Optimizing Probabilistic Models for Relational Sequence Learning

This paper tackles the problem of relational sequence learning selecting relevant features elicited from a set of labelled sequences. Each relational sequence is firstly mapped into a feature vector using the result of a feature construction method. The second step finds an optimal subset of the constructed features that leads to high classification accuracy, by adopting a wrapper approach that uses a stochastic local search algorithm embedding a Bayes classifier. The performance of the proposed method on a real-world dataset shows an improvement compared to other sequential statistical relational methods, such as Logical Hidden Markov Models and relational Conditional Random Fields.