Chebyshev Graph Convolution

The problem, from first principles

Chebyshev Graph Convolution uses Chebyshev polynomials of the graph Laplacian to define convolutional filters. This avoids eigendecomposition while being expressive.

Where T_k is the k-th Chebyshev polynomial and tilde(Lambda) is normalized Laplacian eigenvalues.

From theory to code

Implement:

chebyshev_conv(features, adj, weight, k=3)

Applies k-order Chebyshev convolution.

Constraints

• features shape: (num_nodes, input_dim).
• adj shape: (num_nodes, num_nodes).
• weight shape: (k+1, input_dim, output_dim).
• Return shape: (num_nodes, output_dim).