Amazon: Death Cab, The Shins
iTunes: Imogen Heap, Feist, M.I.A.
Perceptron: The Knife, Dntel, Bright Eyes
Eric B and Rakim:
Amazon: EPMD, Big Daddy Kane, Public Enemy
iTunes: EPMD, Boogie Down Productions, KRS-One
Perceptron: No recommendations
Fleetwood Mac:
Amazon: Journey, Bob Seger, Steve Miller
iTunes: Rod Stewart, Eric Clapton, The Doobie Brothers
Perceptron: Vetiver, The Beach Boys, The Everly Brothers
Sepultura:
Amazon: Pantera, Slayer, Metallica
iTunes: Soulfly, Machine Head, Fear Factory
Perceptron: Moonspell, Soulfly (x3), Fear Factory
In none of these head-to-heads does Perceptron do the best job (the Dntel pick ties it with Amazon's Death Cab pick). I can intuitively guess that iTunes will do a better job of recommending for me than your website, because millions of people's buying habits make iTunes smarter every day.
It's not just that this space is crowded; it's that it structurally resists your entrance. You need features, lots of them, compelling features, or this isn't going anywhere.
Finally: "astute music recommendations" is a terrible tagline. It doesn't convey value. It tells me why you like your code, but not why your code matters to me.
Thanks very much for doing those side-by-side comparisons. I am not a fan of any of the bands you entered, but, to my ill-informed eye, the recommendations don't seem too bad. Which ones were off the mark?
And thanks for the heads-up on the XSS exploit. Will need to sanitise my URLs more thoroughly.
Postal Service: Indie pop; you did fine, but not better than the majors.
Eric B and Rakim: seminal NYC hip hop; guessing classic hip hop is a blind spot in your data set.
Fleetwood Mac: mainstream classic AOR; Vetiver? An indie folk band? The Beach Boys are similar only as contemporaries --- also, not a very useful recommendation, since there are 3 different Beach Boys your recommendation could be keying on, two of them execrable.
Sepultura: insider-y mainstream metal; you did fine, I guess, but not better than the majors.
In two cases your recommendations are bad, and in two cases they're just not excellent. But if any online music retailer will do this for me, why would I use your service? To listen to some of the music online?
You should consider:
* Allowing people to upload their entire iTunes library files so you can analyze them and make targeted recommendations to stuff they already own.
* Allowing people to vote on the recommendations, or doing something "social" (gag) to let people adopt recommendations or biases from tastemakers who care about this stuff.
* Link to or repurpose content from Wikipedia to give more background for the artists you're displaying.
It still seems like this app is really just 10% of what Last.fm or Pandora already does.
Thank for you the clarification on where the recommendations fell down.
I have resisted letting people upload their record collection because I want to focus on making the site work well when I don't know anything about the user.
Letting users spy on the recommendations of people they admire or feel they are similar to is in the works. Once I get enough users, I plan to have a thing that says, "Hi Jill, your taste is similar to Bill. Would you like to be updated on his recommendations?"
I already pull in the band descriptions from Wikipedia.
Thanks again for the detailed feedback - it's really useful.
It's not just that this space is crowded; it's that it structurally resists your entrance. You need features, lots of them, compelling features, or this isn't going anywhere.
Also, you're trivially XSS-able:
http://theperceptron.com/recommendation/artist/h%3Cscript%3E...('hi');%3C%2fscript%3E?h=true
Finally: "astute music recommendations" is a terrible tagline. It doesn't convey value. It tells me why you like your code, but not why your code matters to me.