Easy Wins with Network Label Smoothing for Function Prediction

Idea: One smoothing step over a normalized PPI graph can yield consistent gains before you build a full GNN.

Method

Let P0 ∈ [0,1]^{NxC} be class probabilities from your sequence model and A the symmetrically normalized adjacency:

[ P_1 = (1 - \alpha) P_0 + \alpha \, \hat{A} P_0, \quad \alpha \in [0.1,0.3] ]

from scipy.sparse import csr_matrix
A = load_ppi_csr()                    # NxN
# symmetric normalization D^{-1/2} A D^{-1/2}
deg = np.array(A.sum(1))[:,0]; Dm12 = 1.0/np.sqrt(np.maximum(deg,1e-6))
A_norm = A.multiply(Dm12).T.multiply(Dm12).tocsr()

alpha = 0.2
P1 = (1 - alpha) * P0 + alpha * A_norm.dot(P0)
P1 = np.clip(P1, 0, 1)

Practical notes

• Hubs: cap degree or use personalized smoothing to reduce bias.
• Disconnected nodes: fallback to P0.
• Calibration: re-calibrate after smoothing.

Expected gains

Small but robust lifts in Fmax/auPRC, especially for mid-frequency terms. If no gains, inspect graph quality and degree distribution.

When to upgrade: if smoothing helps, consider GAT or edge-weighted GCN with confidence-aware edges.