To solve the issue of poor fault diagnosis accuracy induced by unbalanced and limited dissolved gas analysis (DGA) samples of oil-immersed transformer faults, the method based on cuckoo search optimized categorical boosting algorithm (CS-CatBoost) and improved synthetic minority over-sampling technique (SMOTE) is proposed. First, the center offset weight (COW) is used to optimize the SMOTE and enhance unbalanced fault samples, obtaining a balanced dataset. Then, a base classifier based on an ensemble learning framework is constructed using CatBoost. The classification performance of the CatBoost model is significantly influenced by its initial parameters or may select incorrect parameters, thereby leading to overfitting or underfitting. Thus CS is introduced to optimize its initial parameters, further enhancing its classification performance. Experimental results demonstrate that under conditions of small sample size and imbalance, the proposed SMOTE-CS-CatBoost model significantly improves fault diagnosis accuracy compared to other methods, accurately identifying transformer fault information.