How does Roon Valence AI work for recommendations?

Hi copying a thread from another topic which I feel deserves it’s own topic.

@Craig_Palmer @danny so how does Valence work ref recommendations?

Here is a pic of Phil Ochs American Troubador album, take a look at the recommendations. Now a cursory glance shows they are both singer-songwriters, and in a way this recommendation surprised me, I had not heard of either, but I would have expected to see eg Joan Baez, Bob Dylan, Pete Seeger, other protest singers, not just USA but perhaps from around the world etc, so very interesting choice of recommendations…

For Miles David Kind of Blue, no surprises, indeed week after week it’s typically the same, in this case why are the recommendations always similar and 50s/60s based, why no suggestions for jazz trumpeters eg from the last decade, from other countries etc

In conclusion I feel this is the area where Roon should invest! :sunglasses:

Valence uses a lot more than just genre because Roon genres (which come from AMG … bought by Rovi…now Tivo) are only at the album level vs track. If that is all Roon used it truly would be terrible. It also is taking into considerations the thumbs up / thumbs down selections users are inputting during roon radio selections. Last.fm is also listed in the credits so it’s also likely using it’s scrobbling capability to help in music similarity algorithms. So at minimum it is using all three aspects in some intelligent fashion. Beyond this who knows…

2 Likes

@brian’s the main architect for Valence, and he wrote about the Roon Radio aspect of it in this thread.

I’m not sure how much he wants to expand on that, but to call it “a simple formula based on genre identification” couldn’t be more wrong.

1 Like

Thanks Danny glad to be corrected :slight_smile:

Would be delighted to hear more from Brian about how Valence works!

Happy new year

Watching this thread with amused interest. Crisps and salted nuts at the ready and binoculars in hand to observe flying pigs.

1 Like

I think the answer is that it doesn’t work very well compared to other algos. I ran a Pandora Station for Art Pepper for an hour and took screen shots of all the recommendations/next songs. A great and logical mix of songs and artists. Roon, after a couple of Art Pepper tracks exits out of Jazz completely and over time makes its way into U2? My kids like Spotify so they can pick songs. I like Pandora so I can get stuff done and listen to music I’d pick if I had the time to build playlists. What is going on with Roon Radio? And if I have to subscribe to Tidal to get what I want, why do I even need Roon? I could just as easily use Pine Player or something cheap to play lists of songs. Maybe this isn’t really a feature we should expect to see get better.

1 Like

Interesting comparisons…how it gets to U2 is baffling…maybe Bono likes jazz hehe

As I understand it, and lord knows my understanding is limited, a machine learning algorithm needs to be trained with a large data-set in order to be meaningful. Now you need your customers to train it (like Twitter allows you to tell it which ads you don’t like), and in terms of data sets, you need access to every musical choice and relevant personal data revealed by the customer, and all your customers.

So if I like music of this “style and pattern” and the algorithm spots similar patterns in its data sets from other users and makes a recommendation?

Question for @brian : how is the training happening (how do I influence the algorithms recommendations), and what is the nature/size of the data set you have access to in order to a make an “effective” recommendation?

Cheers
H

It doesn’t seem that much has changed or improved with Valence regarding Roon Radio in the year since @brian posted about it.

Here’s an example of a rather extreme RR failure. Most of the afternoon RR has been playing very nice selections of folk, singer/songwriter, country, etc. and then out of the blue plays a track by a death metal band Cave State - track Dwell. This is the second time this track has been played in 4 hours. No other tracks would I say were ‘bad’ choices save for this one, twice!

1 Like

Did you “bad thumb” it , ASAIK that’s how it learns , if you just accept it thinks it’s correct. It is a dumb computer after all, until trained …

Not the I am suggesting this is a practical solution, I tend to use RR when I am doing things, cooking, eating etc where a remote isn’t handy or appropriate.

Candle lit RR training has low WAF

I skipped it and marked it ‘doesn’t fit’. No thumbs up or down on iOS.