5 ms·

There's a vague RFS3 idea that's been festering in my brain for a few months, and since I have a cofounder problem (i.e. I don't have one) I'd love it if someon

There's a vague RFS3 idea that's been festering in my brain for a few months, and since I have a cofounder problem (i.e. I don't have one) I'd love it if someone with more time would implement this: My problem with Twitter is that I can only select information by author. A lot of stuff people say is uninteresting to me, but I follow anyway because some of it is good. As a corollary, finding new people to follow is tricky. There are the search feeds, and while some topics are easily filtered by a couple of keywords, most of the time this becomes too complex to keep up. I'd like to be able to train a classifier for stuff I'm interested in, weighted by whether the person is my friend, etc. New tweets would be given a score (possibly in more than one dimension) and presented accordingly.

neurotech1 17y ago
IMHO This is a good idea. One reason why I don't use twitter much is that a lot of tweets are uninteresting. Perhaps a Google-style Markov algorithm could work for filtering a feed.
- jlees 17y ago
  Markov? The problem with filtering is that defining interesting is practically NP-complete.
  - neurotech1 17y ago
    I am not at the level of Sergey Brin by any means, but if the its parent feeds are given a value based on how many people follow them, relative to others of similar keywords, there would be a mathematical way to filter them. I could be completely wrong.. but its out there.
benreesman 17y ago
you're idea reminded me of this blog post: http://blog.directededge.com/2009/10/23/facebooks-news-feed-the-beginning-of-a-recommendations-dominated-web/ http://blog.directededge.com/2009/10/23/facebooks-news-feed-... i certainly agree that increasingly, software fortunes will be made by applying ever more intelligent and adaptive criteria to the information that actually makes it to an eyeball.
tialys 17y ago
I'm actually doing something in this neighborhood for a project for a class I'm in. Basically, I'm just looking at tweets in general and trying to classify them as reviews -- e.g. I loved ____ it's great. I had originally intended to make a trainable 'spam' filter for tweets, but it proved to be too large a project for the assignment timeline. Would love to hear more from you. dave @ davelyon . net
- jlees 17y ago
  A review-relevance filter was part of the engine I built for http://festbuzz.com http://festbuzz.com - I can't really give you code but if I can be of help, hit me up.
jlees 17y ago
I started sketching out a system very similar to this a while ago - I wouldn't be surprised if plenty of Twitter-using hackers have had the same thought, TBH. The problem is, as you've correctly identified, time! As well as keyword filtering and friend bonus there are a ton of things you can pull into the mix. I was actually working on two ideas, one to tell you whether you should follow someone based on their twitter fingerprint and your preferences, and one to pull out interesting stuff from the stream. As I replied further down this thread, interesting is just damn hard. Even trying to concretely identify what I found interesting such that it could be expressed as a grammar or decision tree or set of regexps or whatever... yeah, never really got round to it. I'd keep realising new properties of tweets that made them interesting to me. It's a fascinating problem, though, and very closely related to the "Oh dear gods I have ten thousand unread RSS items" one - that I still haven't solved, nor seen an acceptable solution for! We're drilling down into sentiment right now, but that's not a solution in and of itself; it's an interesting filter though, in terms of only showing stuff with extreme sentiment or stuff that deviates from an individual's normal. It's almost a boringness anti-filter. Kinda works, but is in no fit state to release - again, due to time!

neurotech1 17y ago

IMHO This is a good idea. One reason why I don't use twitter much is that a lot of tweets are uninteresting. Perhaps a Google-style Markov algorithm could work for filtering a feed.

jlees 17y ago
Markov? The problem with filtering is that defining interesting is practically NP-complete.
- neurotech1 17y ago
  I am not at the level of Sergey Brin by any means, but if the its parent feeds are given a value based on how many people follow them, relative to others of similar keywords, there would be a mathematical way to filter them. I could be completely wrong.. but its out there.

benreesman 17y ago

you're idea reminded me of this blog post: http://blog.directededge.com/2009/10/23/facebooks-news-feed-the-beginning-of-a-recommendations-dominated-web/ http://blog.directededge.com/2009/10/23/facebooks-news-feed-... i certainly agree that increasingly, software fortunes will be made by applying ever more intelligent and adaptive criteria to the information that actually makes it to an eyeball.

tialys 17y ago

I'm actually doing something in this neighborhood for a project for a class I'm in. Basically, I'm just looking at tweets in general and trying to classify them as reviews -- e.g. I loved ____ it's great. I had originally intended to make a trainable 'spam' filter for tweets, but it proved to be too large a project for the assignment timeline. Would love to hear more from you. dave @ davelyon . net

jlees 17y ago
A review-relevance filter was part of the engine I built for http://festbuzz.com http://festbuzz.com - I can't really give you code but if I can be of help, hit me up.

jlees 17y ago

I started sketching out a system very similar to this a while ago - I wouldn't be surprised if plenty of Twitter-using hackers have had the same thought, TBH. The problem is, as you've correctly identified, time! As well as keyword filtering and friend bonus there are a ton of things you can pull into the mix. I was actually working on two ideas, one to tell you whether you should follow someone based on their twitter fingerprint and your preferences, and one to pull out interesting stuff from the stream. As I replied further down this thread, interesting is just damn hard. Even trying to concretely identify what I found interesting such that it could be expressed as a grammar or decision tree or set of regexps or whatever... yeah, never really got round to it. I'd keep realising new properties of tweets that made them interesting to me. It's a fascinating problem, though, and very closely related to the "Oh dear gods I have ten thousand unread RSS items" one - that I still haven't solved, nor seen an acceptable solution for! We're drilling down into sentiment right now, but that's not a solution in and of itself; it's an interesting filter though, in terms of only showing stuff with extreme sentiment or stuff that deviates from an individual's normal. It's almost a boringness anti-filter. Kinda works, but is in no fit state to release - again, due to time!