This study aims to significantly improve existing quantitative structure-property relationship (QSPR) models for predicting the octanol-water partition coefficient (KOW). This is because accurate predictions of KOW are crucial for assessing the environmental behavior and bioaccumulation potential of chemicals. Previous models have reported determination coefficient (R2) values between 0.9451 and 0.9681, and this research seeks to exceed these benchmarks. Three machine learning (ML) models are explored, i.e., feed-forward neural networks (FNN), Extreme Gradient Boosting (XGBoost), and random forest (RF). Using a dataset of 14,610 solvents (14,580 after data cleaning) and 21 molecular descriptors derived from SMILES representations, we rigorously evaluate these models based on R2, mean absolute error (MAE), root mean squared error (RMSE), and mean relative error (MRE). Notably, the best model developed, the XGBoost-based QSPR, demonstrated exceptional performance, exhibiting an impressive R2 value of 0.9772, surpassing benchmarks set by prior research models. Additionally, shapley additive explanation (SHAP) analysis is also employed for model interpretation, and it is revealed that the top five influential input features include SMR_VSA8, SMR_VSA3, Kappa2, HeavyAtomCount, and fr_furan. This study not only sets a new benchmark for KOW prediction accuracy but also enhances the interpretability of QSPR models.