Exploiting Innocuousness in Bayesian Networks

Boolean combination functions in Bayesian networks, such as noisy-or, are often credited a property stating that inactive dependences (e.g., observed to false) do not "cause any harm" and an arc becomes vacuous and could have been left out. However, in classic Bayesian networks we are not able to express this property in local CPDs. By using novel ADBNs, we formalize the innocuousness property in CPDs and extend previous work on context-specific independencies. With an explicit representation of innocuousness in local CPDs, we provide a higher causal accuracy for CPD specifications and open new ways for more efficient and less-restricted reasoning in (A) DBNs.