The CF model helps determine relationships between songs by collecting data from cultural references and social networks. As it tracks playback history, purchase history, user feedback, and other metadata, the model evolves, becoming smarter as more feedback data is collected.
The problem with CF-only modeling is called "cold start." CF requires a critical mass of user feedback before it can return useful recommendations. So any time there's no data or very limited data about a track (like for a new release, an indie track, an artist or song that has limited feedback associated with it), the CF-only model has no way to classify or recommend that track.
Audio Similarity (AS)
Using a machine learning application, the AS model "listens" to each song and analyzes it, generating a kind of thumbprint, or "audio signature" for the song. By analyzing over 100 different sonic characteristics like key, tempo, overall loudness and instrumentation, the model provides recommendations based on similarities in the audio signatures generated.
Unlike CF, AS-only models can generate useful recommendations right away. The problem is that AS modeling only knows that two tracks sound similar. AS-only modelling has no information about how two songs relate to each other out there in the real world of music. As a result, AS-only modeling can generate recommendations which have sonic similarities but make no sense from a cultural standpoint.
The One Llama Approach
One Llama uses a combination of CF and AS modeling to generate recommendations. Our model harvests cultural references and social networking data about each track, and listens to the audio using an advanced "virtual ear." The result is a stronger combined logic for all our recommendations.
The One Llama method has the advantage of being able to give intelligent recommendations for new audio tracks immediately while becoming increasingly smarter as additional information is collected about the tracks from playlists, downloads, user feedback, etc.
Additional One Llama Advantages
Multiple seed search: Find songs similar to both track A and track B.
Positive and negative seeds: Find songs like tracks A and B but NOT like track C.
Search on keywords and metadata: A search for songs that are "uplifting" will return songs also tagged as inspiring, happy, bright, etc.
Search on audio fragments: Find results similar to a section of a song, rather than the entire song thumbprint. This function can be used to to generate intelligent audio previews by finding a track's most "memorable" fragments.
Combination of search types: Search can be performed on any combination of title, artist, keywords, audio similarity, and audio fragment.
Smart search with fuzzy text matching: Fixes spelling mistakes, includes "sounds like," and refines search result mapping for greater accuracy.
Model integration: Search model can easily integrate into existing development platforms (SOAP, REST, XML).
Advanced user interface: Customizable user interface design to match existing tools.
Fast model searching: Tested with over 3 million tracks on modest hardware configurations.
Download iTunes Plug-in [ALPHA](Windows only, Mac version coming soon!)Don't have the One Llama iTunes plug-in yet? Download it now to discover and create playlists directly in iTunes.