The development of homogeneous catalytic reactions is hindered by the resource-intensive nature of traditional experimental and computational methods, particularly when dealing with small, sparse datasets and complex multi-objective optimization. While machine learning (ML) offers a promising alternative, prevailing methods rely on large-scale datasets and often fail to address the challenges of small-sample datasets, high-dimensional descriptor spaces, and strong nonlinear relationships inherent in catalytic processes. To address this issue, we present a nonlinear multi-objective ML workflow, PSO_CRP, for predicting reaction categories (e.g., conversion or yield) and quantitative outcomes (e.g., enantioselectivity or site selectivity), as well as the interpretability analysis. Experimental results on 4 small-sample datasets demonstrate that our models, relying only on simple RDKit-derived molecular parameters rather than costly DFT calculations, achieve consistently higher predictive accuracy than at least five common ML models. Model-based permutation feature importance (PFI) and partial dependence plot (PDP) analyses identified the key molecular descriptors governing reaction outcomes. They quantified their contributions, producing results aligned with previous studies while providing added mechanistic insight to enhance model interpretability and guide rational catalyst design. As a high-precision, low-cost, and interpretable framework, this PSO-based workflow offers valuable insights for forward prediction in homogeneous catalytic systems.