3 ms·

I agree with zdw in that it's very much like the Netflix Challenge, in the way of focus. If you don't narrow enough, on specific topics, you let irrelevant dat

by jimminy 16y ago

I agree with zdw in that it's very much like the Netflix Challenge, in the way of focus. If you don't narrow enough, on specific topics, you let irrelevant data in to the results, if you narrow too much, you cut out posts that were tangential but possibly even more relevant. Also to compare this type of tracking to either Netflix or Amazon's recommendations, is hard, because they use a fairly static set of items to analyze, and thus don't suffer from long-tail analysis that would need to occur in a constantly updated set of data. For a basic search recommendation tool, this isn't as hard, but trying to pick up and determine someone's interests, without over or undercutting the topics, in a constantly changing set is very hard. You also have to deal with shifts in user interest, Amazon gets by by only having to recommend for the latest few entries, in most cases. Trying to provide recommendations of new data for a user, you have to be able to be able to refocus to adjustments in the user's interests. Also, as far as the words being a goldmine, not always, are they a great resource, yes, but only slighly more than the source that the user likes. Sources are great for initial narrowing, words are not. Two is not really easier, even with prior art, the prior art is great to look at for inspiration and ideas, but when it comes down to it, it's a very different set of requirements.

jamesbritt 16y ago
"If you don't narrow enough, on specific topics, you let irrelevant data in to the results, if you narrow too much, you cut out posts that were tangential but possibly even more relevant." I like Amazon's recommendation engine, but I'm surprised when I check out suggested future releases and get no more than a single page. I'm interested in a whole lot of things, and I think my purchase history and such would support that, but for whatever reason Amazon sometimes doesn't show me very much. I'm guessing it's a side-effect of also marking some items as "not interested", with the software unable to make reasonable distinctions between likes and dislikes. It's not uncommon, for example, to start getting recommendations for damn near every new DVD release simply because I put a DVD into my wish list. As if liking "a movie" == "liking movies" => "show all movies." But if I then mark "Saw XII" as "not interested" it infers "does not like movies." Or something; it's hard to tell. The absence of metadata in my choices make this quite hard. I like some rock music, but probably dislike most of it. Same for electronica or folk or whatever. Discerning just what I like about this or that artist or piece is hard, even for me. Same goes for books and movies. There are clearly common threads, but deducing them is non-trivial. In many ways I like the same music and art I liked 20 years ago, but I don't always like the same musicians or artists, nor do I always like hearing or seeing things represented or presented in the same way. So, it's the same, but different. For example, I loved "Metal Box"-era Public Image Ltd., but do not want to listen to "Metal Box" forever; I want to know what is the 2011 equivalent of "Metal Box". But that requires understanding current musical norms and conventions, not just the tonal and rhythmic qualities of a particular album. The whole "When are two things the same?" question is at the root of AI, and a seriously tough nut to crack.
- gwern 16y ago
  > I like Amazon's recommendation engine, but I'm surprised when I check out suggested future releases and get no more than a single page. I'm interested in a whole lot of things, and I think my purchase history and such would support that, but for whatever reason Amazon sometimes doesn't show me very much. My theory is that this is evidence that the engine is using http://en.wikipedia.org/wiki/Collaborative_filtering http://en.wikipedia.org/wiki/Collaborative_filtering - the way I think of it is as a giant Venn diagram. At the beginning, I fit in almost every one of the little circles because the service is so ignorant about me. It will recommend me stuff from just about any overlapping circle. But as I select and anti-select items, more circles become invalid, and I slowly narrow myself down into a cramped little circle of just the items I've decided about, until finally no one else's selections overlap with my own and the service can't find anyone similar to steal recommendations from to give to me.
  - jamesbritt 16y ago
    Good observation. Also a bit sad, if true, as it suggests that anyone outside the mainstream will get fewer recommendations. Of course, it's tricky, and I appreciate an approach that takes into account that not everyone who likes Foo will like the same things; other likes and dislikes give important context. But I'd prefer a somewhat larger, if less precise, set of suggestions over a small set.

by jimminy 16y ago

jamesbritt 16y ago

"If you don't narrow enough, on specific topics, you let irrelevant data in to the results, if you narrow too much, you cut out posts that were tangential but possibly even more relevant." I like Amazon's recommendation engine, but I'm surprised when I check out suggested future releases and get no more than a single page. I'm interested in a whole lot of things, and I think my purchase history and such would support that, but for whatever reason Amazon sometimes doesn't show me very much. I'm guessing it's a side-effect of also marking some items as "not interested", with the software unable to make reasonable distinctions between likes and dislikes. It's not uncommon, for example, to start getting recommendations for damn near every new DVD release simply because I put a DVD into my wish list. As if liking "a movie" == "liking movies" => "show all movies." But if I then mark "Saw XII" as "not interested" it infers "does not like movies." Or something; it's hard to tell. The absence of metadata in my choices make this quite hard. I like some rock music, but probably dislike most of it. Same for electronica or folk or whatever. Discerning just what I like about this or that artist or piece is hard, even for me. Same goes for books and movies. There are clearly common threads, but deducing them is non-trivial. In many ways I like the same music and art I liked 20 years ago, but I don't always like the same musicians or artists, nor do I always like hearing or seeing things represented or presented in the same way. So, it's the same, but different. For example, I loved "Metal Box"-era Public Image Ltd., but do not want to listen to "Metal Box" forever; I want to know what is the 2011 equivalent of "Metal Box". But that requires understanding current musical norms and conventions, not just the tonal and rhythmic qualities of a particular album. The whole "When are two things the same?" question is at the root of AI, and a seriously tough nut to crack.

gwern 16y ago
> I like Amazon's recommendation engine, but I'm surprised when I check out suggested future releases and get no more than a single page. I'm interested in a whole lot of things, and I think my purchase history and such would support that, but for whatever reason Amazon sometimes doesn't show me very much. My theory is that this is evidence that the engine is using http://en.wikipedia.org/wiki/Collaborative_filtering http://en.wikipedia.org/wiki/Collaborative_filtering - the way I think of it is as a giant Venn diagram. At the beginning, I fit in almost every one of the little circles because the service is so ignorant about me. It will recommend me stuff from just about any overlapping circle. But as I select and anti-select items, more circles become invalid, and I slowly narrow myself down into a cramped little circle of just the items I've decided about, until finally no one else's selections overlap with my own and the service can't find anyone similar to steal recommendations from to give to me.
- jamesbritt 16y ago
  Good observation. Also a bit sad, if true, as it suggests that anyone outside the mainstream will get fewer recommendations. Of course, it's tricky, and I appreciate an approach that takes into account that not everyone who likes Foo will like the same things; other likes and dislikes give important context. But I'd prefer a somewhat larger, if less precise, set of suggestions over a small set.