Playlist environmental analysis for the serendipity-based data mining

The real recommendation will involve not only unsurprised, but some surprised elements sometimes. We are constantly experiencing this serendipity element in our daily life. However, today´s data mining technology cannot support this function adequately. In this paper, we hypothesize that there should be some relevant serendipity elements on a radio station´s music playlist. The radio station makes the playlist for the programme by the producer or director based on his experience and huge knowledge of music. In this research, we first determine the optimal number of clusters to be used by using BIC and AIC, and then we apply EM model clustering technique for the collection of more than 3000 playlists from the radio station. By analyzing the resulting clusters obtained, significant dependency between clusters and all the playlist metadata (i.e. time, genre, music era, and popularity) are discovered. Finally, we found that the time of playing the music and the music era (composed-year of music) has a strong correlation. This implies that music in some particular music era is popular in certain particular time of the day. For example, more classical music pieces from modern era are played at the night from 18:00 until 2:00. Furthermore, more classical music from the early Baroque era has been played in the afternoon around 12.00 to 13.00. However, we think that some music, which belong to the music era that are less popular in that particular time, will contribute to the serendipity function of the playlist to certain extend. With these studies, this serendipity would be widely used for an excellent recommendation service, which will include the personal radio station for the digital music player, marketing and knowledge capital.