papers/zoo-habitat-modeling/sections/14-supplementary-hyperparameter-sensitivity.tex

\section{Supplementary: Hyperparameter Sensitivity}

\label{app:hyperparameters}

\begin{table}[H]
\centering
\caption{Sensitivity of HabitatNet AUC to key hyperparameters.}
\label{tab:hyperparams}
\begin{tabular}{lcccc}
\toprule
\textbf{Parameter} & \textbf{Values tested} & \textbf{Best} & \textbf{AUC range} \\
\midrule
Message-passing layers $L$ & 2, 4, 6, 8, 10 & 6 & 0.931--0.943 \\
Hidden dimension & 64, 128, 256 & 128 & 0.937--0.943 \\
Learning rate & $10^{-3}$, $3\times10^{-4}$, $10^{-4}$ & $3\times10^{-4}$ & 0.935--0.943 \\
Spatial loss $\lambda_1$ & 0.01, 0.05, 0.1, 0.5 & 0.1 & 0.938--0.943 \\
Connectivity loss $\lambda_2$ & 0.01, 0.05, 0.1 & 0.05 & 0.940--0.943 \\
\bottomrule
\end{tabular}
\end{table}

The model is most sensitive to the number of message-passing layers and
least sensitive to the connectivity loss weight, suggesting that the
spatial structure provides the dominant signal.