What makes
a good music recommendation?
At MediaUnbound, we have spent over six years pondering this question. The answer is not a single magical algorithm. Nor is it a secret combination of sonic features. The answer is complex and changes based on context, audience, and customer requirements.
When making recommendations to music listeners, MediaUnbound follows these guidelines:
At MediaUnbound, we have spent over six years pondering this question. The answer is not a single magical algorithm. Nor is it a secret combination of sonic features. The answer is complex and changes based on context, audience, and customer requirements.
When making recommendations to music listeners, MediaUnbound follows these guidelines:
- Focus on the Listener - Understanding and learning about the listener is at the center of MediaUnbound's recommendation philosophy. This approach has resulted in formally modelling not listeners or music, but the complexity of the music-listening experience. Most other personalization or recommendation systems typically have their basis in music categorization -- extracting and describing the features of songs to help users find similar sounding songs. Listener centricity produces a more relevant set of data and analytic techniques for determining which songs a person will like. Song similarity is a useful metric for searching and recommendation but is insufficient as a sole predictor of preference.
- Use human music analysts to complement computer technology - Computers and humans have unique strengths. A computer can catalog, arrange, and archive large data sets in efficient ways. But, it cannot understand why people enjoy music. MediaUnbound believes that robust personalization is achieved by blending human music analysts with advanced computer technology.
- Seek out inefficiencies in listener's and world's knowledge - When recommending songs, MediaUnbound strives to understand the listener's boundary of musical knowledge to provide a mixture of known and new material. Human music analysts constantly correct the computer technology by adding selections which are relatively unknown to the general population.
- Build listener trust - Musical preference is an emotional affinity. MediaUnbound research demonstrates that the source of a recommendation is sometimes as important as the music itself. Therefore, the company demonstrates specific knowledge of the music domain and an individual's preferences to the listener. Trust is also built by avoiding false selections of popular music and carefully exposing more cautious listeners to new material.
- Silver-bullet solutions do not exist - There is no single algorithm (or class of algorithms) that can always make the best recommendation. MediaUnbound uses multiple techniques and is not afraid to try new approaches for new problems.
- Sometimes it is better to know what not to recommend than to know what to recommend.
