Chlorinated organophosphate esters (Cl-OPEs), a major class of organophosphate ester pollutants, are ubiquitously detected in industrial soils, yet quantitatively apportioning these contaminants remains a critical challenge. Traditional machine learning-based source apportionment approaches are often undermined by the spatial clustering of monitoring data, a phenomenon known as geographical bias. To address this limitation, this study first identified geographical bias through exploratory data diagnostics and subsequently constructed a robust random forest-based machine learning prediction framework that excludes geographical features to eliminate the confounding effect of spatial bias. The optimized model exhibited excellent reliability in predicting soil Cl-OPEs concentrations (test set R2 = 0.90). Model interpretability analysis via feature importance assessment confirmed E-waste processing (site function category_E-waste processing) as the primary driver of Cl-OPEs pollution. To validate the model and explore overlooked sources, we also conducted an empirical case study of the petrochemical industry. Comparative analysis not only confirmed E-waste processing areas as the primary source but also identified the petrochemical industry as a previously underestimated yet significant source of emissions, with its coil Cl-OPEs levels markedly exceeding those of other traditional industrial areas. Overall, this study proposes a novel, geographical-bias-free machine learning source apportionment method and provides new preliminary empirical evidence supporting pollution control of Cl-OPEs.