. Adding an unconstrained residual pathway to the model's predictive process can overcome the consequences of incompleteness. However, this introduces an inscrutable, leak-prone prediction pathway that may weaken concept interventions, as these often leave residual variables unchanged. To address this limitation, we introduce Residual Stochastic Concept Bottleneck Models (Res-SCBMs), a family of concept-based models that jointly model concept and residual variables through a multivariate Gaussian distribution. Res-SCBMs learn a covariance structure that explicitly captures dependencies between the known concepts and learnt residual variables, allowing interventions to propagate to statistically dependent residual variables through Gaussian conditioning. This extends the intervention capabilities of previous CMs without requiring concept completeness during training. Our results show that, across multiple concept-incomplete benchmarks, Res-SCBMs improve predictive performance over competing baselines while retaining effective test-time concept interventions.