Inter-element dependency models for sequence classification

Naive Bayes is a fast to train model for sequence classification. We develop and experiment with two methods that are equally fast (they require only one pass through the training data), but they are also able to models interactions among close neighbours in the sequence (unlike the Naive bayes independence assumption). The first method basically runs Naive Bayes on overlapping k-grams obtained from the sequence. The second method, called NB(k), constructs a classifier associated with an undirected graphical model over the sequence that takes into account k-wise dependencies. We test our algorithms on protein function classification tasks based on functional families from the Gene Ontology (GO) database. Our results show significant improvements of the proposed methods over Naive Bayes with NB(k) leading over NB k-grams in all test cases. The two proposed methods despite having improved modelling accuracy and demonstrated improved test accuracy maintain the “one pass through the data only” training property of Naive Bayes thus yielding fairly accurate and efficient classifiers.