This study constructs 196 transition metals (TM)@S_xN_y single-atom catalysts (SACs) (x = 0–4 and y = 0–4) and employs the eXtreme Gradient Boosting (XGBoost) classification model in machine learning (ML) for effectively distinguishing qualified and unqualified catalysts. The prediction accuracy rate is high, up to 95%. The SHapley Additive exPlanations (SHAP) analysis reveals that the N≡N bond length and the number of outermost d electrons (N_d) can well describe the nitrogen (N₂) reduction reaction (NRR) activity. The relationships between N≡N, N_d, the adsorption energies of different intermediates ( $\mathrm{\Delta }{E}_{{\ast }_{{\text{N}}_2}}$, $\mathrm{\Delta }{E}_{{\ast }_{{\text{N}}_2\text{H}}}$, and $\mathrm{\Delta }{E}_{{\ast }_{{\text{N}\text{H}}_2}}$), the general descriptor (φ), and the Gibbs free energy of key steps ( $\mathrm{\Delta }{G}_{{\ast }_{{\text{N}}_2}}$, $\mathrm{\Delta }{G}_{{\ast }_{{\text{N}}_2-}{\ast }_{{\text{N}}_2\text{H}}}$, and $\mathrm{\Delta }{G}_{{\ast }_{{\text{N}\text{H}}_2-}{\ast }_{{\text{N}\text{H}}_3}}$) indicate that moderate nitrogen activation can enhance the reaction activity. Among the 17 screened SACs, Mo@S₃N₁, and W@S₃N₁ demonstrate the best catalytic performance, with limiting potential (U_L) values of only −0.26 and −0.25 V under implicit solvation conditions. The electronic properties and variations in N≡N and TM–N bond lengths are investigated to reveal the origin of NRR activity. This study provides the decisive features and NRR dataset for ML research, as well as a feasible strategy for rational design of NRR SACs.