Algorithm 1.

Intra-site Posterior update in EXPLORER

1:	Initialize : each approximate likelihood function ${\stackrel{\sim }{f}}_i^j({\mathit{\beta }}_j)$ and approximate posterior $q({\mathit{\beta }}_j)=m_{h_j\to B_j}({\mathit{\beta }}_j){\prod }_{i=m_{j-1}+1}^{m_j}{\stackrel{\sim }{f}}_i^j({\mathit{\beta }}_j)$n
2:	Repeat
3:	for all records i = mj−1 + 1, · · ·, mj do
4:	Get partial posterior function q\i (βj) by removing the approximate factor ${\stackrel{\sim }{f}}_i^j({\mathit{\beta }}_j)$ from the approximate posterior: $q^{\backslash i}({\mathit{\beta }}_j)=q({\mathit{\beta }}_j)/{\stackrel{\sim }{f}}_i^j({\mathit{\beta }}_j)$
5:	Update q(βj) by incorporating a single true likelihood function $f_i^j({\mathit{\beta }}_j)$ according to Assumed Density Filtering (ADF) [33]: $q({\mathit{\beta }}_j)=\mathit{Proj}[\frac{1}{Z_i}{q}^{\backslash i}({\mathit{\beta }}_j)f_i^j({\mathit{\beta }}_j)]$
6:	Set the approximate factor: ${\stackrel{\sim }{f}}_i^j({\mathit{\beta }}_j)=\frac{Z_{\text{i}}q({\mathit{\beta }}_{\text{j}})}{q^{\backslash \text{i}}({\mathit{\beta }}_{\text{j}})}$
7:	end for
8:	Until parameters converge