Graph Laplacian and horseshoe-regularized Bayesian latent component analysis of hydrocarbon fingerprints and...

Graph Laplacian clustering and horseshoe-regularized Bayesian latent component regression were applied to n-alkane, PAH, toxic equivalent (TEQ), and benzo[a]pyrene equivalent potency (BaPE) data from 12 Volta Lake sediment stations sampled in April 2025. The strongest numerical support was for a two-cluster graph partition (silhouette = 0.343; Davies-Bouldin index = 1.017; modularity = 0.447). Σn-alkanes were highest at VL3, VL5, VL1, and VL6, whereas Total PAHs were highest at VL9, VL11, VL10, and VL8. The aromatic-leaning second component showed a positive posterior direction for TEQ (β = 0.202; Pr[β > 0] = 0.965; posterior odds = 27.30) and BaPE (β = 0.124; Pr[β > 0] = 0.863; posterior odds = 6.29), but both 95% credible intervals crossed zero. Posterior ranking identified VL10, VL8, and VL9 as priorities for repeat sampling, not confirmed regulatory hotspots. The primary two-cluster partition was unchanged across Gaussian-kernel bandwidth multipliers of 0.50-1.50 and across 1000 multiple-imputation datasets; a conservative 30% analytical-perturbation bootstrap reproduced the exact partition in 96.0% of 1,000 replicates. Because the design comprised only 12 stations, one sampling period, a time-integrated 0-60 cm sediment interval, and no measurements of total organic carbon, black carbon, or grain size, the workflow remains an exploratory screening analysis rather than validated source apportionment or risk classification.