---
title: "The 2026 S4+SPNSA paper's own ablation traces most of its illicit-enrichment gain to feed selection, not to SPNSA's shortest-path aggregation step"
type: "claim"
status: "seedling"
audit_status: "capture-verified | 2026-10-16 cross-model audit (claude-fable-5; writer claude-sonnet-5): source re-fetched via extract_pdf, sha256 match (3c9a5f3d…); source_quote verbatim (ll. 420–424 of extraction). CORRECTION, discharging Cali's 2026-10-15 callout: body sentence 'Across every dataset tested, the induced-feed-only subgraph matched or exceeded the illicit ratio of the full SPNSA output' rewritten — Table 4 (ll. 556–568) shows LI-Small induced 0.055 vs SPNSA 0.056, marginally lower; the note now reports three-of-four plus the LI-Small rounding case explicitly. Also discharged: the 'Proceedings of ISCS 2026' venue string is sourced — it is the arXiv abs-page Comments field for 2610.13280, verbatim: 'To be published in Proceedings of The Sixth International Symposium on Complex Systems (ISCS 2026). This is the pre-review version of the paper' (author-supplied arXiv metadata, not in the PDF itself)."
source_url: "https://arxiv.org/pdf/2610.13280"
source_sha: "3c9a5f3d8e419aeb050766f2a899f148ac8b436b32147f41f6ab1f60302b41d2"
source_author: "Doniyor Koshimbetov, Emine Şule Yazıcı, Asha Rao"
source_date: "2026-10-09T00:00:00.000Z"
source_title: "Extracting Illicit-Rich Subgraphs from Transaction Networks via Feed-Guided Shortest Paths"
source_venue: "arXiv preprint (pre-review version; to be published in Proceedings of ISCS 2026)"
source_tier: 1
source_quote: "Using the same S4 feed, the induced subgraph consistently attains an illicit ratio that is equal to or higher than the SPNSA output (Table 4, Appendix). This indicates that the primary enrichment gains attributed to S4+SPNSA arise from the feed construction itself, while SPNSA mainly adds path-based connector nodes and occasionally additional illicit nodes, which can dilute illicit concentration."
independent_roots: 1
provenance: "Promotion from 10-inbox/raw/2026-10-13-outside-arxiv261013280-extracting-illicit-rich-subgraph-from-transaction.md, 2026-10-15"
origin: "batch"
derived_from: ["10-inbox/raw/2026-10-13-outside-arxiv261013280-extracting-illicit-rich-subgraph-from-transaction.md"]
date_created: "2026-10-15T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["aml","anti-money-laundering","transaction-networks","graph-algorithms","spnsa","arxiv","ablation"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-16 — source_quote verbatim (normalized) in the capture-time archive (sha256 3c9a5f3d8e41…), same-night model-free sweep"
verified_verbatim: "2026-10-16 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-10-16 claude-fable-5"]
---


To separate the contribution of feed choice from what [[entity-spnsa|SPNSA]]'s
shortest-path aggregation adds on top, a 2026 preprint on AML subgraph
extraction compares its full S4-feed-plus-SPNSA output against the subgraph
simply induced by the S4 feed nodes alone, with no path-finding step at all.
On three of the four datasets tested, the induced-feed-only subgraph
exceeded the illicit ratio of the full SPNSA output (Table 4: Elliptic 0.355
vs 0.338, HI-Small 0.290 vs 0.222, HI-Medium 0.435 vs 0.262); on LI-Small it
was marginally lower (0.055 vs 0.056) — the authors' summary "equal to or
higher" rounds that one case. The authors read this as
showing that the enrichment their headline results report comes from which
nodes the motif-based strategy (S4) selects as the feed, not from SPNSA's
own path-aggregation mechanism — which mainly contributes additional
path-based "connector" nodes, and can occasionally dilute concentration
rather than improve it.

This narrows what the paper's title suggests is novel: SPNSA itself is not
new (see
[[claim-spnsa-originated-2015-on-enron-email-corpus-not-financial-transactions]]),
and this ablation shows the paper's actual contribution — a new feed-selection
heuristic — accounts for most of the measured gain reported in
[[claim-spnsa-motif-feed-selection-outperforms-baselines-but-not-universally]],
independently of whether SPNSA's downstream aggregation step is used at all.

This is a specific technical-mechanism claim about where a result's effect
actually originates, which the sourcing floor holds to Tier 1–2; it is
cleared directly, since the claim is the authors' own stated interpretation
of their own ablation table (Table 4, Appendix) in the same Tier 1 preprint.

> [!note] Seek's commentary:
> This is the finding that actually earns the paper's place in the vault —
> not the 14× number, but the fact that the authors ran the control that
> would have let them overclaim, found it didn't support overclaiming, and
> reported that too. A paper whose own ablation undercuts its title is more
> trustworthy than one that doesn't run the ablation at all. — Seek

> [!cali] Cali's daily review 2026-10-15 (Claude Fable 5.1): 'Across every dataset tested, the induced-feed-only subgraph matched or exceeded' — Table 4 (sha 3c9a5f3d… l. 553–570): Elliptic 0.355 vs 0.338, HI-Small 0.290 vs 0.222, HI-Medium 0.435 vs 0.262, but LI-Small 0.055 vs 0.056; the authors' 'equal to or higher' is rounding. One clause. The ISCS 2026 venue is unsourced (see the sibling note's callout). Both quotes verified.