Gravar-mail: Impact of training sets on classification of high-throughput bacterial 16s rRNA gene surveys