That’s the Time

Last.fmI’d just like to know: how many of you out there have been burned — or, I guess, delighted — by recommendation engines? You know, when you buy one product, a mercenarily convenient notice will present itself with a recommendation for another similarly minded product that you might like to buy as well., of course, is the most famous online retailer who’s implemented this ‘upsell’ technique for just about anything they sell, but I’m specifically talking about recommendations for music. By and large, I’ve found the recommendations engine at Netflix to be very satisfactory, as there’s something more easily quantifiable about offering up movies than music, tastes for which can be so capriciously subjective.

By contrast, I finally decided to give a try, and I’ve been more or less fully dissatisfied with the results so far. I’d heard a lot of good things about the service, which monitors the songs you play in iTunes (and through other computer-centric music playing methods) and presents recommendations based on your listening habits. I’ve been impressed with almost none of the bands that it’s shown me, having already been familiar with most of them or finding the others to be almost universally bland.

I guess I’m feeling particularly burned because, in a fit of optimism when I first started trying the service, I took’s word for it and actually purchased one of the albums that was suggested to me: Rilo Kiley’s “Under the Blacklight.” I don’t know what I was thinking; this band is so boring I almost fell asleep typing out their name. From now on, when I go looking for new tunes, it’s only supercilious, human-penned music criticism for me.

  1. Have you tried Pandora? I find that to be a much better music-recommendation service.

    But I rely a lot on human critics, too.

  2. I fully agree. I only use to track my music (for the hell of it); it’s been entirely, 100% useless to me in terms of recommending new music.

    When it comes to music, I find that the best method for finding new music that you’ll love is to have a close friend who has very similar but slightly broader music tastes than yourself.

  3. I’m mostly of the Statler and Waldorf school when it comes to recommendation services, but I think I might be persuaded to come out of my shell if there was something that more closely resembled a Mechanical Turk.

    I know people say Netflix is great at recommendations, but I just don’t truss it! I’m a fan of rationale, and it seems like a person writing prose can always do that more succinctly, especially compared to such a sophisticated system as Netflix’s! (The list of contributing movies doesn’t really satisfy me.)

    It also seems like actual person could certainly more ably answer the query “Rilo Kiley but more exciting”, even if you did have all of pitchfork’s reviews indexed by excitingness. (I won’t hazard a guess here, though.)

  4. +1 on Pandora.

    A few of the artists I’ve produced have actually sold a few records to new fans who have discovered them through Pandora’s technology.

  5. On I’ve found that I get a more enjoyable stream of music by using the ‘similar artists’ feature and typing in the name of someone fairly obscure that I like. This works better than the personalized radio stuff, and you get a more consistent selection of things. Type in a Scottish guitar band and you get something that sounds like Scottish college radio.

  6. For what it’s worth, Rilo Kiley hasn’t always sounded like this. I highly recommend returning “Under the Blacklight” and trying “Execution of All Things” instead. I don’t know when this relentless drive to over-produce every song came from, but I don’t like it.

  7. Actually, I’ve found to be quite helpful in discovering new bands. The ‘Similar Artists’ feature works great too, I find. Maybe you just need to poke around and find some of the better groups? ‘London Scene’ is one of my favorites…

  8. I prefer the old fashioned methods myself, like talking with friends. Recommendation engines rely on objective criteria, but friends are subjective. When choosing music, art, food, wine, etc. I almost always find that a friend’s subjectivity is a powerful filter.

  9. I discovered lots of great artist through The way I use it, is a rather indirect approach: After a while (~5000 tracks maybe),’s neighbour list is populated with people who have a very similar taste in music. From then on, I just track their “recently listened” lists, and if a new name pops up, I either preview it, or I ask my neighbours directly what they think about this artist and whether or not they would recommend it. In a way, through, I now have access to very personalized human criticism, which is a good thing, because almost every discovery made that way is a sure hit in my ears, the downside might be that I won’t find anything substantially different than the kind of music I’m already listening too. But to sum it up,, works great for me, not because of its automatic suggestions, but because of the network of neighbours and friends.

  10. I found it took me a little while before the results started to get relevant and interesting enough, but now I love it. And the “Similar Artists” radio is definitely a great way to explore music on the site.

  11. I’ll join the “trust your friends” crew. When you know your friends good enough, you know when you’ll (you won’t) like what they (don’t) like.
    That’s why I recently downloaded a 586 track.

    Second way is the cover design. Sometimes it’s nice to pick up records for the sleeve, even if the music is not exactly what you like best.
    That’s why I recently downloaded a Morr Music compilation.

    Third way is the accident, and it’s really exciting. I found an Adam & The Ants LP last week-end, didn’t know shit about him, but why know give it a try?
    That’s why I bought it for 2ђ.

  12. I understand where you’re coming from but perhaps you could give it more of a chance? (I don’t work for, honest). I find typing in an artists name is a really good way of finding new music.

    It took me a while but I really like it now. I’ve recently been typing in the name of musicians who score films, e.g. Thomas Newman, this method has been very fruitful for me.

  13. Most of the new bands I’ve discovered over the past couple of years has been through, but never through their recommendations engine, which i haven’t found much use for. Listening to the “My Neighborhood” radio and then looking up “Similar Artists” of the ones that catch your ear is an excellent way of finding new stuff. I tend to just skip most of what Pandora throws up, but unlike it doesn’t have 3 years of my listening habits to draw on.

  14. Khoi, I think’s recommendation system gets better for you as you listen to more music. That is, it may be your status as a new user that made your recommendations inappropriate.

    I think one of the central problems with music recommendation is that I may be into bands X, Y, and Z because they share C characteristic, but you’re into X, A, and B because they share Q quality. I think it’s nearly impossible to blindly predict which reasons you like a band, and thus the bands that you might also like, but as you built a repertoire of tracks submitted the recommendation algorithm can get a better sense of your taste.

    From what I understand, Pandora uses the Music Genome project to give you recommendations based on the phonic qualities shared by certain songs, but even that’s quirky. I may, for example, like a certain type of melody, but not with a certain type of beat. Pandora wasn’t very good at guessing for me.

    I’ve never actually listened to radio, but after reading your post I tried it out, and I think the “Neighbor Radio,” like Caspar mentioned, ain’t half bad. It collects from the users who listen to the same sort of music you do, not just one band or song.

    My gripe with is how often the submission server goes down and how difficult it is to get my data from the site and onto another one in the way I choose. I do like checking in on my statistics, though.

    Anyway, caveat emptor; that new Rilo Kiley may have just been a lemon. 😉

  15. I have been a Rilo Kiley fan for ages, and the album is overproduced and boring as hell. As Andy said, try “Execution of All Things.” It is quite excellent.

  16. I’ve used Launchcast, which was bought by Yahoo Music a few years ago, for a long time and find their recommendations spot on. I never understood why it didn’t catch on. It pre-dates all these services. I’ve tried them all and always return to Launchcast.

  17. LastFM tends to be really generic with their recommendations. If you want to find better comparisons, look up a band on

  18. I’ve been using Pandora for a couple of years. When you create a new station on Pandora it takes some “tuning” to get it right (thumbs up / thumbs down, like TiVo). I found bands on there that no one else I know has heard of. (I don’t know any real music buffs, tho… at least not with my tastes.) However, once the station is tuned you’ll hear less new stuff (not no new stuff… just less).

    If the objective is to discover new music, the trifecta is obviously best — friends,, and Pandora. If the objective is just to listen to a stream of music that fits your mood, a well-tuned Pandora station works best for me.

  19. I’m with the several other people that like Pandora over Last.FM.

    With Pandora I regularly get those: “How did it know I’d like this song so much?” moments. It’s just great.

    – Mike

  20. To Last.FM’s defense, you could have listened to their stuff for free either through Last.FM or in iTunes store before making the purchase!

    I use but only for scrobbling right now. I haven’t used the recommendations or the radio in a long time. Trying to build up my reputation as a bif DCFC fan, ahaha

  21. I’ve used for quite a while, and their recommendations are pretty poor. But when it comes to music, I find all music recommendation engines to be severely lacking, Amazon included. Music blogs are much more reliable. Find a blogger who has similar tastes and follow along — that’s worked for me.’s best feature is their charting. The rest needs work.

  22. Another nod to Pandora. As Matt pointed out, there are some drawbacks (for instance, it can’t tell the difference between traditional Irish folk music and bluegrass even though they are worlds apart to the human ear), but it works extremely well beyond its annoyances. Plus with the rating system, you can give songs a thumbs-up or -down and the system narrows the results as a result. I have a few stations that only play a certain type of song, from a broad range or artists. I’ve found quite a few very good albums from listening to Pandora.

  23. I tried a while back, too, and I found it to be quite weak. I used it for about a week, grew bored of it, then disabled my account. I had heard lots of good stuff, too, but I guess it just wasn’t for me. Good luck finding a better service.

  24. One thing that does better than any other service I’ve found is to generate a list of local shows by bands I enjoy. I still can’t believe someone (*cough* upcoming *cough*) hasn’t done this better — there’s no way to toss these into yahoo/google/ical calendars, and the ‘who else is attending’ is boring.

    For some reason the music matching on is next to useless. I think it’s because they match against your most popular artists, when they should be matching against your long tail — thus the success of “typing in the name of someone fairly obscure that I like” (or really, leverage network graph theory to find statistically improbable matches). I like Johnny Cash a lot, but so do many others, and it doesn’t say much about my tastes. I need to get recommendations from people who like Steinski, Jedi Mind Tricks /and/ Gorilla Biscuits.

    Really, though, the best reason to use is the third-party listening history visualization. Here’s three and a half years of my musical progression.

    Also: I like Rilo-Kiley-singer Jenny Lewis’ solo album better than any RK album

  25. works well assuming you’ve let it collect enough data. a week or two isn’t much. I’ve been there three years 🙂

  26. I don’t trust friends with music recommendations, since our taste in music is very different. Previewing songs before you buy them is the safest bet. And notice I said songs–Unless we’re talking about a trance copilation, I usually don’t care for entire albums.

  27. If you’re looking for a list of what are the most-played tracks by all users, I suppose’s model is fine. I think one of the key problems is that’s model probably considers a song that’s been listened to by a large set of people with overlapping taste to yours to be a good recommendation, just because there’s a large number of people.

    When they recommend something based on their data, what question are they asking the data? Are they looking for some kind of popularity ranking — which, say, corrects for novelty and looks for the number of plays by individual users over time — or are they just looking for the songs with the highest number of plays within a certain group of users whose tastes overlap with yours? It seems like is doing the latter.

    If they weren’t, they’d probably be better at knowing that “Under The Blacklight” is a terrible album even though Rilo Kiley isn’t a terrible band, and if you asked any of those users for a personal recommendation they’d probably say “The Execution of All Things” or even “More Adventurous”.

  28. I reckon that’s a bit harsh Khol.

    I’m biased because recently I’ve got to know some people at the company, but does get surprisingly good given a reasonable period of time.

    I used it for a bit a year or so ago, then didn’t use it for ages, then went back to it early this year and gave it a good go. Now, it really hits the spot 99% of the time.

    Listening to ‘taste-neighbours’ radio is the best way to discover new stuff I’ve found, just out of interest.

    Given the nature of recommendation engines you really have to give it a chance to get to know you, same as you get to know your friends.

  29. Does use collaborative filtering in their recommendation engine? CF is about as close as you can get to “taking recommendations from friends”, and I’ve found it to be realy great. Amazon’s recommendations, once you’ve bought enough books or music, are excellent because they do not rely on “dumb” criteria such as “other books by the same author”, or “other movies with the same star” (a terrible method of choosing a movie to see), or “other songs from the same time period” (even worse). It’s pure CF, where recommendations are based only on what other people whose patterns closely match yours are interested in.

    It’s very telling that most of the defenders are suggesting to look at the lists of your “taste neighbors”.. this is basically’s users making their own makeshift CF system.

  30. It appears that does use CF. But it bases recommendations on what you’ve listened to, not what you say you like. In any given week, about half the songs I will listen to are likely to not be songs I’d say I love. I’ll save my very favorite music, the stuff I want to give 11 stars to, tpo only listen to occasionally. And I will also listen to new stuff that I ultimately don’t like and will allow to fade away in a few days. And also, I like to shuffle music, which is likely to retrieve my B and C list songs from my collection. Maybe that’s’s fatal flaw. — what you actually listen to doesn’t always have a lot to do with what new stuff you want to hear.

    One thing I love about Netflix and Amazon is that you can seed your profile by simply rating songs/movies/products whether or not you rent, buy, or otherwise transact with the product. After spending 15 minutes with Amazon rating every record it showed me, the CF engine got way smarter.

  31. I haven’t really used that much due to the fact that most of my music listening happens on my iPod and, up until recently, it never kept track of that, but I will agree that recommendations can be a bit out there at times. With that said, it begs to be asked, what do you like to listen to then? Perhaps the collective mass of readers here at Subtraction can be of a better recommendation service than was.

  32. That’s a great point, Omar. Coming up: a post about what’s playing in iTunes on my Mac. I’ll try and get that done in the next few days. Thanks for the suggestion.

  33. Jumping on the “Yay, Pandora!” bandwagon — I like, but I use it as a archival device rather than as a music recommender. Also, the Music Genome Project is pretty damn cool.

  34. Yay for Pandora! I think that must have a bad algorithm, because I’ve had similar driftwood-channel style experiences.

    Yeah, if you want some serious suck-upage, you ought to list your favorite bands/songs, and we all can rec. I know I’ll be trying to prove my undeniable prowess of underground nouveau.


  35. I find that I’m just so particular about music that atuomated services like and Pandora just don’t work for me. I tend to rely on friends for music recommendations because I know them and what they listen to – so those recs often work out for the better. I still listen to the album first before purchasing though (or at least the iTunes previews…)

  36. Perhaps this isn’t the fault of an app like but more of a fault of the music industry in general? I found myself thinking the other day, I haven’t really been excited about a new band since John Peel was around.

    My recommendation would be The National, any of their albums.

  37. does have a lot of great uses. I don’t think you can judge them on one bum album.

    I learn about a lot of new bands that I never would have heard of if not for by using their Neighbor Radio system, and the personal playlist feature is a giant boon.

    And once recommended my own album to me as one I might enjoy, so they can’t be all wrong!

  38. I read zines for reviews of material before I even think of getting an album from a band, or I search for a website and see if they have any audio I can hear. I’ve made bad choices in the past when someone has recommended a band to me. So I don’t really follow anyone’s advice, unless it’s a “trusted source”.

    But regardless about what anyone thinks about Last.FM, it’s one of the best ideas ever! I love that site and support it fully. 🙂

  39. I agree with both Yuda and Sage’s remarks. has limited use to me in the automatic recommendation of music but can prove useful otherwise. Finding musically compatible friends and dabbling into their music list can reveal some cool stuff. Pandora is quite good at helping you find similar artists/styles by doing it’s best to learn your musical tastes. Through creation of channels, which you yourself can fine-tune via an easy click feedback system, one can bury disliked music/artist and focus on more of the favored. Unlike’s radio whereby you end up stumbling much too often onto music that had nothing to do with your initial choice plus it’s poor method of filtering. In the end nothing is perfect. I use a combination of, Pandora, and Discogs for tracking, seeking, and cataloging mah beatz nukkah.

  40. Recommendations is always a difficult one to get right, but I think its also important to realize that you need to be part of the process – if something which is suggest is something you don’t want, like or care about, then use the feedback mechanism in place.

    Its the same for the likes of Amazon and its product suggestions – initially, they are very generic, often throwing oddball suggestions at you, but over time its amazing how finely tuned to your tastes that they become.

    Give LastFM a bit of time to monitor what you listen do, then give the recommendations a whirl.

  41. I usually skip most of the recommendation stuff, and instead use the similar artists or genre tags. Works much better, and has let to many great finds.

    Thanks for reminding me how bad the recommendations are though, I gave up on them so long ago that I’ve been recommending without putting that big caveat in front of it.

  42. I realize this is old, but I love, so I thought I’d throw my .02 in. I find the automatic recommendations to be crap, since I have highly eclectic taste (lounge music, drum & bass, a little folk, happy hardcore). The ‘similar artists’ is gold, however – putting in artists like Pendulum, John B, or ILS has shown me a variety of awesome bands I never would have known about. I usually find Pandora or Amazon just throw ‘big name’ bands at me – no matter what kind of techno I enter, it recommends the Chemical Brothers, when I don’t like a lot of their music. is willing to stick to genres based on similar artistry and tags, which I really appreciate.

    That said, it’s not for everyone. I think rock music is harder to gauge – honestly, I don’t know much about bands with lyrics. 🙂 But recommendation engines seem to have a much harder time with punk/ indie/ rock stuff than techno.

Thank you! Your remarks have been sent to Khoi.