MediaUnbound’s media recommendations are a hybrid between human-driven media psychology and machine learning technology.

Instead of utilizing a single algorithm or algorithm-class, we have created multiple techniques to overcome traditional flaws in recommendation technology and solve distinct music-related problems. Each technique sits as a module on top of the MediaUnbound personalization platform, allowing us to mix-and-match on the application- and user-level. Each technique additionally has a number of tweakable parameters to allow flexibility in the range and type of possible outputs.

Some of our patent-pending machine technologies include:

• Rankshift^TM: A robust, statistical technique used to analyze large sets of usage data to find patterns of association. MediaUnbound developed this technique to overcome the popularity effect present in many collaborative filtering algorithms which overemphasize popular elements in recommendations. Rankshift^TM is particularly useful on datasets that have 5,000 to 200,000 elements distributed among 10,000 or more user collections. The technology works best on domains which feature a Zipf-like distribution (i.e. power law probability distribution).
• N-dimensional clustering analysis: A method by which a domain is factored into an N-dimensional vector space to calculate a global similarity metric. This technique is often used on datasets that have over 200,000 elements distributed amongst 10,000 or more user collections.
• The MediaUnbound Matcher^TM: A highly customizable fuzzy lexical matcher. This module is used to match content collections to MediaUnbound’s canonical catalog. The Matcher can be optimized to run at different speed or accuracy levels depending on the application and content set.
• Profile Cruncher/Expander^TM: A method to take large amounts of raw usage data and create a lightweight listener profile for real-time calculation and storage within a production environment. The Expander allows for use of the entire data history in offline and more complex applications.
• Playlist Engine: A sophisticated rules engine to allow creation of dynamic playlists which meet customer-defined arrangement and selection rules. We apply completely different approaches in making playlist recommendations to adventurous listeners vs. more mainstream listeners.