<p>Data censoring is a common problem during the process of survival data collection. To maximize the usable information in dataset with censoring, Cox model has been proposed and becomes of a benchmark model in censoring data modelling. However, as data size grows, challenging raises on learning Cox model and its modern extensions. Existing algorithms for training Cox model facing the problems of insufficient sample pair usage and control sample unbalance. In this work, the authors introduce the batch recombination algorithm: A chain-based method that better uses sample pairs while keeping control sample balance. The authors show that, under mild conditions, parameter estimates from stochastic gradient descent using our recombinated batch are consistent. Confidence interval can also be established using asymptotic distribution. Extensive numerical experiment both on simulated data and real data, linear and neural network Cox model show efficiency and accuracy of the proposed method.</p>