netfm/data/structural_features.py

import torch
import networkx as nx
import numpy as np
from torch_geometric.data import Data
from torch_geometric.utils import to_networkx, degree as pyg_degree


LARGE_GRAPH_THRESHOLD = 50_000


def _compute_degree(data: Data) -> np.ndarray:
    """Compute node degree directly from edge_index (fast, no NetworkX)."""
    n = data.num_nodes
    deg = pyg_degree(data.edge_index[0], num_nodes=n) + pyg_degree(data.edge_index[1], num_nodes=n)
    return deg.numpy().astype(np.float32) / 2


def compute_structural_features(data: Data) -> torch.Tensor:
    """Compute 6 domain-agnostic structural node features for a PyG Data object."""
    n = data.num_nodes
    is_large = n > LARGE_GRAPH_THRESHOLD

    degree = _compute_degree(data)

    G = to_networkx(data, to_undirected=True)

    clustering_dict = nx.clustering(G)
    clustering = np.array([clustering_dict.get(i, 0.0) for i in range(n)], dtype=np.float32)

    pagerank_dict = nx.pagerank(G, max_iter=100, tol=1e-04)
    pagerank = np.array([pagerank_dict.get(i, 0.0) for i in range(n)], dtype=np.float32)

    triangles_dict = nx.triangles(G)
    triangles = np.array([triangles_dict.get(i, 0) for i in range(n)], dtype=np.float32)

    core_dict = nx.core_number(G)
    core_number = np.array([core_dict.get(i, 0) for i in range(n)], dtype=np.float32)

    if is_large:
        eigenvector = np.zeros(n, dtype=np.float32)
    else:
        try:
            eig_dict = nx.eigenvector_centrality(G, max_iter=300, tol=1e-06)
            eigenvector = np.array([eig_dict.get(i, 0.0) for i in range(n)], dtype=np.float32)
        except (nx.PowerIterationFailedConvergence, nx.NetworkXError):
            eigenvector = np.zeros(n, dtype=np.float32)

    features = np.stack(
        [degree, clustering, pagerank, triangles, core_number, eigenvector], axis=1
    )
    return torch.from_numpy(features)


def normalize_features(features: torch.Tensor) -> torch.Tensor:
    """Z-score normalize features column-wise."""
    mean = features.mean(dim=0, keepdim=True)
    std = features.std(dim=0, keepdim=True).clamp(min=1e-8)
    return (features - mean) / std