Leakage-aware multimodal GeoAI for residential rent valuation: evidence from Arabic property listings in four Saudi cities

Abstract Automated valuation models (AVMs) increasingly combine structured property attributes with listing text and location, yet their reported accuracy depends on how the data are cleaned and partitioned. This study evaluates a multimodal GeoAI workflow for annual villa rents in Riyadh, Jeddah, Dammam and Al-Khobar using public Aqar listings that pair structured attributes with Arabic descriptions. Listings were linked to a national district gazetteer (98.5% of the analysis sample matched), from which district centroids, district area, distance to the city centre and an inverse-distance spatial lag of neighbouring district rents were derived. A data audit showed that 2,197 of the 3,718 published records (59.1%) were exact duplicates, rising to 86–92% in Dammam and Al-Khobar. Gradient boosting with structured, text and geospatial inputs was compared with hedonic and random-forest baselines under four validation designs. With duplicates retained, random cross-validation gave a median absolute percentage error (MdAPE) of 1.5% and an R² of 0.82 on log rent; after de-duplication these became 15.4% and 0.60, and under district-grouped validation 19.3% and 0.55. Under district-grouped validation, geospatial context reduced mean absolute error by 20.6% (95% district-bootstrap interval 14.6–26.1%) relative to structured attributes alone and removed significant residual autocorrelation in Riyadh (Moran’s I from 0.40 to − 0.01). Text reduced error by 6.2% on its own but added no significant gain once geospatial context was present. Split-conformal intervals reached 91.6% coverage at the nominal 90% level, while leave-one-city-out transfer degraded all models (MdAPE 26–34%). De-duplication and spatially grouped validation should accompany reported AVM accuracy.