Learning Markov networks with context-specific independences