Hyperbolic Linear Layer Parameters: manifold (Manifold): The manifold to use for the linear transformation. in_features (int

Parameters: manifold (manifold): The manifold to use for the linear transformation. in_features (int): The size of each input sam

Hyperbolic Linear Layer Parameters: manifold (Manifold): The manifold to use for the linear transformation. in_features (int

Our MixHop layer

Our MixHop layer

Hyperbolic Layer Normalization Layer Parameters: manifold (Manifold): The manifold to use for normalization. in_features (in

Hyperbolic Layer Normalization Layer Parameters: manifold (Manifold): The manifold to use for normalization. in_features (in

Adapted from https://github.com/snap-stanford/ogb/

Optimizer for Euclidean and Hyperbolic parameters Parameters: model (nn.Module): The model containing the parameters to optimize.

GPRGNN, from original repo https://github.com/jianhao2016/GPRGNN propagation class for GPR_GNN

GPRGNN, from original repo https://github.com/jianhao2016/GPRGNN propagation class for GPR_GNN

Hyperbolic Activation Layer Parameters: manifold (Manifold): The manifold to use for the activation. activation (function):

Hyperbolic Activation Layer Parameters: manifold (Manifold): The manifold to use for the activation. activation (function):

Hyperbolic Dropout Layer Parameters: manifold (Manifold): The manifold to use for the dropout. dropout (float): The dropout

Hyperbolic Dropout Layer Parameters: manifold (Manifold): The manifold to use for the dropout. dropout (float): The dropout

Numerically stable arccosh that never returns NaNs. Returns acosh(x) = arccosh(x) = log(x + sqrt(max(x^2 - 1, eps))).

Numerically stable arccosh that never returns NaNs. Returns acosh(x) = arccosh(x) = log(x + sqrt(max(x^2 - 1, eps))).

Numerically stable arccosh that never returns NaNs. Returns acosh(x) = arccosh(x) = log(x + sqrt(max(x^2 - 1, eps))).

Numerically stable arctanh that never returns NaNs. x = clamp(x, min=-1+eps, max=1-eps) Returns atanh(x) = arctanh(x) = 0.5*(log(1+x)-log

Numerically stable arctanh that never returns NaNs. x = clamp(x, min=-1+eps, max=1-eps) Returns atanh(x) = arctanh(x) = 0.5*(log(1+x)-log

Numerically stable arctanh that never returns NaNs. x = clamp(x, min=-1+eps, max=1-eps) Returns atanh(x) = arctanh(x) = 0.5*(log(1+x)-log

GPRGNN, from original repo https://github.com/jianhao2016/GPRGNN

GPRGNN, from original repo https://github.com/jianhao2016/GPRGNN

Neighborhood aggregation step

Neighborhood aggregation step

Parameters: manifold (manifold): The manifold to use for the linear transformation. in_features (int): The size of each input sam

Hyperbolic Normalization Layer Parameters: manifold (Manifold): The manifold to use for normalization. manifold_out (Manifol

Hyperbolic Normalization Layer Parameters: manifold (Manifold): The manifold to use for normalization. manifold_out (Manifol

logistic regression on adjacency matrix

logistic regression on adjacency matrix

adapted from https://github.com/CUAI/CorrectAndSmooth/blob/master/gen_models.py

adapted from https://github.com/CUAI/CorrectAndSmooth/blob/master/gen_models.py

our implementation of MixHop some assumptions: the powers of the adjacency are [0, 1, ..., hops], with every power in between each co

our implementation of MixHop some assumptions: the powers of the adjacency are [0, 1, ..., hops], with every power in between each co

label propagation, with possibly multiple hops of the adjacency

label propagation, with possibly multiple hops of the adjacency

Optimizer for Euclidean and Hyperbolic parameters Parameters: model (nn.Module): The model containing the parameters to optimize.

Use MLP instead of a single linear layer.

Use MLP instead of a single linear layer.