Research On Predicate Invention Based On Regularized Sparsity

Inductive Logic Programming(ILP)is inductive inference in a restricted first order logic framework.A problem in ILP is how to extend the hypothesis language in the case that the vocabulary given initially is insufficient.One way to adapt the vocabulary is to introduce new predicates.Predicate invention in ILP means automatic introduction of new,hopefully useful predicates during the process of learning from examples.In statistical learning domain,this problem is known as hidden or latent variable discovery.Both hidden variable discovery and predicate invention are considered quite important in their respective communities,but are also very difficult,with limited progress to date.The usual predicate invention method is to rewrite a group of closely-related rules to use a common invented predicate as a "subroutine".Predicate invention is difficult,since a poorly-chosen invented predicate may lead to error cascades.Until the soft version of predicate invention proposed,it effectively overcomes the shortcomings of the traditional predicate invention,but the experimental performance of the method is not ideal.In this paper,a predicate invention method based on regularized sparsity is proposed,it is based on the first-order extensible logical platform ProPPR(Programming with Personalized PageRank),predicate invention by introducing of regularized sparsity model,in order to improve the efficiency of predicate invention.Instead of explicitly creating new predicates,this predicate invention method implicitly group closely-related rules by using regularized sparsity to regularize their parameters together.In addition,the influence of elastic network is researched emphatically.The approach proposed in this paper can effectively overcome the difficulty of error cascades.Experiment results show that this method can improve the mean average precision in the process of predicate invention and shorten the query time of the knowledge base.