I don't have a CS or math background, so my understanding of these concepts is relatively limited.
I'm working on an idea for a recommendation website/service, and I'm trying to figure out first if it's feasible, and if it is, what types of people should I be looking for to help build it, and what the most practical approach might be. Without getting too specific (yet), I'm thinking of something along the lines of Pandora, Netflix, and Amazon.
From what I understand, Pandora actually has real people listening to every song and choosing (from a predefined list) which attributes best characterize it. When you "like" or "dislike" a song or artist, it uses that data to make future recommendations. However, others such as MusicBrainz/MusicIP use acoustic fingerprinting and feature extraction to gather data.
What are some of the advantages and disadvantages of each method, and when might you use one over the other?
How about when applied to other multimedia, such as images and video?
Does the challenge lie more in figuring out how to choose meaningful attributes, the learning algorithm itself, or both equally?
What are the advantages and disadvantages of a top-down method (pre-defined attributes typically hidden from the enduser) versus a bottom-up approach (something along the lines of tagging or user-defined attributes)? The most sucessful recommendation algorithms seem to use the former method, is there a reason for that?
Any insight would be very much appreciated.
[link] [1 comment]