Fine-grained culture-aware music recommender systems

Having tens of millions of musical works available at a listeners fingertips requires novel recommendation and interaction techniques for music consumption. Thereby the success of a music recommender system, a system that proposes users what to explore or listen to next, depends on its ability to propose the right music, to the right user, at the right moment (i.e., in the right context). However, this task is extremely complex, as various factors influence a users music preferences. Amongst others, cultural aspects and characteristics (e.g., different requirements regarding diversity of a playlist or familiarity with its music tracks) have been shown to affect music perception, preferences, and listening behavior. Calling on this, the project entitled Fine-grained culture-aware music recommender systems investigates how music recommender systems could and should integrate cultural aspects in order to provide better recommendations. The research findings will answer the question how music recommender systems have to be designed to reflect cultural diversity and will provide insights into cross-cultural music perception, preferences, and listening behavior. Specifically, the project will investigate the cultural requirements on music recommender systems as concerns what listeners in different cultures expect with regard to the recommended music. Thereby, we postulate that different granularity levels of culture (e.g., individual, regional, national, or global level) have to be considered to improve music recommender systems. We hypothesize that the various cultural levels of different granularities have to be combined in a comprehensive way to transcend limitations of current music recommender systems. And we will investigate its impact on recommendation quality in cross-cultural studies with users from Austria, the United States, and Korea. Our scientific approach comprises four methodological orientations: (i) a combination of surveys and user panels, (ii) user modeling, (iii) designing and implementing prototypes of culture-aware music recommender systems, and (iv) cross-cultural studies with users to investigate their performance. The samples will include users from the United States, Austria, and Korea; we will focus on national culture, but also consider regional cultures (e.g., urban vs. suburban vs. countryside areas). In contrast to past research in the field of culture-aware music information retrieval and recommendation, the project follows an approach that is driven by user needs and preferences. The project aims to design and implement music recommender systems that are able to meet those requirements by considering different granularity levels of cultural aspects in a comprehensive way.

The project "Fine-grained culture-aware music recommendation system" investigated how music recommendation systems could leverage cultural. The project analyzed culture- specific differences in music preferences, for which we relied on data on listening behavior on music platforms. These findings have been incorporated into algorithmic music recommendation approaches, whereby we have shown that the inclusion of culture-specific differences leads to lower error rates in the recommendations thus, delivering better results. The projects main findings are summarized below: 1. A users music preferences can be described in terms of the degree to which they prefer music items that are currently popular (the mainstream) or rather ignore such trends; termed a users mainstreaminess. Here, mainstream may be defined globally, but also on a country-specific level; thereby, the country-specific mainstream does not necessarily correspond to the global mainstream. 2. With regard to artist popularity, the country-specific music listening behavior may deviate from the global one. In some countries, the listening behavior corresponds to the global mainstream; some countries have developed their own country-specific mainstream in addition to the global mainstream; a third group of countries shows clear deviations from the global mainstream, although a country-specific mainstream is not clearly noticeable. 3. Comparing the top charts of different countries, it is typically the same artists who are represented. For considering the country-specific nuances in defining mainstreaminess, we require approaches that downweigh popular superstars or give more weight to country- specific artists. We found approaches based on the Kullback-Leibler divergence and approaches based on Kendall`s tau being suitable. In combination with matrix factorization, the approach based on Kendall`s rank correlation coefficients is particularly successful in music recommendations experiments. 4. For low-mainstreamy users, we achieve particularly strong improvements in recommendation results (measured in the error rate) when compared to a generic approach that does not consider mainstreaminess or country. 5. First results indicate that there are differences in music preferences between urban and rural regions in conglomerations worldwide. Furthermore, countries can be grouped based on similarities in the users music listening behavior. Including such information as input for music recommendation, results can be improved. Building on this, future research may investigate the similarities in listening behavior based on, for example, sociological or economic aspects. Findings could then flow back into novel music recommendation approaches. 6. In music playlist creation in groups, users show different behavior patterns when confronted with a majority opinion. With a favored song, a single counter-opinion is enough to change a users mind and vote against the song for the playlist. In case of a disliked song, however, a majority opinion in favor of this song is required to change the user`s mind.