State Aggregation Learning from Markov Transition Data