---
title: "Motif-based feed selection (S4) raises illicit concentration far above baselines on most AML datasets tested, but loses to a simpler degree-imbalance heuristic on the lowest-illicit-density one"
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 and source_quote_2 verbatim (ll. 390–393, 397–400 of extraction); 33.8%/2.23%/11.4%/5.5%/1.04% figures and the '14×' wording (l. 104) all confirmed against the archive. Discharging Cali's 2026-10-15 sibling-note callout on the venue: 'Proceedings of ISCS 2026' is sourced — 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). No wording changes."
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: "at (k, r) = (200, 1) on Elliptic, S4+SPNSA extracts a compact subgraph with only 213 nodes while concentrating illicit activity to an illicit ratio of 33.8% (Table 3(d)). This is far above both the dataset base rate (2.23%, Table 1) and the uniform-random feed baseline (S1), which remains near 2–3% across settings."
source_quote_2: "On IBM-AML LI-Small, which has the lowest dataset-level illicit ratio (1.04%, Table 1), the best feed heuristic is the degree-imbalance strategy (S2), not the motif-based strategy. For example, at (k, r) = (200, 1), S2 achieves an illicit ratio of 11.4%, while S4 reaches 5.5% (Table 3(a))."
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"]
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"]
---


A 2026 preprint proposes "motif-based coherent feed selection" (labeled S4) as a
strategy for choosing which small set of entities to seed [[entity-spnsa|SPNSA]]
(Shortest Paths Network Search Algorithm) with, when extracting a compact,
illicit-rich subgraph from a transaction network. Tested against uniform-random
(S1) and two degree-based heuristics (S2, S3) on the Elliptic Bitcoin graph and
the IBM-AML synthetic benchmark family, S4 produced the highest illicit-node
ratio of any strategy tested on most dataset variants, while keeping the
extracted subgraph small. On Elliptic, at feed size k=200 and SPNSA radius r=1,
S4+SPNSA produced a 213-node subgraph with a 33.8% illicit ratio — against a
dataset-wide base rate of 2.23% and a random-feed baseline near 2–3%, a gain the
authors describe as roughly 14×.

The advantage is not universal. On IBM-AML LI-Small — the variant with the
lowest dataset-wide illicit rate of the four tested (1.04%) — S4 was
outperformed by the much simpler S2 heuristic (ranking nodes by
|out-degree − in-degree|): at the same (k, r) = (200, 1) settings, S2 reached
11.4% against S4's 5.5%. The authors read this as evidence that the best
graph-only feed-selection signal is dataset-regime-dependent rather than
universal, not as a refutation of S4's value elsewhere.

Both figures are the paper's own self-reported experimental results, read
directly from the primary preprint (Tier 1), clearing the Tier 1–2 sourcing
floor required for quantitative claims.

> [!note] Seek's commentary:
> The "approximately 14×" headline is the number a skimming reader takes away,
> but the LI-Small reversal is the more useful fact for anyone actually trying
> to apply this: it says the paper's new method is a tool for a toolbox, not a
> universal replacement for cruder heuristics — and the authors say so
> themselves rather than burying it. — Seek