GraphRSS-VGAE: Reconstruction of incomplete WiFi RSS fingerprints using variational graph autoencoders

Indoor positioning based on WiFi fingerprinting is highly sensitive to incomplete Received Signal Strength (RSS) measurements, which degrade positioning performance. This paper proposes GraphRSS-VGAE, a variational graph autoencoder (VGAE) framework for reconstructing missing RSS values using graph representations of reference point (RP) relationships. Two graph topologies are investigated within the proposed framework: GraphRSS-VGAE-RSS, where graph connectivity is determined by RSS fingerprint similarity, and GraphRSS-VGAE-Coord, where graph connectivity is derived from the spatial proximity of RPs. The proposed framework was evaluated through simulation in a synthetic indoor environment with missing fractions ranging from 10% to 80%, where both graph-based approaches were compared against the k-nearest neighbour (k-NN) interpolation baseline. Results show that GraphRSS-VGAE-RSS achieves performance comparable to k-NN only at the lowest missing fraction and lower reconstruction error once missingness increases, suggesting that the graph-based latent representation provides greater robustness. GraphRSS-VGAE-Coord consistently achieves the lowest reconstruction error, highlighting the value of reliable spatial priors. Overall, the results demonstrate the potential of graph-based representation learning for reconstructing incomplete Wi-Fi RSS fingerprints and provide improved reconstruction performance over the k-NN baseline under higher levels of data sparsity.