Genes, Features and Recommendations, Oh My!
The Music Genome Project was created seven years ago with the goal of establishing a set of characteristics (such as “Intricate Rhythms”, “Acoustic Guitar Riffs”) of music which can be used to compare and contrast music in a systematic way. These characteristics are called “genes” by the founder, Tim Westergren, but I’m going to use the term “features” to describe them. Why? Because features are a classic concept in Machine Learning, the field who’s goal it is to automatically extract information and make predictions based on large sets of data. Tim Westergren’s decision was to establish features, then to go through the laborious process of extracting/identifying in a large body of music was a stroke of genius. With such features in hand, any number of Machine Learning algorithms can be applied to user feedback data in order to classify, cluster, organize, or even to… recommend music. Pandora is Westergren’s venture to do exactly that.
Hand-labeling information about songs has its advantages, but also has its drawbacks. According to Wikipedia, the process of labeling a song takes a musical expert 20-30 minutes of time. Is this overhead necessary? Read/WriteWeb compares Pandora to other recommendation engines, prompting the question, “Can tags become genes?” They note that del.icio.us‘s “related tags” serve as a sort of recommendation. But, is there more that one can do with tags? Our answer to this here at StyleFeeder is an emphatic “Yes!” StyleFeeder is a social shopping bookmarking site, allowing users to keep track of their shopping “finds”. We provide recommendations based on users’ bookmarks and ratings. Recently, we updated our system to include additional user and item information to enhance our recommendations: item tags, as well as user gender, age and sign-up source and selections. Much like Westergren’s “genes”, these “features” (as we call them) allow us to more quickly provide relevant recommendations. They also have a subtle advantage—no hand-labeling by an expert required. Are tags noisy? Sure. But, the large majority of them are accurate and relevant. One challenge of being a product bookmarking site is that, unlike Amazon.com, we don’t have a boat-load of meta-data to go along with each product. But, there are still many ways of gathering information about a product. With each product bookmark, the user provides us with an image and a web page. Both the image and web page text are sources of information about the product. The key is knowing how to extract this information. Automatic feature extraction is a sub-field of Machine Learning and has seen much progress of late. Here at StyleFeeder, we can apply techniques for extracting features from text/web (e.g. Blei/Bagnell/McCallum ’02) and images (e.g. Tieu/Viola ’99). Westergren’s Pandora site may be able to take advantage of techniques developed specially for sound/audio (e.g. Mierswa/Morik ’05). It’s not trivial to apply these approaches, but they show great promise for enhancing a wide range of learning applications, especially recommendation systems.
It sounds fantastic. The question is – this post is absolutely new and original, isn’t it? It seems to me I’ve saw it somewhere before.
Hi Mike,
Yes, this post is original. I should know—I wrote it :) Of course, as any machine learning expert will tell you, the ideas I’m discussing/describing are not. People have been using non-hand-labeled features of various sorts (including text, image and sound) for learning for decades. The exact form of that learning has varied, and recommender systems are are the more recent end of that scale, but these are old ideas (even though there is no end of marketers purporting such ideas as “new” and “original”).
Jason