The classification of polyphonic music still presents challenges for current music data mining methods. In this paper we explore the performance of classifiers specifically created for melody on the polyphonic classification task. On a small dataset of string quartet movements of Haydn and Mozart, the melodic n-gram model outperforms the melodic global feature model for composer recognition. Furthermore, a simple model that combines the predictions made from different instrumental parts outperforms models created from any single voice. The results indicate that models taking into account polyphonic information achieve higher classification accuracy.