A Bayesian analysis of co-training algorithm with insufficient views

The co-training algorithm can be applied if a dataset admits a representation into two different feature sets (two views). However, its optimality is proved only under the conditions a) sufficiency of each view, and b) conditional independence given the class. We address the case where condition a) doesn't hold, as often happens in concrete applications. In such cases the co-training is unable to converge to the optimal Bayesian classifier, because samples added in the training set are not distributed according to the class-conditional distributions, even if their assigned label is correct. These results help to better understand the behavior of the co-training algorithm when the classes are only `statistically' separable.