basin-retrieval / code /payload_graph /bench_payload_graph_refinement.py
Bakobiibizo's picture
Basin retrieval paper: code, data, results, paper
d1cfb66 verified
Raw
History Blame Contribute Delete
15 kB
"""Payload graph projection refinement.
Question:
Can semantic/content connections stored inside a keyed recurrence motif be
projected as a payload graph and used to refine a broad retrieved bundle?
Architecture under test:
label-free recurrence motif key
-> broad bundle retrieval
-> project semantic/content payload graph stored inside each motif
-> refine using target payload graph topology/relations
Memory content: LDGR historical findings (same source as behavioral relevance).
Important control:
For each real memory item, create a rewired decoy with the SAME payload nodes
and SAME relation-label multiset, but different edges. Node-bag/token overlap
and relation-topology-only cannot distinguish original from decoy. Exact
content-graph connections can. This directly tests whether the semantic
connections inside the motif are doing work beyond flat overlap.
"""
from __future__ import annotations
import json
import os
import random
import statistics
import sys
from collections import Counter, defaultdict
from dataclasses import dataclass
from pathlib import Path
_HERE = Path(__file__).resolve().parent
_CORE = _HERE.parent / "core"
_BEHAV = _HERE.parent / "behavioral_relevance"
for _p in (str(_CORE), str(_BEHAV)):
if _p not in sys.path:
sys.path.insert(0, _p)
from bench_behavioral_relevance import MemoryItem, build_corpus
HERE = Path(__file__).resolve().parent
OUT_DIR = HERE
SCORERS = ("coarse_only", "node_bag", "relation_topology", "content_graph")
QUERY_MODES = ("core", "partial", "noisy", "topic_mechanism")
@dataclass(frozen=True)
class PayloadGraph:
family: str
source: str
is_decoy: bool
nodes: tuple[str, ...]
edges: tuple[tuple[str, str, str], ...] # src, relation, dst
# ---------------------------------------------------------------------------
# graph projection from motif payload data
# ---------------------------------------------------------------------------
def first_node(item: MemoryItem, prefix: str) -> str:
for node, _et, _dir in item.walk:
if str(node).startswith(prefix):
return node
return f"{prefix}absent"
def all_nodes(item: MemoryItem, prefix: str) -> list[str]:
out, seen = [], set()
for node, _et, _dir in item.walk:
if str(node).startswith(prefix) and node not in seen:
out.append(node); seen.add(node)
return out
def payload_graph(item: MemoryItem) -> PayloadGraph:
"""Project semantic/content connections stored inside the motif payload."""
topic = first_node(item, "topic::")
mechanism = first_node(item, "mechanism::")
metric = first_node(item, "metric::")
kind = first_node(item, "kind::")
source = first_node(item, "source::")
boundary = first_node(item, "boundary::")
terms = all_nodes(item, "term::")
while len(terms) < 4:
terms.append(f"term::absent_{len(terms)}")
edges = [
(topic, "USES_MECHANISM", mechanism),
(mechanism, "HAS_METRIC", metric),
(topic, "HAS_BOUNDARY", boundary),
(boundary, "EVIDENCED_BY_SOURCE_KIND", kind),
(source, "SUPPORTS_BOUNDARY", boundary),
(topic, "MENTIONS_TERM", terms[0]),
(topic, "MENTIONS_TERM", terms[1]),
(mechanism, "SUPPORTED_BY_TERM", terms[0]),
(mechanism, "SUPPORTED_BY_TERM", terms[2]),
(terms[0], "CO_OCCURS_WITH", terms[1]),
(terms[1], "CO_OCCURS_WITH", terms[2]),
(terms[2], "CO_OCCURS_WITH", terms[3]),
(metric, "QUALIFIES_TERM", terms[3]),
(kind, "CONTAINS_TERM", terms[0]),
]
nodes = tuple(sorted({n for e in edges for n in (e[0], e[2])}))
return PayloadGraph(item.family, item.source, False, nodes, tuple(edges))
def rewire_decoy(g: PayloadGraph, rng: random.Random) -> PayloadGraph:
"""Same nodes + same relation multiset, different connections."""
nodes = list(g.nodes)
edges = []
for i, (_src, rel, _dst) in enumerate(g.edges):
# deterministic-ish rewire: rotate endpoints by different offsets.
src = nodes[(i + 3) % len(nodes)]
dst = nodes[(i * 2 + 5) % len(nodes)]
if src == dst:
dst = nodes[(i * 2 + 6) % len(nodes)]
edges.append((src, rel, dst))
# If by chance any edge survived, perturb it.
original = set(g.edges)
fixed = []
for i, e in enumerate(edges):
if e in original:
src, rel, dst = e
dst = nodes[(nodes.index(dst) + 1) % len(nodes)]
if src == dst:
dst = nodes[(nodes.index(dst) + 1) % len(nodes)]
e = (src, rel, dst)
fixed.append(e)
return PayloadGraph(f"decoy_{g.family}", f"decoy_of:{g.source}", True, g.nodes, tuple(fixed))
# ---------------------------------------------------------------------------
# query projection
# ---------------------------------------------------------------------------
def query_edges(g: PayloadGraph, mode: str, rng: random.Random) -> tuple[tuple[str, str, str], ...]:
if mode == "core":
keep_rels = {"USES_MECHANISM", "HAS_METRIC", "MENTIONS_TERM", "SUPPORTED_BY_TERM"}
return tuple(e for e in g.edges if e[1] in keep_rels)[:6]
if mode == "partial":
return tuple(rng.sample(list(g.edges), k=min(6, len(g.edges))))
if mode == "noisy":
base = list(rng.sample(list(g.edges), k=min(6, len(g.edges))))
base.append((f"foreign::{rng.randint(0,9999)}", "FOREIGN_RELATION", f"foreign::{rng.randint(0,9999)}"))
return tuple(base)
if mode == "topic_mechanism":
keep_rels = {"USES_MECHANISM", "HAS_BOUNDARY"}
return tuple(e for e in g.edges if e[1] in keep_rels)
raise ValueError(mode)
def nodes_of(edges):
return {n for e in edges for n in (e[0], e[2])}
# ---------------------------------------------------------------------------
# scoring
# ---------------------------------------------------------------------------
def containment(query_counter: Counter, cand_counter: Counter) -> float:
denom = sum(query_counter.values())
if denom == 0:
return 1.0
hit = sum(min(cand_counter[k], query_counter[k]) for k in query_counter)
return hit / denom
def score_candidate(candidate: PayloadGraph, q_edges, scorer: str) -> float:
if scorer == "coarse_only":
return 1.0
if scorer == "node_bag":
q = Counter(nodes_of(q_edges))
c = Counter(candidate.nodes)
return containment(q, c)
if scorer == "relation_topology":
q = Counter(e[1] for e in q_edges)
c = Counter(e[1] for e in candidate.edges)
return containment(q, c)
if scorer == "content_graph":
q = Counter(q_edges)
c = Counter(candidate.edges)
return containment(q, c)
raise ValueError(scorer)
def top_bundle(scored, frac=0.9):
if not scored:
return []
top = scored[0]["score"]
threshold = top * frac
return [r for r in scored if r["score"] >= threshold] or scored[:1]
def refine(candidates: list[PayloadGraph], target_family: str, q_edges, scorer: str) -> dict:
scored = []
for cand in candidates:
scored.append({"family": cand.family, "source": cand.source, "is_decoy": cand.is_decoy,
"score": round(score_candidate(cand, q_edges, scorer), 6)})
scored.sort(key=lambda r: r["score"], reverse=True)
top = scored[0]["score"] if scored else None
top_rows = [r for r in scored if top is not None and abs(r["score"] - top) < 1e-9]
top_families = {r["family"] for r in top_rows}
dominant = next(iter(top_families)) if len(top_families) == 1 else None
bundle = top_bundle(scored)
target_rank = None
for i, r in enumerate(scored, start=1):
if r["family"] == target_family:
target_rank = i
break
return {
"scorer": scorer,
"dominant_family": dominant,
"correct_top1": dominant == target_family,
"target_in_top_tie": any(r["family"] == target_family for r in top_rows),
"target_rank": target_rank,
"top_score": top,
"top_tie_size": len(top_rows),
"bundle_size": len(bundle),
"bundle_reduction": round(len(candidates) / len(bundle), 4) if bundle else 0,
"top_decoy_rate": sum(r["is_decoy"] for r in top_rows) / len(top_rows) if top_rows else 0.0,
"ranked_top5": scored[:5],
}
# ---------------------------------------------------------------------------
# experiment
# ---------------------------------------------------------------------------
def run_pool(real_graphs: list[PayloadGraph], candidates: list[PayloadGraph], pool_name: str, seed=20260706):
rng = random.Random(seed)
rows = []
for g in real_graphs:
for mode in QUERY_MODES:
q = query_edges(g, mode, rng)
for scorer in SCORERS:
r = refine(candidates, g.family, q, scorer)
rows.append({"pool": pool_name, "family": g.family, "source": g.source,
"mode": mode, "n_query_edges": len(q), **r})
return rows
def summarize(rows):
out = {}
for pool in sorted({r["pool"] for r in rows}):
out[pool] = {}
for mode in QUERY_MODES:
out[pool][mode] = {}
for scorer in SCORERS:
xs = [r for r in rows if r["pool"] == pool and r["mode"] == mode and r["scorer"] == scorer]
out[pool][mode][scorer] = {
"n": len(xs),
"top1_accuracy": mean([x["correct_top1"] for x in xs]),
"target_in_top_tie": mean([x["target_in_top_tie"] for x in xs]),
"mean_target_rank": round(statistics.fmean([x["target_rank"] for x in xs]), 4),
"mean_top_tie_size": round(statistics.fmean([x["top_tie_size"] for x in xs]), 4),
"mean_bundle_size": round(statistics.fmean([x["bundle_size"] for x in xs]), 4),
"mean_bundle_reduction": round(statistics.fmean([x["bundle_reduction"] for x in xs]), 4),
"mean_top_decoy_rate": round(statistics.fmean([x["top_decoy_rate"] for x in xs]), 4),
}
# aggregate by scorer across modes
out["aggregate"] = {}
for pool in sorted({r["pool"] for r in rows}):
out["aggregate"][pool] = {}
for scorer in SCORERS:
xs = [r for r in rows if r["pool"] == pool and r["scorer"] == scorer]
out["aggregate"][pool][scorer] = {
"top1_accuracy": mean([x["correct_top1"] for x in xs]),
"target_in_top_tie": mean([x["target_in_top_tie"] for x in xs]),
"mean_target_rank": round(statistics.fmean([x["target_rank"] for x in xs]), 4),
"mean_top_tie_size": round(statistics.fmean([x["top_tie_size"] for x in xs]), 4),
"mean_bundle_size": round(statistics.fmean([x["bundle_size"] for x in xs]), 4),
"mean_bundle_reduction": round(statistics.fmean([x["bundle_reduction"] for x in xs]), 4),
}
return out
def mean(xs):
return round(statistics.fmean([float(x) for x in xs]), 4) if xs else None
def render_report(summary, rows, n_real, n_decoy):
lines = [
"# Payload Graph Projection Refinement",
"",
"Memory content: LDGR historical findings. Coarse key: generic label-free recurrence motif. Payload: semantic/content graph connecting data inside the motif.",
"",
f"Real payload graphs: {n_real}",
f"Rewired decoys: {n_decoy}",
"",
"## Aggregate Results",
"",
"| pool | scorer | top1 | target in top tie | target rank | top tie | bundle | reduction |",
"|---|---|---:|---:|---:|---:|---:|---:|",
]
for pool, scorers in summary["aggregate"].items():
for scorer, s in scorers.items():
lines.append(f"| {pool} | {scorer} | {s['top1_accuracy']} | {s['target_in_top_tie']} | "
f"{s['mean_target_rank']} | {s['mean_top_tie_size']} | {s['mean_bundle_size']} | {s['mean_bundle_reduction']} |")
lines += ["", "## Decoy Pool By Query Mode", ""]
for mode in QUERY_MODES:
lines += [f"### {mode}", "", "| scorer | top1 | target in top tie | top tie | bundle | reduction | decoy rate |",
"|---|---:|---:|---:|---:|---:|---:|"]
for scorer in SCORERS:
s = summary["real_plus_rewired_decoys"][mode][scorer]
lines.append(f"| {scorer} | {s['top1_accuracy']} | {s['target_in_top_tie']} | "
f"{s['mean_top_tie_size']} | {s['mean_bundle_size']} | {s['mean_bundle_reduction']} | {s['mean_top_decoy_rate']} |")
lines.append("")
lines += [
"## Interpretation",
"",
"The rewired-decoy pool is the decisive control. Each decoy has the same payload nodes and same relation-label multiset as its paired real item, but different semantic connections. If node-bag and relation-topology tie while content_graph isolates the original, then the stored payload graph connections are doing refinement work beyond flat overlap.",
]
return "\n".join(lines) + "\n"
def main():
print("building LDGR historical payload graphs...")
items = build_corpus(max_items=32, per_topic_cap=8)
real = [payload_graph(it) for it in items]
rng = random.Random(20260706)
decoys = [rewire_decoy(g, rng) for g in real]
rows = []
rows.extend(run_pool(real, real, "real_only"))
rows.extend(run_pool(real, real + decoys, "real_plus_rewired_decoys"))
summ = summarize(rows)
out = {"experiment": "payload-graph-refinement-profile",
"branch": "payload-graph-refinement",
"architecture": "label-free recurrence motif -> projected semantic payload graph -> refinement",
"n_real": len(real), "n_decoys": len(decoys),
"scorers": SCORERS, "query_modes": QUERY_MODES,
"summary": summ, "rows": rows,
"graphs": {"real": [g.__dict__ for g in real], "decoys": [g.__dict__ for g in decoys]}}
json_path = OUT_DIR / "payload_graph_refinement_results.json"
json_path.write_text(json.dumps(out, indent=2))
md_path = OUT_DIR / "payload_graph_refinement_report.md"
md_path.write_text(render_report(summ, rows, len(real), len(decoys)))
print("\nPAYLOAD GRAPH PROJECTION REFINEMENT")
print("=" * 78)
print(f"real graphs={len(real)} decoys={len(decoys)}")
print("\naggregate:")
for pool, scorers in summ["aggregate"].items():
print(f" {pool}")
for scorer, s in scorers.items():
print(f" {scorer:<18} top1={s['top1_accuracy']:<6} tieIncl={s['target_in_top_tie']:<6} "
f"rank={s['mean_target_rank']:<7} bundle={s['mean_bundle_size']:<7} red={s['mean_bundle_reduction']}")
print(f"\nwrote {json_path.name}")
print(f"wrote {md_path.name}")
if __name__ == "__main__":
main()