<p>Under the context of increasing urban functional mix and spatial heterogeneity, accurately characterizing regional features requires overcoming the limitations of traditional spatial proximity assumptions. Existing studies predominantly rely on POI mixed networks to extract global regional features; however, the spatial distribution bias of high-frequency POI categories can overshadow the interaction information of low-frequency categories, making it difficult to capture deep features. To address this issue, we propose a regional representation method based on modeling bidirectional interrelationships between pairwise POI categories. This approach decouples regional features into interactions between pairwise POI categories by independently constructing association networks for each pair, thereby eliminating interference from multiple categories. And we designed the Positional Overlap Degree (POD) metric to quantify asymmetric interaction intensity between categories from two perspectives: the influence of the target category on its surroundings (forward) and the dependence of the surroundings on the target category (reverse), to analyze mixing features in pairwise categories within regions. Finally, using Xiamen Island as the experimental area, our validation based on POI data demonstrates that our model significantly outperforms baseline models in tasks such as population density prediction and housing price estimation and superior performance in identifying mixed distributions of urban functional zones. This proves that accurately modeling the bidirectional asymmetric associations between low-frequency categories (such as hospitals and shopping malls) and high-frequency categories (such as restaurants) better represents regional characteristics and effectively reveals coupling patterns among latent functions.</p>