ZSH: A Zeta-Weighted Hybrid Semantic Clustering Framework for Behavioral Profiling of Blockchain Transactions

This study introduces ZSH (Zeta-weighted Semantic Hybrid), an unsupervised clustering framework designed to convert raw Bitcoin transaction records into interpretable behavioral profiles for financial crime and forensic investigation. The framework uses a rank-power feature weighting scheme, normalized by a generalized harmonic number, to prioritize the 27 structural and monetary attributes most informative under mutual information ranking, combined with a hybrid centroid initialization that blends rule-derived semantic seeds with Ward-linkage micro-clusters and a balance-constrained refinement step to prevent majority-class dominance. Evaluated on 11.3 million Bitcoin transactions, ZSH produced 43 statistically validated behavioral profiles (bootstrap Silhouette 0.2159, permutation p < 0.001) and, while conceding intrinsic Euclidean compactness to classical baselines such as KMeans++, demonstrated superior concentration of behavioral families relative to an unseeded baseline, a result that held even when families were withheld from seeding and that replicated on the externally labelled Elliptic corpus, where one cluster reached 82.1% illicit concentration at 8.41 times the base rate. An independent Isolation Forest anomaly layer was also rigorously tested against ground-truth illicit labels and found to be anti-correlated with illicit status, a negative result the authors report transparently rather than omit, reframing the layer as a general atypicality signal rather than a fraud detector. Through extensive ablation, sensitivity analysis, and raw-corpus replication, the study locates the source of ZSH's concentration advantage precisely in the zeta weighting mechanism and characterizes the associated trade-off against broader semantic alignment, offering the fintech and forensic analytics community a transparent, task-aligned alternative to purely geometric clustering for transaction-level behavioral profiling.