Integration of Fuzzy Matching and Domain Rules for Identifying Bali's Indigenous Banjar-Based Addresses in Last-Mile Delivery Without Predefined Gazetteers

Identifying residential addresses in regions that depend on culturally embedded locality markers presents a significant challenge for geocoding and last-mile logistics, particularly when such references are absent from administrative gazetteers. In Bali, shipment records frequently incorporate indigenous Banjar-based address components, which introduce ambiguity and diminish courier-routing accuracy. This study proposes a hybrid framework that integrates fuzzy matching with domain-specific rules to identify Banjar references from unstructured address texts without relying on predefined gazetteers. Three similarity algorithms, namely Levenshtein Distance, Partial Ratio, and Token Sort Ratio, were combined into a Hybrid Mix Score to generate robust candidate matches. Domain rules, including prefix normalization, Banjar-Village-District hierarchy validation, and semantic disambiguation filters, were applied to eliminate linguistically similar but geographically invalid candidates. Using 17,354 cleaned delivery records from Pos Indonesia, the hybrid framework significantly enhanced interpretation reliability, with approximately 95% of all addresses converging to a single Highest Valid Candidate (HVC). The final predictions were linked to verified geographic centroids, enabling operationally meaningful location references. The results demonstrate that combining multi-metric fuzzy similarity with contextual domain constraints provides an effective and reproducible solution for geocoding indigenous Banjar-based addresses in last-mile delivery environments that lack standardized gazetteers.