Learning Structure of Bayes Networks with K2

So, I recently implemented Bayes Networks for discrete variables, and implemented the K2 algorithm for learning the structure of a network ( http://web.cs.wpi.edu/~cs539/s05/Projects/k2_algorithm.pdf ) .

Has anyone else implemented this? I ask, because its results are just horrible. It will usually infer only one or 2 variables as having a relationship with the target class, with very few nodes ending up in the Markov blanket. Naive Bayes often ends up giving better results then the inferred graph. Has anyone else observed this?