Dynamic heterogeneity is one of the hallmarks for the dynamics of glass-forming liquids, strongly associated with non-exponential relaxation. The non-exponential relaxation can lead to the structural relaxation increasing as a super-Arrhenius formula with quenching, while with only minor change in structures. Whether there is a relationship between the statistic structure and the drastically changed dynamics is still in debate in glassy community. In this study, we combine large-scale molecular dynamics simulations with the supervised machine learning to investigate the dynamics of the binary glass-forming system Cu₅₀Zr₅₀. The local atomic environment is characterized using symmetry functions, and atomic mobility, i.e., the logarithm of the migration rate ln(m_i). We develop an interpretable atomic-scale structural parameter, termed ”softness” (S_i), using an XGBoost model to elucidate the structure-dynamics relationships at the atomic level. Our results show that the probability distribution of the atomic mobility evolves from a unimodal to a bimodal distribution with quenching, consistent with the emergence of dynamical heterogeneity. The softness mapping reveals the coexistence of high- and low-softness atomic populations. For atoms with a given softness S_i, the activation of local structural rearrangements follows Arrhenius behavior, with the activation energy approximately linearly correlated with S_i. In contrast, the softness dependent α relaxation time exhibits a clear non-Arrhenius behavior, indicating the strong collectivity between different structural units. This distinct response behavior indicates that S_i effectively separates the “intrinsic structural activation energy” from the “collective cooperative effects”, providing new insights into the structure-dynamics coupling during the glass transition. The high and low softness atomic clusters identified by S_i can be regarded as distinct structural units with different local activation tendencies. With temperature decreasing, the distribution of softness parameter exhibits a transition from a high-softness state to a low-softness state, indicating a pronounced redistribution of the activation propensity of local atomic structures in the supercooled liquid regime. These results provide new insights on the structural evolution and the structural significance for the dynamic heterogeneity during glass transition.