Learning Selective Sum-Product Networks
Robert Peharz, Robert Gens, Pedro Domingos · 2014
We consider the selectivity constraint on the structure of sum-product networks (SPNs), which allows each sum node to have at most one child with non-zero output for each possible in-put. This allows us to find globally optimal max-imum likelihood parameters in closed form. Al-though being a constrained class of SPNs, these models still strictly generalize classical graphical models such as Bayesian networks. Closed form parameter estimation opens the door for struc-ture learning using a principled scoring function, trading off training likelihood and model com-plexity. In experiments we show that these mod-els are easy to learn and compete well with state of the art. 1.