Answers are grounded in this paper and cite their sources.
Wi-Fi fingerprinting is usually scored on the phones that built the radio map and with every access point alive.We state four hypotheses about what happens otherwise and test them on a public multi-building dataset, holding out one phone at a time and then scoring a campaign recorded months later.A nearest-neighbour locator is run on five fingerprint representations; a distance that discounts access points present in the map but absent from the query is introduced; and three ways of attaching an error radius to each fix are compared.Holding out a phone multiplies the mean error by more than four, most of it owed to unseen locations, while the loss in floor identification is owed mainly to the device.A rank-based fingerprint improves floor identification for twelve of sixteen phones and lowers the later-campaign error by a fifth, at a price when the phone is known.Offset correction does not help.Discounting absence keeps the error within a quarter of its clean value when three access points in ten are silent, where the standard distance more than doubles it.Neighbour spread alone is a poor guide to the error; a four-cue quantile model is better.
No takes yet. Share an insight, caveat, or question.
Saddik et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: