<p>Post-construction settlement of soft soil embankments significantly affects the serviceability and long-term maintenance of roadways. This study proposes a probabilistic prediction method for soft soil embankment settlement considering the pile-soil interaction. The pile-soil interface parameters and soil parameters are adaptively updated by assimilating the monitoring data, leading to enhanced probabilistic predictions of settlements. A random forest-based surrogate model is developed to improve the computational efficiency. The probability distributions of key parameters are updated by incorporating multi-point settlement monitoring data within a Bayesian framework, and the uncertainty in settlement prediction is quantified. A case study of the Saga embankment project in Japan is used to illustrate the proposed method. The results show that the random forest surrogate model can effectively replace the finite element model, with a high <i>R</i>² of 0.921. Sobol global sensitivity analysis indicates that the modified compression index, the modified creep index, and the interface reduction factor of the soft clay layer are the dominant parameters governing the settlement response. As monitoring data are progressively incorporated, the prediction errors decrease, and the forecasted settlement in later stages closely matches the observed values. This research provides an efficient and accurate solution for settlement prediction of soft soil embankments reinforced with cement-mixed piles.</p>