In this paper, graph attention network (GAT) is firstly utilized for the
channel estimation. In accordance with the 6G expectations, we consider a
high-altitude platform station (HAPS) mounted reconfigurable intelligent
surface-assisted two-way communications and obtain a low overhead and a high
normalized mean square error performance. The performance of the proposed
method is investigated on the two-way backhauling link over the RIS-integrated
HAPS. The simulation results denote that the GAT estimator overperforms the
least square in full-duplex channel estimation. Contrary to the previously
introduced methods, GAT at one of the nodes can separately estimate the
cascaded channel coefficients. Thus, there is no need to use time-division
duplex mode during pilot signaling in full-duplex communication. Moreover, it
is shown that the GAT estimator is robust to hardware imperfections and changes
in small-scale fading characteristics even if the training data do not include
all these variations.

本文利用图注意力网络（GAT）进行信道估计，在高空平台站（HAPS）上使用可重构智能表面协助的双向通信，获得低开销和高规范化均方误差性能，并证明 GAT 估算器在全双工信道估计中优于最小二乘估计。