I’m really amazed at the combination of the Music Genome Project and my iPhone. Pandora Radio is a free service that allows you to stream customized music “stations” to your computer or your iPhone. Today, I was able to stream it to my iPhone and pipe it into my car through the Auxiliary plug. By selecting a specific song or genre, it is able to use its’ information on similar other songs and create an unlimited series of songs similar to the one you like. Better still, you can have it mix across channels. And any new channels you create on either your iPhone or computer are automatically available on all your devices.
As each song plays, you’re able to give it a “thumbs up” or “thumbs down” to indicate whether or not you like the song. Favorable and unfavorable ratings help the system adjust future selections to be in the same general domain as the original song, but will be more similar to those you liked. These binary or dichotomous selections can be considered the most basic form of data useful for productive Rasch Measurement. It’s an excellent example of why the Rasch model has so much potential in the Web 2.0 world where people are constantly sharing and rating things on Facebook, LinkedIn, Amazon.com and other social websites.
The current version is useful, but appears to be uninformed by psychometrics. It is based on ratings of over 400 attributes to describe songs, and an engineering algorithm to organize them. Each song is represented by a list of attributes. Wikipedia’s entry describes relatively atheoretical approach used to parse the attributes into “genes”, and then recommend similar alternatives to create a “channel”. The firm has several patents, both for the algorithm, playlist technology and interface on the technology.
John Michael Linacre
But apparently 10% of songs are rated by more than one “technician” to try to improve reliability. From the description on Wikipedia, it appears the firm isn’t using it’s data to the fullest. Pandora has over 2 million users, as of December 2, 2008. First, even if there were only 500 or so ratings of each song, or genre, the users “thumbs up” and “thumbs down” would be excellent fodder to analyze with Rasch Measurement software that could further improve both song and listener measurement, on the same “ruler” of interest. Second, whether technicians or listeners, some people are likely to be easy thumbsters and others hard. This leniency and severity bias in ratings can be adjusted with Mike Linacre’s Many Facet Rasch Model (MFRM) method. Linacre also happens to write the two most popular Rasch analysis programs, Winsteps and Facets. Highly robust to missing data, this approach could really help Pandora improve assessment of both the songs, Genres, and listeners simultaneously – especially with a form of Computer-Adaptive Testing.
While Pandora Radio is a fantastic overall offerring today, my only other suggestion for improvement is to allow it to be used outside the USA. When I was in India a few months ago, I learned the hard way that Pandora’s geolocation software could detect that my IP address was abroad, and refused to send me my normal stations, until I returned to the states. I didn’t try US-based proxies to try to still connect, but hopefully, both Rasch Measurement and Global Licensing will be future features they’ll be adding in the future.
The Scientific Leader Blog 