Recommender systems have transformed how people find films, restaurants and products online, but in the financial sector of fast-growing cities they still often fall back on a surprisingly crude method: find the nearest branch on a map and send the customer there. A new open-access study published in Discover Data on 13 April 2026 argues that this proximity-only logic routinely misdirects customers in dense urban environments, and it proposes a remedy built on geographic information systems. The framework, called Geo-Banking, was developed and evaluated using selected banks in Accra, Ghana, and combines spatial indexing, network-based route optimization, service validation and location awareness within a single architecture.
The problem the authors identify is concrete rather than abstract. A customer can be referred to the geographically closest bank branch that, in reality, does not offer the requested service, or is practically unreachable because of traffic. According to the paper, recommender applications in banking in cities like Accra are characterized by stagnant distance-based search, unreliable GPS positioning, unreliable geocoding accuracy, routes oblivious to congestion, and a lack of branch-scale verification of service availability. Such restrictions, the authors write, augment the risk of misguiding and inconveniencing customers. In a sprawling metropolis where road topology and congestion determine how long a journey actually takes, straight-line distance is a poor proxy for accessibility.
Geo-Banking addresses these weaknesses with a layered technical design. At its core is R-tree spatial indexing, a hierarchical arrangement of spatial features using minimum bounding rectangles that allows logarithmic-time retrieval of candidate branches instead of scanning every branch in the database. The system also performs congestion-aware routing over a weighted road graph using shortest-path algorithms facilitated by ArcGIS Network Analyst, so that recommendations are optimized by realistic travel time rather than Euclidean distance. Additional validation layers are designed to eliminate false positive geographic identifications and enhance the reliability of recommendations before they reach the user.
The recommendation pipeline itself proceeds in stages. First, hard constraints filter candidates: a bank qualifies only if it offers the requested service and lies within a specified search radius of the user. Second, a weighted multi-criteria scoring function ranks the surviving candidates using service matching, an inverse normalized distance score, user rating or reliability, and an availability or contextual factor such as opening hours, with weights summing to one. The top-ranked banks are then returned as recommendations. The authors also describe a confidence scoring mechanism that combines location accuracy, service certainty and routing validity, excluding branches that fall below a predefined threshold.
Robustness receives particular attention, reflecting conditions in rapidly developing urban areas where GPS positioning can be poor, address systems incomplete and connectivity unstable. A dedicated failure-handling workflow considers denied location permission, GPS inaccuracies and changing connectivity, adding verification steps, fallback procedures such as manual location entry and offline caching, and exception handling to maintain service continuity. A geospatial accuracy and validation framework further checks coordinate precision, geocoding uncertainty and spatial error propagation. Rooftop geocoding is preferred over rough street-segment placement for its greater positional accuracy, and low-confidence geocode matches are set for manual verification. A distance consistency check compares straight-line and network distances, flagging large discrepancies as possible geographic mismatches or routing anomalies.
The empirical evaluation used data from the Accra Metropolitan Area, covering central business districts and peri-urban zones. The dataset comprised 15 registered commercial bank branches, 3,000 simulated and real user queries representing different service request types, and 43 categorized banking services such as ATM withdrawals, deposits, loans and foreign exchange. The data combined geocoded branch locations, service availability records and anonymized user location queries. The dataset was split with 80 percent allocated to training and 20 percent to testing, and all models were evaluated on the same data, the same hardware and the same metrics, including precision, recall, F1-score, arithmetic mean, mean absolute error and normalized mean absolute error.
The reported results favor the proposed framework across the metrics. At the initial evaluation coefficient, Geo-Banking attained a recall value of 2.42, compared with 2.40 for the non-GIS filtering baseline, 2.37 for the raw distance-based baseline, 1.69 for the traditional approach and 0.11 for Google Maps. Its F1 score of 1.40 exceeded the corresponding figures of 1.27, 1.25, 0.85 and 0.12. On error measures, Geo-Banking achieved an MAE of 0.02 at the initial evaluation coefficient, lower than 0.14, 0.15, 0.43 and 0.87 for the comparison methods, with a normalized MAE of 0.005 against 0.035, 0.038, 0.108 and 0.218. The authors also benchmarked against AHP multi-criteria ranking and MCLP coverage optimization, reporting top composite scores under AHP across eight evaluation coefficients and the highest coverage values, ranging around 0.87 to 0.93, under MCLP.
Ablation analysis was used to separate the contribution of the GIS components from the underlying recommender logic, with performance gains computed as the difference between the full model and a version without GIS integration, and latency reduction computed between indexed and non-indexed configurations. According to the paper, the ablation results support the conclusion that GIS-based spatial indexing and routing add to the performance improvements over traditional recommender logic, and that spatial indexing improves system responsiveness and decreases query execution time. The authors frame the outcome as a real-time and scalable urban financial service recommendation model suitable for smart-city implementation.
The study is grounded in a survey that motivated the design. Out of a customer population of 25,000, 390 responses were collected, alongside 34 staff responses. The respondents were described as technologically literate, with most customers in the 31-to-35-year age bracket and most holding tertiary education, and more than 90 percent of customers confirmed a prominent level of technology adoption. Yet a structural gap emerged: 70.3 percent of customers reported the lack of a geographical banking platform, 77.23 percent were unaware of Geo-Banking, and 82.35 percent of staff affirmed a lack of spatially enabled banking systems in their institutions. The authors conclude that transactional digital services are established in Accra’s banking ecosystem but spatial decision-support capabilities are absent, and that the problem is not a lack of services but their discovery.
The authors acknowledge limitations, including the geographic restriction to predetermined regions of Accra, a cross-sectional design, reliance on self-reported information and a small staff sample, all of which can restrict generalization. They also note that GIS in banking faces challenges such as prohibitive cost, data inconsistencies, lack of standardization and siloed data. Even so, they argue that the architecture, which requires proximity-based filtering, service matching and latency-optimized retrieval, could extend beyond banking to healthcare, retail, transportation, emergency management and education. For now, the study offers a quantified case that in congested cities, knowing where a bank is matters far less than knowing how to reach it and whether it can actually serve the customer who arrives.
Subject of Research: A geospatial location-aware recommender system for banking services in Accra, Ghana
Article Title: Geospatial-based location-aware recommendation of banking services in the world of big data
Article References: Asabere, N. Y., Asabere, J., Andoh, C., Lawson, G., Peprah, E. O., & Adjei, O. O. (2026). Geospatial-based location-aware recommendation of banking services in the world of big data. Discover Data, 4(1), Article 6. https://doi.org/10.1007/s44248-026-00106-1
Image Credits: AI Generated
DOI: 10.1007/s44248-026-00106-1
Keywords: geographic information systems, recommender systems, banking, spatial indexing, R-tree, route optimization, Accra, Ghana, smart cities, location-based services, geocoding, big data
News Source: Denise Maddox. (October 29, 2026). GIS-Powered Recommender Steers Bank Customers to the Right Branch in Congested Accra. Scienmag.
