Adaptive entropy-guided access point selection for robust multi-floor Wi-Fi fingerprinting

Abstract Wi-Fi fingerprinting is widely used for indoor localization. Although a large number of observable access points (APs) can provide richer spatial information, retaining all APs introduces significant computational overhead, increased scanning latency, and reduced generalization under heterogeneity across floors, devices, and contexts. This paper presents an entropy-guided dimensionality reduction framework that selects the most informative APs while preserving localization accuracy. First, each AP is ranked using Shannon entropy computed from its empirical RSSI distribution. The approach is evaluated on the multi-floor JUIndoorLoc dataset (172 AP features, three floors, four devices, and multiple contextual conditions) using a Random Forest classifier with stratified 5-fold cross-validation, and robustness is further assessed under additive RSSI noise and random AP dropouts. Across floors, the proposed adaptive entropy selection achieves substantial dimensionality reduction with negligible accuracy loss: it selects 17 APs on F3 ( $$\approx $$ 90.11% reduction) with only $$-0.41\%$$ accuracy change, 8 APs on F4 ( $$\approx $$ 95.34% reduction) with $$-0.51\%$$ change, and 26 APs on F5 ( $$\approx $$ 84.88% reduction) with $$-0.42\%$$ accuracy loss compared to the baseline configuration using all APs. Robustness experiments show that entropy-based selection can improve resilience under measurement noise on certain floors (e.g., on F5, at $$\sigma $$ =1–3 dB the selected subset achieves 0.8472/0.7748/0.7070 versus 0.7674/0.7084/0.6659 for using all APs), while also revealing floor-dependent trade-offs where extremely compact subsets may become more noise-sensitive under strong perturbations. Overall, the proposed adaptive entropy framework provides a lightweight, floor-aware AP selection mechanism that reduces inference cost and scanning burden while maintaining competitive accuracy and improving robustness in realistic degradation scenarios. The generalizability of these findings to more complex and dynamic real-world environments warrants further investigation.