qbiocode.apps.quvine.fusion.fuse_fixes module#

Summary#

Functions:

`calibrate_heat_kernel_fast`	Vectorized heat kernel calibration.
`calibrate_polynomial_filter_fast`	Vectorized polynomial filter calibration.
`fuse_embeddings_attention_fixed`	FIXED: Attention-weighted SVD fusion with information-content-aware similarity.
`fuse_embeddings_consensus_gated`	SVD fusion with per-node consensus gating.
`fuse_embeddings_uniform_mean`	Honest baseline: simple mean of views, then SVD compression.
`per_node_consensus_gate`	Compute per-node reliability score based on agreement with consensus.

fuse_embeddings_attention_fixed(Zs, k, temperature=1.0, mode='effective_rank')[source]#

FIXED: Attention-weighted SVD fusion with information-content-aware similarity.

The original implementation always produced uniform weights because similarity = ||Z_norm||_F^2 / n = 1 for any row-normalized Z.

fuse_embeddings_uniform_mean(Zs, k)[source]#: Honest baseline: simple mean of views, then SVD compression.

per_node_consensus_gate(Zs)[source]#: Compute per-node reliability score based on agreement with consensus.

fuse_embeddings_consensus_gated(Zs, k)[source]#: SVD fusion with per-node consensus gating.

calibrate_heat_kernel_fast(L, targets, node_to_idx, t_grid=None, loss='l2')[source]#: Vectorized heat kernel calibration. 3-4x faster than naive loop. Stacks delta vectors and calls expm_multiply once per t.

calibrate_polynomial_filter_fast(L, targets, node_to_idx, K=4, ridge=1e-05)[source]#

Vectorized polynomial filter calibration. 2-3x faster. Builds polynomial basis as block matrix multiplications.