Abstract. Distributed inference schemes for detection, estimation and learning comprise an attractive approach to Wireless Sensor Networks (WSNs), because of properties such as asynchronous operation and robustness in the face of failures. Belief Propagation (BP) is a method for distributed inference which provides accurate results with rapid convergence properties. However, applying a BP algorithm to WSN is not trivial, due to the unique characteristics of WSN networks. Many papers which have proposed using BP for WSNs do not consider all of the constraints which these networks impose. This paper first undertakes a thorough study of the practical challenges of WSNs which are raised in the context of distributed inference. It then presents a framework which implements both localized and data-centric approaches to improve the effectiveness and the robustness of this algorithm in the WSN environment. The proposed solution is empirically evaluated, as applied to the clustering problem, and it can be easily extended to suit many other applications that use BP as an underlying algorithm.