Public abuse portals such as bitcoinabuse.com let victims report a Bitcoin address and pick a scam category ("ransomware", "blackmail scam", "sextortion", "darknet market", and so on), usually with a pasted ransom note or email. Many addresses collect dozens of reports, and those reports often disagree on the category. Forensic models treat the resulting tags as ground truth, and the one large study of this corpus found that categories like "blackmail scam" act as catch-alls because victims confuse extortion, malware and ransomware (A Large-Scale Bitcoin Abuse Measurement and Clustering Analysis Utilizing Public Reports). We propose to treat each victim report as an annotation of the address and to model the disagreement directly. The goal is to separate two very different causes. The first is confusion: victims who received the same ransom-note template label it inconsistently. The second is genuine multi-use: one operator's address serves several distinct campaigns.
The key idea is that the report text itself reveals which reports belong to the same campaign. A heterogeneous document graph is a graph whose nodes are pieces of documents at different granularities (reports, sentences, extracted entities such as contact emails, onion URLs, demanded amounts, other wallet addresses), linked by typed edges. Built over the report corpus, it tells us which victims were hit by the same lure, even when the lure was paraphrased or the payment addresses are unconnected on-chain. Those campaign groups are exactly what the best disagreement model is missing. The result is a cleaner, calibrated category distribution per address, which serves as a training target for on-chain classifiers. The recovered campaigns also link pseudonymous addresses that transaction-graph clustering cannot connect.
Core Claim
Disagreement among victims reporting the same address is structured, and most of the structure is explained by campaign membership that can be inferred from the reports' text and shared entities. A disagreement model that infers these groups from the document graph should do three things better than majority vote, raw vote distributions, standard consensus models, or text clustering without the graph. It should predict the category mix of future reports on an address. It should agree better with expert adjudication on whether an address is confused or truly multi-use. It should also yield better training targets for classifiers that label never-reported addresses from transaction data alone.
Where This Sits
Three lines of prior work motivate the proposal, and each leaves a specific hole. Blockchain account classifiers learn from transaction structure. DIAM models Bitcoin and Ethereum as directed multigraphs and propagates feature differences between an account and its neighbors because illicit accounts connect mostly to benign ones (Effective Illicit Account Detection on Large Cryptocurrency MultiGraphs). Ethident classifies Ethereum accounts from sampled interaction subgraphs with contrastive pretraining to cope with scarce labels (Behavior-aware Account De-anonymization on Ethereum Interaction Graph). Both take crawled explorer tags as ground truth, and Ethident explicitly flags mislabeled accounts as a distortion. We do not propose a better transaction encoder. We change the labels these encoders are trained on.
The second line is the constraint. StableAML shows that on shattered, very sparse stablecoin graphs (density below 0.01), CatBoost over 68 handcrafted features reaches 0.9775 Macro-F1 while GraphSAGE reaches 0.8048 (StableAML). We take this seriously: our downstream on-chain classifier will be a tree ensemble unless a graph model beats it. Our graph lives on the text side, where shared emails and templates provide the connectivity that the transaction graph lacks.
The third line supplies the method. NUTMEG is a Bayesian model that estimates a separate truth per annotator subpopulation plus per-annotator competence. It separates systematic group disagreement from spam and recovers true competence at 0.81 correlation versus 0.58 for MACE in synthetic trials (NUTMEG). It requires known group metadata such as demographics, and it lists inferring groups without metadata as open. Victims have no demographics, but they have campaigns, and campaigns are visible in text. We supply NUTMEG's missing groups from a document graph built the way HeterSumGraph uses shared words as bridges between sentence nodes (Heterogeneous Graph Neural Networks for Extractive Document Summarization). It also follows the typed mention–entity–sentence construction of dual-tier relation-extraction graphs (Document-level Relation Extraction with Dual-tier Heterogeneous Graph). We then use the resulting distributions as soft targets, following the multi-task soft-label recipe of Beyond Black & White. That recipe feeds raw vote distributions to an auxiliary head, where we feed denoised, campaign-aware ones.
Choi et al. is the closest existing work. It joins the same report corpus to on-chain clustering, but it uses flat TF-IDF keyword tables and names label ambiguity as an unaddressed limitation. It builds no model of disagreement and no graph linking report content to addresses.
Research Program
We build a graph over the Bitcoin Abuse corpus (200,507 reports on roughly 58,000 addresses per Choi et al.). It has report nodes, sentence nodes, and entity nodes for extracted contact emails, URLs, onion domains, demanded amounts and quoted wallet addresses. It also has address nodes and on-chain cluster nodes from the multi-input spending heuristic. We avoid change-address merging because Möser shows it causes erroneous cluster collapse. Edges encode report-about-address, report-contains-sentence, sentence-mentions-entity, near-duplicate sentence links, and address-in-cluster.
Each report $r$ about address $a$ carries a victim label $y_r$. We give each report a latent campaign $g_r$. Campaign $k$ has a true category $t_k$, and each report has a latent reliability $c_r$ (the probability the victim labels the campaign correctly rather than guessing from a confusion distribution $\xi$ over categories):
$$ p(y_r \mid g_r = k, c_r) = c_r\,\mathbb{1}[y_r = t_k] + (1 - c_r)\,\xi(y_r). $$
This is NUTMEG's likelihood with the group index made latent. The prior $q(g_r)$ comes from a relational graph neural network over the document graph, trained jointly with the latent-variable model by variational expectation–maximization. The network proposes campaign assignments, and the annotation likelihood rewards assignments under which within-campaign disagreement looks like noise.
An address's output is the distribution over categories implied by its campaigns, weighted by report counts, together with a separate multi-use score (the entropy over the campaigns the address belongs to). We then train an on-chain classifier on transaction features of reported addresses. The primary loss is on the hard label, and an auxiliary loss is inverse KL against our soft targets, as in Fornaciari et al. We apply the classifier to unreported addresses.
Baselines
The naive reference is the majority category per address, a hard label, which is what every forensic paper in the file effectively uses. The second baseline is the raw empirical vote distribution used as a soft target without denoising, which isolates whether modeling adds anything beyond keeping the disagreement. The third is MACE, the standard single-truth consensus model, which NUTMEG shows conflates minority views with spam.
The strongest competitor is NUTMEG given groups by clustering report text with an off-the-shelf sentence encoder and k-means, with $k$ chosen on validation. It shares our likelihood and differs only in whether groups come from the document graph. This is also the key ablation isolating our main ingredient. A further ablation removes entity edges (keeping only text similarity) and removes on-chain cluster edges, to see which relation carries the signal.
For per-report relabeling we run an instruction-tuned language model that reads each report's text and assigns a category, ignoring the victim's choice, and then aggregate per address. On the downstream side, the base learner is CatBoost with on-chain features plus TF-IDF of reports, following StableAML. The only thing varied across arms is the training target, and a DIAM-style encoder is tried as a second learner.
All arms share the same splits, the same entity extraction, and the same hyperparameter search budget (equal numbers of trials on the validation fold). No reported result for this task on this corpus was found on Lacuna. Beating the strongest baseline means lower held-out log-loss on future reports and higher agreement with the expert set than NUTMEG-with-embedding-clusters, with non-overlapping bootstrap intervals. For context only, the file reports NUTMEG at 0.81 competence correlation versus 0.58 for MACE (synthetic, NUTMEG paper) and CatBoost at 0.9775 Macro-F1 on a different stablecoin task (StableAML).
Why not just…?
Why not just cluster the report texts with embeddings and run NUTMEG? This is the real threat. Extortion templates are highly repetitive, and flat clustering may already find them. The graph's claimed advantage is that operators paraphrase lures but reuse contact emails, onion domains and payout addresses. Only entity edges catch these, and they tie reports that embeddings place far apart. The ablation settles it. If embedding clusters match the full graph within bootstrap noise, we will report that the graph is unnecessary and that the contribution reduces to inferring NUTMEG groups from text.
Why not just have a language model relabel every report from its text? The model may beat confused victims per report, and we expect it to be a strong baseline. It still answers only "what does this report describe". It cannot say whether an address's mixed labels reflect confusion or several campaigns, and it inherits the same catch-all taxonomy problem. We also use its per-report output as an alternative feature for $q(g_r)$. If relabel-then-majority matches us on future-report prediction and on the expert confusion-versus-multi-use judgments, the latent model is not earning its cost.
Why not just train CatBoost on on-chain and TF-IDF features with majority labels, as StableAML would suggest? We keep CatBoost as the learner. The question is only which targets it sees, so this alternative is one arm of the downstream experiment, not a competitor to the method. Fornaciari et al. report that soft targets help on ambiguous items (stemming 73.59% to 77.09% accuracy), but that was POS tagging and stemming, not forensic data. Whether the gain transfers is exactly what the downstream comparison tests.
Proposed Experiments
The first decisive check is temporal. We fit every method on reports before a cutoff and score the predicted category distribution per address against the empirical distribution of reports filed after it, using log-loss and Jensen–Shannon divergence. Future victims are independent annotators, so this avoids needing a gold label.
The second check uses an expert set. Two or more security analysts adjudicate a stratified sample of about a thousand high-disagreement addresses as confused or multi-use and assign true categories. Their own agreement is reported as the ceiling.
The third check tests de-pseudonymization. We measure whether pairs of addresses placed in the same inferred campaign also share a multi-input on-chain cluster more often than pairs grouped by embedding clusters, with on-chain edges removed from the graph for this test to avoid circularity. We also check whether the campaigns link addresses that no on-chain heuristic connects.
The fourth check is downstream. On-chain classifiers trained on each target type are evaluated on addresses independently labeled by the ransomware family lists integrated in BitcoinHeist (BitcoinHeist). That dataset reports 16.59 false positives per true positive for its best topological model. Its 2009–2018 window may overlap little with the report corpus, and we will report the overlap size before drawing conclusions.
The proposal fails cleanly if the embedding-cluster ablation matches it on the first two checks.
References
- A Large-Scale Bitcoin Abuse Measurement and Clustering Analysis Utilizing Public Reports
- Effective Illicit Account Detection on Large Cryptocurrency MultiGraphs
- Behavior-Aware Account De-Anonymization on Ethereum Interaction Graph
- StableAML: Machine Learning for Behavioral Wallet Detection in Stablecoin Anti-Money Laundering on Ethereum
- NUTMEG: Separating Signal From Noise in Annotator Disagreement
- Heterogeneous Graph Neural Networks for Extractive Document Summarization
- Document-level Relation Extraction with Dual-tier Heterogeneous Graph
- Beyond Black & White: Leveraging Annotator Disagreement via Soft-Label Multi-Task Learning
- Resurrecting Address Clustering in Bitcoin
- BitcoinHeist: Topological Data Analysis for Ransomware Prediction on the Bitcoin Blockchain