Rapid development of the Internet provided us accessibility to abundant amount of data. However, such abundance in accessible data doesn;;t directly satisfy the user;;s need for information. Therefore, research in automatic analysis of music contents is increasing and many commercial companies provide recommendation service to end users. There are three different types of music recommendation based on the recommending algorithm: 1) Collaborative Filtering music recommender which uses a large database of user;;s preference to recommend music, 2) Content-based Filtering music recommender which analyzes the music content using musical features to discover similar music, and 3) hybrid music recommender which combines the advantages of the two methods described above. The problem with these approaches is that they don;;t consider user;;s situational context. Therefore, we provide a music recommender that use radio episodes. In each radio episodes, the user;;s situational context is provided. In this paper, we propose a music recommendation system based on radio episode analysis. In Korea, radio station DJ introduces the radio episode, written by the audience. These radio episodes are associated with reqeust songs and the DJ plays this song. In some occasions where the request song is missing, the DJ selects a song that suits the radio episode. The main goal of this research is to develop a music recommender that acts as this radio DJ. Our algorithm gathers radio episode data from www.imbc.co.kr via webcrawler and performs morpheme analysis to create a word-document matrix. Then LSA (Latent Semantic Analysis) is performed on the word-document matrix to find the semantic meanings of the episode. Once the semantic meaning is discovered, the similarity between episodes is calculated. Our assumption is that if episodes are similar then the requested song would be similar as well.There are both objective and subjective indicators in the evaluation to make sure the possibility of this system. The objective indicator evaluates if the suggested algorithm finds similar episodes well. The subjective Indicator evaluates if the suggested music matches with the specific episode.The result of the objective indicator showed that the algorithm found similar episodes better than random based approach. This proved that it is possible to recommend music by solely analyzing text data. The result of the subjective indicator showed that there is high similarity between the actual song request and the song provided by our algorithm. This result suggests that the song recommended by our algorithm matched the situational information written in the radio episode. The main contribution of the research is that by analyzing purely the textual data written by individuals, our algorithm was able to extract contextual information of the users and was able to recommend music based on the contextual information. Also, since our algorithm is based on textual information, it will be possible to implement the algorithm in SNS and diaries to recommend music.
PDF
download
878987797
028-85220240
