A Generative-Fuzzy Approach for Transparent Rule Induction from Latent Embeddings

We propose a hybrid Generative-Fuzzy framework to reconcile predictive performance with interpretability in high-dimensional settings by coupling a Variational Autoencoder (VAE) with a Takagi-Sugeno Fuzzy Inference System (FIS) induced via pyFUME. The VAE compresses the original feature space into a smooth latent manifold suitable for fuzzy partitioning, while the FIS translates latent regions into compact human-readable IF-THEN rules. To bridge latent semantics to physical variables, we introduce a feature-attribution mechanism that combines the decoder’s Jacobian sensitivity with latent traversal effects, thereby grounding rule antecedents in the original features. We validate the proposed approach on a single-cell transcriptomics benchmark (FCGR3A+ monocytes under IFN-β stimulation), showing that the induced rule base achieves performance comparable to a Random Forest trained on the full feature space, despite operating on a 10-dimensional latent embedding. In addition, the attribution analysis yields biologically coherent drivers (e.g., ISG15, CXCL10, IFIT family) and a near-perfect agreement between local (Jacobian) and global (traversal) effects for the principal activation axis in the latent space, supporting stable and faithful interpretability.