your streaming music experience

[V Knid esq]

One note of caution: making recommendations work is basically the holy grail. I know a lot of people who have put untold time, money, blood, sweat and tears into making both algorithms AND expert-selector models into something that people will really pay attention to, but nothing has yet risen above the mass of other startups.

[CV]

Originally Posted by kol

http://flattr.com

Could this be a step in the right direction?

Great idea!

[Hectic Cum Balthazar]

Yeah auto-recomendations are wasted effort IMO, I don't think it's impossible, I think it's just a Google or Facebook level task and even they can't get it right yet. For a smaller company I think it falls into the 'fiddling while rome burns' category.

Lovefilm for example have coming up to 8 years data on my rentals (and there have been a lot), they have a database that is the fraction of the size of big music retailers and their recommendations are just awful. It's not even the "HEY YOU SEEM TO LOVE TARKOVSKY SO YOU'LL JUST ADORE THE LATEST BEN STILLER ROMCOM" style.. It actually seems like their terrible recommendations are based off some complicated but utterly flawed algorhythm.

The boomkat style "people who bought this record also bought..." actually puts me off buying things sometimes

[wheezer]

Originally Posted by CV

5- ... a need to filter ALOT of noise from these 15 million song catalogues. For me, these massive catalogues are like an endless "carboot sale" representation of recorded music output... endless drudgery.

Is the solution recommendation, by hand, or by creatively-designed "automagic" algorithms (which are cool only in what they *do* then you interact with them, not what they are) ?

In my experience, a dedicated "crate digger" does rapid identifying of the gems in the crates - they do that without recommendation and without metal detectors. Respect to the expertise of selector! What an amazing talent.

Realistically, I think that you want both. Algorithms are extremely poor at achieving serendipity, while they are pretty good at identifying adjacent relations. Your algorithm should get better at predicting neighbors as it is 'trained', i.e. fed loads of data generated by your user community. But your data set will be small at the start and even very large data sets still don't seem to be that great at recommending me something 'out of the blue' that isn't closely related to something that I already know.

To give a concrete example:
On my last.fm homepage, I am being recommended:

- Senking
Similar to Alva Noto, Mika Vainio & ANBB. I've never actually heard of this artist before, but I can get a pretty good idea what kind of music to expect based on what last.fm thinks it is similar to.

- The Soft Pink Truth
Similar to Matmos, Secret Mommy & Felix Kubin. Well, I even have Soft Pink Truth on vinyl so this is a useless recommendation to me.

- Vladiaslav Delay
Similar to Mouse on Mars, Pole, Alva Noto. Really, his music isn't that similar to MoM, but I get why last.fm thinks that there is a similarity between these 4 artists. This is a true recommendation on consensus of status in the 'electronic music scene' rather than on actual musical similarity. Plus I already know this artist, so nothing new gained.

- Martin Denny
Similar to Les Baxter, Yma Surmac and Esquivel. I have been exploring this lounge/exotica stuff a little while now, so I can see why it would recommend Martin Denny to me. Semi-useful because I hadn't taken the time to track him down yet.

Compare this to, say, Marolo turning me on to the music of Joselito, which had no clear relation to anything I was listening to or we were talking about - yet a real discovery for me in a kind of music I never would have ventured to explore. Hence: serendipity.

[Orang Utan]

this site looks pretty decent for streaming: www.grooveshark.com

[Spandex]

Originally Posted by Hectic Cum Balthazar

Yeah auto-recomendations are wasted effort IMO, I don't think it's impossible, I think it's just a Google or Facebook level task and even they can't get it right yet. For a smaller company I think it falls into the 'fiddling while rome burns' category.

Lovefilm for example have coming up to 8 years data on my rentals (and there have been a lot), they have a database that is the fraction of the size of big music retailers and their recommendations are just awful. It's not even the "HEY YOU SEEM TO LOVE TARKOVSKY SO YOU'LL JUST ADORE THE LATEST BEN STILLER ROMCOM" style.. It actually seems like their terrible recommendations are based off some complicated but utterly flawed algorhythm.

The boomkat style "people who bought this record also bought..." actually puts me off buying things sometimes

Yep, it's one of those things where you can chuck maths at it and generate plausible nonsense - like the current gen of Turing Test machines. They sound right 80% of the time but the other 20% is total nonsense cos, well, cos they fail the Turing Test That idea is fucking deep I tell thee.

I'm sure there's a thousand brilliant computer scientists imagining they can write the program that'll get that last 20% right - but that's cos they simply haven't understood the nature of the problem. It's not a tidy domain like chess or even medicine. You might as well ask a computer to review an album - or ask me to recommend you a Vietnamese novel. Best either of us could do would be to give you the "average answer" after surveying the data.

[Orang Utan]

The Sorrow Of War by Bao Ninh is Bender's favourite book I'll have you know

[Hectic Cum Balthazar]

Originally Posted by Spandex

I'm sure there's a thousand brilliant computer scientists imagining they can write the program that'll get that last 20% right - but that's cos they simply haven't understood the nature of the problem.

Aye, it's that classic problem where the people who have the brains capable of solving the problem genuinely don't understand why the other carbon based lifeforms don't love Season 38 of Babylon5.

[CV]

@Spandex: ¡'ve met a lot of people who wouldnt pass the Turing test

I'm interested in the semantic web initiative - knowledge modelling /ontology engineering using machine readable models, accesed across the internet. Thats a seriously big neural net for computers to navigate, once everyone agrees standards instead of this no-standard open API nonsense.

FYI, my period of work at the startup lasted 4 weeks - they were dargged down by pretty much everything outlined in this thread ... back to touring and looking for a decent day job ...

Orang you mention grooveshark... they just got sued by the majors, and left with no major label content. - thats what happens if you sign a deal with the devil