In the fields of coal dust control and coalbed methane (CBM) development, wettability is a crucial parameter of coal, often determined by the coal–water contact angle (CA). In order to construct an accurate CA prediction model, extensive data on industrial components, element content, and coal CAs were collected. Two sets of data were utilized: a large sample comprising 98 data groups gathered from various sources, and a small sample consisting of 16 data groups collected from a single source. These datasets were employed to develop models using three machine learning (ML) methods: K-nearest neighbor, support vector regression, and back-propagation neural network. The results revealed a significant underfitting phenomenon in all three ML methods when applied to the training and testing datasets of the large sample. This underfitting can be attributed primarily to the variations in coal sample handling by different scholars. Conversely, the three ML methods exhibited pronounced overfitting on the training dataset of the small sample, resulting in limited generalization ability on the testing dataset. This limitation arises from the small amount of data in the small sample. To address this, the synthetic minority oversampling technique was employed to generate augmented samples for the small sample. The correlation of determination of the augmented samples ranges from 0.82 to 0.92, indicating an excellent fit. Additionally, the fit superiority ratios of the training and testing datasets fell between 0.92 and 1. This approach effectively avoids the risk of underfitting in large-sample datasets and overfitting in small-sample training datasets. In the final stage, the developed model was used to predict the wettability of coal samples from three coal mines in the Qinshui Basin, China. The predicted CA values demonstrated a high level of agreement with the CA values measured in the laboratory. This comprehensive study thoroughly analyzed the underlying reasons behind the failure of ML models to effectively handle large- and small-sample data for CAs. It also provides a valuable solution to the above problems by data augmentation with small samples, which holds great significance in enabling quick and accurate prediction of CAs using limited coal parameter data.