Abstract
Microfinance institutions frequently operate with sparse, imbalanced, and incomplete borrower records, which limits the reliability of traditional credit-scoring models. This study proposes an AI-driven micro-loan eligibility framework designed specifically for low-data environments by integrating transfer learning, domain adaptation, and multi-level explainabil-ity. Using a FICO-derived micro-lending dataset (9,578 records, 14 variables; 83.99% vs 16.01% class imbalance), baseline models trained on constrained data fractions (1–100%) exhibited increasing ROC-AUC but declining recall, confirming well-documented challenges of data scarcity in microfinance. (up to 0.6439 ± 0.0145 at 100%) (from 0.2500 ± 0.0559 at 1% to 0.1211 ± 0.0323 at 100%), confirming the well-documented challenges associated with data scarcity in microfinance. To address this limitation, pretrained gradient-boosted models were fine-tuned on limited target samples, and importance-weighted domain adaptation (SSWMD) was employed to mitigate source–target distribution mismatch. These approaches yielded substantial improvements, increasing ROC-AUC from 0.6308 (scratch, 10%) to 0.8061 (fine-tuned) and 0.8484 (domain-adapted), with the pretrained full-data reference achieving 0.9496. This result supports prior evidence on the value of cross-domain knowledge transfer in financial risk modeling. Expanded SHAP analysis (global, dependence, and local explanations) identified consistent key predictors, supporting transparency, practitioner trust, and regulatory accountability. Overall, the results demonstrate that accurate, interpretable, and data-efficient micro-loan eligibility assessment is achievable, even under severe data constraints, through the use of transfer learning and domain adaptation.