Address Clustering

The transaction graph connects actual spent outputs to the transactions consuming them. A cluster additionally groups nodes by inferred control. A payment between two addresses alone is no reason to assign both to one owner; transaction edges and common-control links must remain distinct. [Meiklejohn et al. — A Fistful of Bitcoins]

An input references an earlier UTXO, not a universal sender-address field. The process must retrieve the original scriptPubKey and state the network and block range. Some scripts lack an ordinary address representation; forcing them into an alleged person would confuse a technical object with identity. [Bitcoin Developer Guide — Transactions]

Common-Input Ownership Heuristic connects the input side. Change detection may attach an output; choosing the recipient incorrectly merges both payment parties. Two outputs do not automatically mean payment and change: they can represent two recipients or transfers between personal wallets. Address novelty alone does not decide. [Meiklejohn et al. — A Fistful of Bitcoins] [BIP 78 — A Simple Payjoin Proposal]

Script type, rounded amounts and prior address use can be model features, not proof. CoinJoin and Payjoin violate some common assumptions; exceptions need not have conspicuously equal amounts. Software or user behavior changes can alter error rates, so historical performance cannot be transferred to current data without new measurement. [Möser and Narayanan — Resurrecting Address Clustering in Bitcoin] [BIP 78 — A Simple Payjoin Proposal]

Union-find efficiently forms transitive groups. An ordinary implementation cannot simply delete a past merge and correctly split the resulting cluster. Preserve underlying edges, transactions and rule versions; correction may require recomputation. A membership list alone does not retain the evidence path that produced the merge. [Möser and Narayanan — Resurrecting Address Clustering in Bitcoin]

Möser and Narayanan constrain union-find to prevent merging groups separated by a predicted payment output. This can curb cluster collapse, but the constraint itself depends on payment inference. A conservative rule may reject valid links too; fewer merges do not automatically describe all owners more accurately. [Möser and Narayanan — Resurrecting Address Clustering in Bitcoin]

Labeling one address with a service name and extending that label to a cluster are different claims. Record label provenance, period and propagation path; an exchange group is not one customer. Reference data derived from the same heuristic are not fully independent. Evaluation must separate false merges, false splits and sample coverage. [Meiklejohn et al. — A Fistful of Bitcoins] [Möser and Narayanan — Resurrecting Address Clustering in Bitcoin]

Recognizing an internal transfer can remove self-churn from estimated incoming payments. A different cluster version can therefore revise a historical metric without changing the blockchain. Compute group balances from its still-unspent UTXOs at a specified block, not all historically received outputs. Report model and label versions and the data cutoff. [Meiklejohn et al. — A Fistful of Bitcoins] [Möser and Narayanan — Resurrecting Address Clustering in Bitcoin]