Hey pclark, Feedscrub co-founder here. We're still waiting to see what our conversion numbers will look like. I imagine it'll take users a few days to train the filter and see how it works with their feeds. We're thrilled to have so many new trial users already!
"It filters out the posts that are least interesting to you and leaves only the stuff that is relevant."
How well does this service work in practice? I imagine telling spam and non-spam emails apart is far easier then "predicting" which blog post I will like.
Hi Maro, I invite you to try it out yourself. I've released 500 "hackernews" invites so everyone can give it a try.
You're right, it's a lot easier to filter out spam emails with the word "V1AGRA" than it is to filter out news posts. You'll want to take some time to train the filter properly before you unsubscribe from your junk feed.
Is there anything fancier than a naive Bayes classifier going on?
Do you do any calculus on the relative values of false positives and false negatives? It seems that a person with a large number of feeds would put a greater importance on limiting the number of false positives (non-interesting messages classified as interesting by your program), whereas a person with a small number of feeds would be more worried about false negatives (interesting stories classified as uninteresting).
Also, see this comment in this thread: http://news.ycombinator.com/item?id=435837 . For something like this, I doubt a machine learning algorithm will work as well as social algorithms. Web search is a very similar problem (find the most interesting link given input from the user) and social algorithms won out. Of course, the kind and quality of input that you are getting from the user is different. However, your problem is much less well-defined (finding interesting things in general instead of interesting things related to a particular topic).
Interesting points! We are using a naive Bayes classifier right now, and we're working with some social aspects, especially as they relate to Latent Semantic Analysis. We've considered doing a mini Netflix Prize for further improving our algorithm :-).
We've also done a lot of tweaking to get the current filter as smart as possible (thanks PG for your thoughts in A Plan for Spam!) Tweaked things like the number of interesting words to count, weighting of title vs body, etc.
We'd love to talk with some machine learning experts, drop me a note if you are one!
I'm not an expert, just interested in the subject.
I thought of an idea, and this is completely fresh so take it with a grain of salt, but you could use social data from all users subscribed to the same feed to modify the prior probability that any given post was interesting, and then use each individual's data to modify the conditional posterior probability terms.
I'm not sure how effective it would be, but it would create a hybrid sort of classifier that might be interesting.
I'm studying natural language processing this quarter, so I'll let you know in three months if I come up with something more clever.
I think it's an interesting idea. I track about 200 feeds, but probably look at only a small percentage of the stories. I definitely would appreciate help in blocking out the least interesting stories.
That said, limiting the free account to only 3 feeds makes it difficult to judge the site's usefulness. I guess that I would probably want to try the site out with at least 50 feeds to see if I wanted to pay for the ability to track even more feeds.
Consider that it may be more valuable to give out fewer invitations but raise with numbers of free feeds per user.
PostRank isn't personalized, which is the big difference with Feedscrub. FS gives you unique results depending on what you like, not just based on how much people are talking about something.
Out of curiosity how many feeds do others subscribe to?
This post (http://news.ycombinator.com/item?id=248623) suggest you, soundsop, might be on the high end for HN readers--that's a _good thing_, you're our target market. Are there a handful of feeds that have the lowest signal-to-noise ratio? Maybe Feedscrub could help you with those in particular?
just some unrelated feedback about how i have been dealing with this feed information overload issue for now: i use feedly which utilizes among other things google reader friends.. so pretty much i rely on hundreds of other people to decide for me what is interesting and i check the stories they upvote/recommend/share
16 comments
[ 3.0 ms ] story [ 62.3 ms ] threadCouldn't find the pricing page on their website...
Trial accounts limited to 3 feeds to help scale for the launch. Very soon we'll be releasing a way for users to earn more free feeds.
How well does this service work in practice? I imagine telling spam and non-spam emails apart is far easier then "predicting" which blog post I will like.
You're right, it's a lot easier to filter out spam emails with the word "V1AGRA" than it is to filter out news posts. You'll want to take some time to train the filter properly before you unsubscribe from your junk feed.
Do you do any calculus on the relative values of false positives and false negatives? It seems that a person with a large number of feeds would put a greater importance on limiting the number of false positives (non-interesting messages classified as interesting by your program), whereas a person with a small number of feeds would be more worried about false negatives (interesting stories classified as uninteresting).
Also, see this comment in this thread: http://news.ycombinator.com/item?id=435837 . For something like this, I doubt a machine learning algorithm will work as well as social algorithms. Web search is a very similar problem (find the most interesting link given input from the user) and social algorithms won out. Of course, the kind and quality of input that you are getting from the user is different. However, your problem is much less well-defined (finding interesting things in general instead of interesting things related to a particular topic).
We've also done a lot of tweaking to get the current filter as smart as possible (thanks PG for your thoughts in A Plan for Spam!) Tweaked things like the number of interesting words to count, weighting of title vs body, etc.
We'd love to talk with some machine learning experts, drop me a note if you are one!
I thought of an idea, and this is completely fresh so take it with a grain of salt, but you could use social data from all users subscribed to the same feed to modify the prior probability that any given post was interesting, and then use each individual's data to modify the conditional posterior probability terms.
I'm not sure how effective it would be, but it would create a hybrid sort of classifier that might be interesting.
I'm studying natural language processing this quarter, so I'll let you know in three months if I come up with something more clever.
That said, limiting the free account to only 3 feeds makes it difficult to judge the site's usefulness. I guess that I would probably want to try the site out with at least 50 feeds to see if I wanted to pay for the ability to track even more feeds.
Consider that it may be more valuable to give out fewer invitations but raise with numbers of free feeds per user.
Out of curiosity how many feeds do others subscribe to?
This post (http://news.ycombinator.com/item?id=248623) suggest you, soundsop, might be on the high end for HN readers--that's a _good thing_, you're our target market. Are there a handful of feeds that have the lowest signal-to-noise ratio? Maybe Feedscrub could help you with those in particular?