Abstract: Agricultural production services are an important means of transforming agricultural production and management, integrating smallholders into modern agriculture, and improving the performance of large-scale farming. However, the factors shaping growers’ perceptions of relative performance may differ across production stages. Based on 381 paired questionnaires from large-scale rice growers covering the seedling raising and harvesting stages, this study develops a random forest and extreme gradient boosting (RF-XGBoost) residual cascade model and uses Shapley additive explanations (SHAP) to compare the contributions of service organization mode, direct costs, labor costs, land conditions, perceived policy support, transaction costs, technical costs, and socioeconomic conditions to perceived relative performance. The results show that: 1) a paired-sample t-test reveals no statistically significant difference in mean perceived relative performance between the seedling raising and harvesting stages; 2) perceived policy support ranks first in feature importance in both stage-specific models; and 3) the most important features in the seedling raising model are perceived policy support, socioeconomic conditions, transaction costs, and land conditions, whereas those in the harvesting model are perceived policy support, direct costs, service organization mode, and technical costs. These findings indicate that stage-specific heterogeneity in rice production services is primarily reflected in differences in the composition and ranking of key explanatory features within the respective models, rather than in differences in mean perceived relative performance. Accordingly, support for agricultural production services should be tailored to the distinct technological characteristics, transaction arrangements, and service market conditions of the seedling raising and harvesting stages.