Continuous casting process control is crucial for enhancing the quality of steel. Center segregation, a key indicator of quality, plays a significant role in ensuring the quality of cord steel. However, the complexities of the continuous casting process, such as data silos, delayed inspections, uneven sampling and non-linearity, pose challenges to making timely improvements in quality. This study introduces an ensemble learning algorithm called LOF-KPCA-XGBoost, which combines Local Outlier Factor (LOF) sample weight, Kernel Principal Component Analysis (KPCA) and eXtreme Gradient Boosting (XGBoost) for online prediction of the center segregation coefficient of continuous casting billets. The algorithm achieves an accuracy of 90% and 98% within ±3% and ±5% relative error, respectively, with an average relative error of 1.6%. A sensitivity analysis method that leverages multivariable feature fluctuations is proposed to identify the process parameters that affect billet quality. The proposed method assists steel enterprises in improving the quality of cord steel billets.