Surprise VS scikit-learn

I spent the latter half of 2019 trying to build this as a startup. Ultimately I pivoted (now I do newsletter recommendations instead), but if I hadn't made some mistakes I think it could've gotten more traction. Mostly I should've simplified the idea to make it easier to build. If anyone's interested in working on this, here's what I would do:

(But first some background: The way I saw it, you can split music recommendation into two tasks: (1) picking a song you already know that should be played right now, and (2) picking a new song you've never heard of before. (Music recommendation is unique in this way since in most other domains there isn't much value in re-recommending items). I think #1 is more important, and if you nail that, you can do a so-so job of #2 and still have a good system.)

Make a website that imports your Last.fm history. Organize the history into sessions (say, groups of listen events with a >= 30 minute gap in between). Feed those sessions into a collaborative filtering library like Surprise[1], as a CSV of `, , 1` (1 being a rating--in this case we only have positive ratings). Then make some UI that lets people create and export playlists. e.g. I pick a couple seed songs from my listening history, then the app uses Surprise to suggest more songs. Present a list of 10 songs at a time. Click a song to add it, and have a "skip all" button that gets a new list of songs. Save these interactions as ratings--e.g. if I skip a song, that's a -1 rating for this playlist. For some percentage of the suggestions (20% by default? Make it configurable), use Last.fm's or Spotify's API to pick a new song not in your history, based on the songs in the current playlist. Also sometimes include songs that were added to the playlist previously--if you skip them, they get removed from the playlist. Then you can spend a couple minutes every week refreshing your playlists. Export the playlists to Spotify/Apple Music/whatever.

As you get more users, you can do "regular" collaborative filtering (i.e. with different users) to recommend new songs instead of relying on external APIs. There are probably lots of other things you could do too--e.g. scrape wikipedia to figure out what artists have done collaborations or something. In general I think the right approach is to build a model for artist similarity rather than individual song similarity. At recommendation time, you pick an artist and then suggest their top songs (and sometimes pick an artist already in the user's history, and suggest songs they haven't heard yet--that's even easier).

This is the simplest thing I can think of that would solve my "I love music but I listen to the same old songs everyday because I'm busy and don't want to futz around with curating my music library" problem. You wouldn't have to waste time building a crappy custom music app, and users won't have to use said crappy custom music app (speaking from personal experience...). You wouldn't have to deal with music rights or integrating with Spotify/Apple Music since you're not actually playing any music.

If you want to go further with it, you could get traction first and then launch your own streaming service or something. (Reminds me a bit of Readwise starting with just highlights and then launching their own reader recently). I think it'd be neat to make an indie streaming service--kind of like Bandcamp but with an algorithm to help you find the good stuff. Let users upload and listen to their own MP3s so it can still work with popular music. Of course it'd be nicer for users in the short term if you just made deals with the big record labels, however this would help you not end up in Spotify's position of pivoting to podcasts so you can get out of paying record labels. And then maybe in a few decades all the good music won't be on the big labels anyway :).

Anyway if anyone is remotely interested in building something like this, I'll be your first user. I really need it. Otherwise I'll probably build it myself at some point in the next year or two as a side project.

[1] http://surpriselib.com/

I'm planning to build a business on this, so probably won't open-source it--but I'm always looking for interesting things to write about! I write a weekly newsletter called Future of Discovery[1]; I might write up some more implementation details there in a week or two. In the mean time, most of the heavy lifting is done by the Surprise python lib[2]. It's pretty easy to play around with, just give it a csv of , , and then you can start making rating predictions. Also fastText[3] is easy to mess around with too. Most of the code I've written just layers things on top of that, e.g. to handle exploration-vs-exploitation as discussed in another thread here.

Recently I've been factoring out the ML code into a separate recommendation service so it can different kinds of apps (I just barely made this essay recommender system[4] start using it for example).

I'm happy to chat about recommender systems also if you like, email's in my profile.

[1] https://findka.com

[2] http://surpriselib.com/

[3] https://fasttext.cc/

[4] https://essays.findka.com

Perhaps this would be of interest to you: NOCATS

Are you using scikit-learn for your training? If so, you may try running the models on one another. If you're using custom kernels, you may want to use a different set of them for the test set.

Scikit-learn

In short, you can read through the code, it's open-source. For instance, you can find LogisticRegression here.

http://scikit-learn.org/ - Machine learning with Python https://www.tensorflow.org/ - Deep learning with Python https://www.djangoproject.com/ - https://www.python.org/dev/peps/pep-0008

My only gripe is that the Labs are in R and not Python, but honestly the [scikit-learn](https://scikit-learn.org/) user guide & docs have been straightforward enough to apply the same knowledge in Python for me with some trial and error.

I'd say, the first question you'd need to ask yourself is "Why do I want to do statistical tests" and "what kind of statistical tests do I want to do?". Most of them rely on a bunch of assumptions and just winging it will produce a number that will be reported and used but is terribly wrong. Funnily enough, scikit-learn does not directly give you p-values for this very reason and advise you to run the same regression in statsmodels.

Machine learning and statistical analysis? http://scikit-learn.org

Our next step is to create a new machine learning model based on this list. We’ll use Python’s excellent scikit learn framework to build our model. We’ll store our training data into two data frames: one for the set of features to train in and the second with the desired class labels. We’ll then split our dataset into 70% training data and 30% test data.