Credit default prediction framework for heterogenous ecosystems using A cross-dataset fuzzy hierarchical graph transformer

Credit default prediction is complicated by severe class imbalance, heterogeneous financial attributes, uncertain borrower relationships, and temporal changes in repayment behavior. Existing tabular and conventional graph-learning models generally represent financial relationships using fixed binary edges and are commonly evaluated on a single dataset, limiting their ability to capture relational uncertainty and demonstrate cross-dataset generalizability. This study proposes CredFuzz-HGT, a cost-sensitive fuzzy hierarchical graph Transformer for multi-benchmark credit default prediction. First, a dataset-adaptive financial graph constructor transforms heterogeneous application, credit-history, repayment, and borrower records into a unified typed graph. A fuzzy relation encoder then assigns learnable membership degrees to structural and temporal edges using borrower similarity, repayment consistency, relational confidence, and temporal recency. Subsequently, a hierarchical graph Transformer models applicant, loan, behavioral, and institutional-level dependencies, while a temporal attention pathway captures the evolution of repayment behavior. The structural and temporal representations are integrated using uncertainty-conditioned gated fusion. Minority-default detection is further improved through a cost-sensitive loss whose class weights are calculated exclusively from each training fold. CredFuzz-HGT was evaluated on three publicly available benchmarks: Home Credit Default Risk, LendingClub Loan Data, and UCI Default of Credit Card Clients. Leakage-controlled stratified five-fold validation produced ROC-AUC values of 0.842 ± 0.003, 0.891 ± 0.004, and 0.792 ± 0.006, respectively. The corresponding minority-class F1-scores were 0.421 ± 0.007, 0.684 ± 0.006, and 0.566 ± 0.009. Compared with the statistical, gradient-boosting, Transformer, and graph-learning baselines, CredFuzz-HGT improved ROC-AUC by 1.8–2.7 percentage points and minority-class F1-score by 2.6–4.3 percentage points. Ablation experiments demonstrated that fuzzy relation encoding, hierarchical message propagation, temporal attention, uncertainty-conditioned fusion, and cost-sensitive optimization each contributed to the final performance.