Knewton and Pandora

Here at Knewton, our adaptive learning engine is centered around the idea that education is optimized when students receive exactly the content that they need at exactly the right time. We’re able to achieve this by mapping educational concepts at the atomic level, and adapting students’ learning experience around how they perform on the various tests and assignments.

So we’re intrigued that the atomized, adaptive approach is manifesting itself in a similar way in the music industry. Pandora, for example, generates unique custom playlists on the fly, based on the musical elements that recur in the songs and bands that the user specifies. Just as we have a concept tree to map your proficiency on, for example, right triangles or inferences, Pandora has a musical concept tree with items such as “electronic influences” or “use of ambient sounds.” And just as our algorithm generates questions based on your previous performance on a certain concept, Pandora generates its stations based on the common musical features of the songs you’ve selected.

The New York Times recently gave some insight into Pandora’s process. They analyze music along objective structural dimensions like melody, harmony and rhythm. They break “voice quality” down into descriptors like “smooth,” “gravelly” or “nasal.” They use similarly nuanced tags for every instrument, as well as lyrics and production. Pandora is especially intriguing in its efforts to quantify factors like “emotional intensity.” Here the subjectivity of their taggers inevitably comes into play. The company tries to impose consistency on the human factors using examples, so Tom Waits stands for the maximum “gravelly” vocal score.

Any adaptive system will require a lot of human effort up front creating a knowledge database rich with tags and descriptors that can then interact with the user flexibly. If you want to develop an adaptive science teaching system, first you need to tag all of the science concepts. If you want to build an adaptive recommendation system for restaurants, you first need to tag all of the aspects of a restaurant experience. We know from experience that the creation of all these tags, descriptors, concept trees and other metadata is labor-intensive. It requires a lot of human intelligence and effort up front. Once the metadata is in place, though, an adaptive engine can show remarkable flexibility and a richness of experience for the end user.