basin-retrieval / code /topology /control_generator.py
Bakobiibizo's picture
Basin retrieval paper: code, data, results, paper
d1cfb66 verified
Raw
History Blame Contribute Delete
4.88 kB
"""Automorphism-aware control generation for temporal topology experiments.
Controls are generated only after canonicalization. Candidate negatives that
canonicalize to the target motif are rejected as automorphic/topologically
equivalent rather than counted as false activations.
"""
from __future__ import annotations
import random
from collections import Counter
from collections.abc import Callable, Sequence
from dataclasses import dataclass
from recurrence_topology_experiment import TemporalSignature, canonical_signature
@dataclass(frozen=True)
class GeneratedControl:
name: str
sequence: tuple[str, ...]
signature: TemporalSignature
attempts: int
def simple_degree_profile(signature: TemporalSignature) -> tuple[tuple[int, int], ...]:
in_degree = Counter[int]()
out_degree = Counter[int]()
for source, target in signature.edges:
out_degree[source] += 1
in_degree[target] += 1
return tuple(
sorted((in_degree[node], out_degree[node]) for node in range(signature.node_count))
)
def weighted_degree_profile(signature: TemporalSignature) -> tuple[tuple[int, int], ...]:
in_degree = Counter[int]()
out_degree = Counter[int]()
for source, target, count in signature.edge_counts:
out_degree[source] += count
in_degree[target] += count
return tuple(
sorted((in_degree[node], out_degree[node]) for node in range(signature.node_count))
)
def is_automorphic_equivalent(candidate: Sequence[str], motif: TemporalSignature) -> bool:
"""Return true when canonicalization makes the candidate the same motif."""
return canonical_signature(candidate).key() == motif.key()
def is_degree_matched_non_isomorphic(candidate: Sequence[str], motif: TemporalSignature) -> bool:
"""Check length, node count, degree profile, and non-isomorphism after canonicalization."""
signature = canonical_signature(candidate)
return (
len(signature.trace) == len(motif.trace)
and signature.node_count == motif.node_count
and signature.key() != motif.key()
and simple_degree_profile(signature) == simple_degree_profile(motif)
and weighted_degree_profile(signature) == weighted_degree_profile(motif)
)
def _trace_to_symbols(trace: Sequence[int], prefix: str) -> tuple[str, ...]:
return tuple(f"{prefix}_{node}" for node in trace)
def _all_nodes_used(trace: Sequence[int], node_count: int) -> bool:
return set(trace) == set(range(node_count))
def sample_degree_matched_controls(
motif: TemporalSignature,
*,
family: str,
count: int,
seed: int,
max_attempts: int = 200_000,
) -> tuple[GeneratedControl, ...]:
"""Sample canonicalized, degree-matched, non-isomorphic controls.
The generator deliberately checks every candidate after canonicalization. This
prevents automorphic equivalents from entering the negative set.
"""
rng = random.Random(seed)
accepted: list[GeneratedControl] = []
seen: set[str] = set()
attempts = 0
length = len(motif.trace)
node_count = motif.node_count
while len(accepted) < count and attempts < max_attempts:
attempts += 1
raw_trace = tuple(rng.randrange(node_count) for _ in range(length))
if not _all_nodes_used(raw_trace, node_count):
continue
candidate_sequence = _trace_to_symbols(raw_trace, f"G_{family}_{len(accepted)}")
signature = canonical_signature(candidate_sequence)
signature_key = signature.key()
if signature_key in seen or signature_key == motif.key():
continue
if not is_degree_matched_non_isomorphic(candidate_sequence, motif):
continue
seen.add(signature_key)
accepted.append(
GeneratedControl(
name=f"{family}_degree_matched_non_isomorphic_{len(accepted):03d}",
sequence=candidate_sequence,
signature=signature,
attempts=attempts,
)
)
if len(accepted) < count:
msg = (
f"generated {len(accepted)} of {count} controls for {family} after {attempts} attempts"
)
raise RuntimeError(msg)
return tuple(accepted)
def enumerate_automorphic_equivalent(
motif_sequence: Sequence[str],
*,
name: str,
transform: Callable[[tuple[str, ...]], tuple[str, ...]],
) -> GeneratedControl:
"""Create a known automorphic equivalent and validate it canonicalizes to the motif."""
motif = canonical_signature(motif_sequence)
candidate = transform(tuple(motif_sequence))
signature = canonical_signature(candidate)
if signature.key() != motif.key():
raise ValueError("candidate is not automorphic/topologically equivalent")
return GeneratedControl(name=name, sequence=candidate, signature=signature, attempts=1)