Neural Basis Models are a class of inherently interpretable machine learning architectures within the generalized additive model framework that represent feature effects using a shared set of basis functions learned by a neural network. Instead of learning an independent non-linear shape function for every individual input variable, a neural basis model utilizes a single neural network to jointly learn a compact dictionary of continuous basis functions across all features, reconstructing each feature effect as a linear combination of these shared bases. This basis decomposition substantially reduces parameter complexity and computational overhead, allowing the architecture to scale efficiently to high-dimensional datasets while maintaining exact, transparent interpretability and supporting higher-order feature interactions.
