Bias correction for distributed Bayesian estimators

Dealing with the whole dataset in big data estimation problems is usually unfeasible. A common solution then consists of dividing the data into several smaller sets, performing distributed Bayesian estimation and combining these partial estimates to obtain a global estimate. A major problem of this approach is the presence of a non-negligible bias in the partial estimators, due to the mismatch between the unknown true prior and the prior assumed in the estimation. A simple method to mitigate the effect of this bias is proposed in this paper. Essentially, the approach is based on using a reference data set to obtain a rough estimation of the parameter of interest, i.e., a reference parameter. This information is then communicated to the partial filters that handle the smaller data sets, which can thus use a refined prior centered around this parameter. Simulation results confirm the good performance of this scheme.