Hierarchical supervised contrastive learning with AP-attention encoding for Wi-Fi fingerprinting indoor localization

Indoor localization based on Wi-Fi fingerprinting has attracted significant research attention due to its infrastructure-free deployment and the ubiquitous availability of IEEE 802.11 access points (APs). A widely adopted deep learning architecture employs a stacked autoencoder (SAE) for unsupervised feature pre-training followed by a supervised multi-label classifier for hierarchical building, floor, and reference-point estimation. However, the SAE pre-training optimizes a mean squared error (MSE) reconstruction objective that is fundamentally misaligned with the downstream localization task, and it treats all 520 AP measurements uniformly despite many being uninformative at any given location. This study proposes HiSCL-AE, a two-component architectural modification: (1) an AP Importance Gate (AIG) that learns a per-AP sigmoid attention weight to adaptively suppress irrelevant APs end-to-end; and (2) a hierarchical supervised contrastive pre-training strategy (HiSCL) that replaces the MSE reconstruction objective with the supervised contrastive loss applied simultaneously at the building, floor, and reference-point granularities, directly optimizing the embedding geometry for localization. Experiments on the public UJIIndoorLoc benchmark demonstrate that HiSCL-AE achieves 90.73% building-floor accuracy and a mean positioning error of 11.38 m, outperforming the SAE baseline on both primary metrics (90.37% accuracy and 11.48 m error). Beyond accuracy, the AIG produces an interpretable AP importance map, revealing that fewer than 60 of 520 APs account for the majority of the discriminative signal—a finding with direct implications for radio map compression and network planning. We compared HiSCL-AE with six additional Wi-Fi fingerprinting methods under an identical experimental protocol; HiSCL-AE simultaneously achieves the highest Building+Floor accuracy and the lowest positioning error among the six methods compared.