Music Recommendation: Audio Neighbourhoods to Discover Music in the Long Tail

Abstract

The final publication is available at Springer via http://dx.doi.org/10.1007/978-3-319-24586-7_6 Abstract. Millions of people use online music services every day and recommender systems are essential to browse these music collections. Users are looking for high quality recommendations, but also want to discover tracks and artists that they do not already know, newly released tracks, and the more niche music found in the 'long tail' of on-line music. Tag-based recommenders are not effective in this 'long tail' because relatively few people are listening to these tracks and so tagging tends to be sparse. However, similarity neighbourhoods in audio space can provide additional tag knowledge that is useful to augment sparse tagging. A new recommender exploits the combined knowledge, from audio and tagging, using a hybrid representation that extends the track's tag-based representation by adding semantic knowledge extracted from the tags of similar music tracks. A user evaluation and a larger experiment using Last.fm user data both show that the new hybrid recommender provides better quality recommendations than using only tags, together with a higher level of discovery of unknown and niche music. This approach of augmenting the representation for items that have missing information, with corresponding information from similar items in a complementary space, offers opportunities beyond content-based music recommendation.

The Effects of Consumer Preference and Peer Influence on Trial of an Experience Good

“…In addition, we can observe the list of neighbors on each user's web page. Last.fm recommends 50 nearest neighbors to users on the basis of preference similarity and encourages them to visit neighbors’ personal web pages to discover new songs (Chen, Boring, and Butz 2010; Craw, Horsburgh, and Massie 2015). In this regard, neighbors are different from peers, which are bidirectional relationships formed with mutual consent.…”

A switching multi-level method for the long tail recommendation problem

“…Graph-based algorithms have also been proposed in [22] for the long tail recommendation by using user-item information along with undirected edge weighted graphs for long tail item recommendation. In [26] a case base reasoning method presented and showed that the recommendations were based on unknown artists and tracks. The proposed system in that study could identify whether an item resided in the long tail and if it were attempting to improve its provided meta-data through the addition of tag knowledge information.…”

Recommender systems as “tastemakers”: collaborative filtering as a market strategy for online cultural products

“…The classification of symbolic goods as "products of belief " endorses the argument about the existence of a strong "social demand" for recommender systems of online cultural products. In the online environment, despite the wide variety of products offered to consumers, excessive options can discourage people from concluding their search and selection processes for a purchase (Craw et al, 2015;Martínez-López et al, 2010;Haynes, 2009;Iyengar & Lepper, 2000;Schwartz 2004). The potential influence of these systems in guiding demand in the cultural market and their impact on e-commerce may be illustrated with a few examples, presented below.…”