Since mineralogical and geophysical conditions of lithofacies are available to provide a deeper insight of reservoirs' storing mechanism, diagenetic facies gradually acts as a more significant role in the petroleum-based geological study. Most kinds of carbonate diagenetic facies universally have similar responses on some conventional logs, and then for the issue of their logging-based classification an excellent predictor is normally needed. Light gradient boosting machine (LightGBM) is applied as a potential predictor in this study due to its state-of-the-art classifying capability, while its perfect working performance usually needs the support of high-quality inputs and optimal parametric setting. Thus, to guarantee the reliability of outputs of LightGBM, continuous restricted Boltzmann machine (CRBM) and artificial fish swarm algorithm (AFSA) are adopted as assistants to process inputs and optimize parametric setting, respectively, and a hybrid predictor named CRBM-AFSA-LightGBM therefore is proposed. Data for validation of the proposed predictor is cored from the carbonate reservoirs of the Santos Basin. Four experiments are designed to orderly verify feasibility, improvement, robustness, and generalization of the hybrid predictor. To highlight testifying effect, three sophisticated classifier are introduced as competitors, including k-nearest neighbors (KNN), support vector machine (SVM), and random forest (RF). The achieved experimental outcomes demonstrate three key points: 1) LightGBM-cored predictor presents more feasibly than competitors in any validation, and more importantly its predicting capability can be further enhanced under the training of a larger-size dataset; 2) a better robustness is testified for LightGBM-cored predictor since a satisfactory classification is still accessible based on the learning of a sparse-condition dataset; 3) transfer learning is proved effective in improving generalization of LightGBM-cored predictor.