The Optimization of Sum-Product Network Structure Learning

Sum-Product Network (SPN) are recently introduced deep tractable Probabilistic Graphical Models providing exact and tractable inference. SPN have been successfully employed as density estimators in some artificial intelligence fields, however, most of the proposed structure learning algorithms focus on improving the performance of a certain aspect of model, at the cost of reducing other performance. This is due to the fact that there is no effective balance between network width and depth during learning process. In this paper, we propose two clustering analysis algorithms to replace the clustering part of LearnSPN. We improve the structure quality of the generated model by deepening the network while adjusting the network width adaptively, trying to find a balance between the expressive power, representation ability, inference accuracy and simplicity. Experimental results prove that LearnSPN equipped by our clustering method has different degrees of improvement in various performances.