Finally, given this g(·), we build a new f : Rd2 → Y to
locate the heterogeneous devices. Note that, finding a g(·)
to satisfy Eq.(1) is challenging – because of unknown yt, we
cannot align xt with xs, and thus directly learn a mapping
function between them as g like [4] does. Is it possible to
find a device-robust feature only from xs? Interestingly, our
answer is yes, as we show through the insights below.
Insight. Firstly, a pairwise feature that compares RSS values
from two APs is robust to device heterogeneity. We are motivated
by the observation that, RSS value is often inversely
proportional to the distance between AP and device; i.e., the
closer the device is from the AP, the larger the RSS value is.
Formally, this observation is characterized by a log-normal
shadowing model [6] in radio propagation theory. Denote
xk| as a RSS value at an arbitrary distance from APk,
then according to shadowing model,