Objectives With the continuous expansion of urban complexes, subterranean infrastructures, and multilevel transportation hubs, building interiors have become increasingly three-dimensional and intricate. As a consequence, traditional GPS signals are severely attenuated by floor slabs, rendering them inadequate for reliable indoor positioning. In this context, indoor points of interest (POI) have emerged as pivotal spatial anchors for seamless navigation and location-based services, and the rigorous selection of salient indoor POI has consequently become a prominent research focus.

Methods To address the challenge of quantifying indoor POI saliency, we propose a data-balance-aware enhanced random forest (RF) framework. The existing approaches suffer from evident shortcomings in class imbalance and in capturing the highly non-linear dependencies between saliency and multi-dimensional predictors. Therefore, a comprehensive indicator set comprising 34 fine-grained features is systematically constructed from three complementary perspectives,including visual saliency, semantic saliency, and spatial saliency. The synthetic minority over-sampling technique (SMOTE) was employed to mitigate class imbalance, and a feature selection scheme driven by importance weighting was subsequently implemented. The final RF model, trained on the rebalanced and optimized feature space, delivers high accuracy and strong generalization for indoor POI saliency evaluation.

Results Empirical research results show that the proposed model demonstrates outstanding performance on the indoor POI dataset. The accuracy, precision, recall, weighted F1 score, and area under the curve have reached 0.987, 0.984, 0.987, 0.987, and 0.999, respectively. Compared with the RF model without data balancing, the performance of the proposed model has been doubled. Compared with other traditional models, such as support vector machine and genetic programming, the performance of the proposed model has been increased by 15% and 5%, respectively. Moreover, the proposed model exhibits good generalization performance on the test set.

Conclusions We introduce a data-balance-aware RF model, offering a unified framework for indicator selection, imbalanced-sample treatment, and robust model construction in indoor POI saliency assessment, thereby conferring both theoretical and practical impetus to high-precision indoor positioning and intelligent navigation services.