Pore pressure (PP) information is essential for gas hydrate exploration, field development, and other geological modelling studies. The prediction of PP from borehole data is difficult due to the complex (non-linear) relationship between the input (e.g., well log; over-burden stress) and target variables (PP). An ensemble learning-based interpretable method (ELBIM) is proposed to model PP using multivariate borehole data of the IODP expedition 372 drill site U1517A of the Northern Hikurangi Margin, New Zealand. The predictive model is configured by taking input of seven petrophysical parameters (e.g., NMR porosity, caliper, gamma-ray, density, neutron porosity, temperature, and sonic transit time) with lags based on auto-correlation function (ACF) and partial autocorrelation function (PACF) analyses of well log and target PP using empirical techniques (Eaton’s and Porosity method). Comparative analysis showed that the ELBIM with extreme gradient boosting (XGB) regression outperformed gradient boosting machine (GB), random forest regression (RFR), and decision tree regression (DTR) based on root-mean-square errors (RMSEs) RMSEXGB ~0.0087, RMSEGB ~0.0090, RMSERFR~0.0123 and RMSE DTR ~0.0165, and coefficients of determination (R2) of R2XGB ~ 0.9954, R2GB ~ 0.9950, R2RFR ~ 0.9908 and R2DTR ~ 0.9834 between the observation and the prediction of PP with 95% confidence interval (CI). The present analysis suggests that the ELBIM with XGB yields more precise forecasts of PP and exhibits improved detection of the bottom-simulating reflector (BSR) at the drill site. The proposed ELBIM with XGB has the potential to predict the PP from multivariate well logs and could be applied to other complex areas. Highlights Pore pressure data is vital for field planning, drilling, and geological studies. An ensemble learning-based interpretable method (ELBIM) is proposed for the prediction of PP from well-log data. Autocorrelation and partial auto-correlation analyses were performed. XGB regression outperformed GB, RRF, and DTR regression. The proposed method can be applied to other complex ocean areas. Ensemble-based machine learning approaches can be a handy tool to achieve reasonable accuracy with limited data exposure.