We introduce BAMF+HS (Bayesian Autocorrelated Matrix Factorisation+Horseshoe), an enhanced Bayesian receptor model for particulate matter (PM) source apportionment. By applying a regularised horseshoe prior to the chemical composition matrix, BAMF+HS enforces sparsity, filtering out irrelevant species and improving source separation. Tests on synthetic and real datasets show BAMF+HS consistently outperforms previous models in accuracy and clarity.