Hyperparameterless k-NN for Wi-Fi Fingerprinting

Fingerprint-based solutions mostly rely on variants of the
$k$
-NN algorithm. Despite the core operation of the model being quite simple, several implementation details can be exploited to enhance positioning accuracy. Recent works have been focusing on dynamically setting the value of
$k$
as, depending on the location, the algorithm may require different levels of reference data to provide a good position estimate. This paper explores two alternatives to settle the best configuration for
$k$
-NN (value of
$k$
and distance metric) depending on the operational fingerprint that is being processed. In contrast to other dynamic models published in the literature, the proposed Dynamic Configuration Hyperparameterless k-NN (DCHPL k-NN) does not introduce any new hyperparameter to set, neither at database nor sample levels. The results indicate that selecting the best configuration based on the dominant AP of the operational fingerprint can reduce the error from 6.59 m to 6.22 m, representing a relative reduction of approximately 5 %.