FuzzyCAL: A Fuzzy-Logic Enhanced Causal Attention GNN for Robust Cocaine Use Disorder Classification

Fuzzy logic has emerged as a powerful tool for handling uncertainty and imprecision in machine learning, enabling models to make robust predictions in complex, noisy domains. In this work, we integrate fuzzy-set theory with causal attention learning to improve the classification of Cocaine Use Disorder (CUD) from resting-state functiona Magnetic Resonance Imaging (fMRI) brain networks. Building on the Causal Attention Learning (CAL) framework, we introduce FuzzyCAL, which augments node- and edge-level causal masks with fuzzy membership functions that quantify “weak,” “medium,” and “strong” relevance of graph components. This fusion of fuzzy logic and causal Graph Neural Networks (GNNs) not only preserves the interpretability of causal attention but also adapts to the inherent variability of neuroimaging data. Through extensive ۵-fold cross-validation on the SUDMEXCONN dataset, FuzzyCAL achieves a mean test accuracy of ۸۶.۲۰% (±۷.۱۸%) and F۱-score of ۸۵.۸۷% (±۷.۲۳%), outperforming both a non-causal GNN baseline and the original CAL model. We further demonstrate how fuzzy-weighted causal masks reveal anatomically meaningful biomarkers in temporal and sensorimotor cortices. Our results suggest that embedding fuzzy reasoning into causal graph models enhances both predictive performance and neuroscientific interpretability, offering a promising direction for precision diagnostics in substance use disorders.

کلیدواژه ها:

Fuzzy logic ، Causal Attention ، Graph Neural Networks ، Cocaine Use Disorder ، Functional Connectivity