Automagical suggestions..

[nikos]

Ok, I know this is a hard time for the photo.net machinery and

Brian's probably busy getting the standard functionality streamlined

again, but I had this idea and I thought I'd set it under

consideration for the future:

<p>

The rating system is mostly a <i>community-driven filtering

mechanism</i> that delivers a good selection of photographs to look

at. As it is now, it works on a principle of <i>blind aggregation</i>

yielding <i>suggestions out of a non-weighted public vote</i>. It

does a decent job at it, but it would be more interesting to have it

<b>suggest photos based on match of user tastes</b>. How? Two ways:

<p>

1. Some nifty (but maybe heavy) SQL should be able to provide, for

each photo in the gallery, <b>a list of links to photographs rated

high by the same users who rated this photo</b>. So when I rate a

photo above my average rating, PN should be able to tell me: "you

liked this photo, so you should also like THESE photos, rated high by

those who also liked it"

<p>

2. More nifty SQL should be able to <b>compare my ratings against

other users' ratings and find one or more users with similar ratings

to mine</b>. If out of 500 ratings, some other user has rated 100

photos with similar values as I have, we probably have similar taste,

so I should be told to look at his top-rated page. There are

information retrieval algorithms that could be used to make this very

accurate.

<p>

As I said, this is not a first priority, but if implemented, it would

be a great resource, and it would help people spot more photos in the

style that interest them. It would really be harvesting the power of

the community.

<p>

Brian, if you ever consider an implementation, I can help either by

coding the IR engine or suggesting algorithms and formulae. It should

be too heavy for online computation, but with the right data exported

and periodically computed it could be possible. I'm willing to spend

some time contributing to a feature like this.

[briany]

Nikos, I too would be interested in contributing to an idea like this with some coding and algorithm work. You've set some ideas spinning around in my head, but I won't clutter up this thread with them; maybe a future discussion. Personally I'm not sure how much I'd like seeing a lot of stuff similar to what I'm "comfortable" with. I suppose a Lot of the final effect would come down to the algorithmic details. It would certainly be an interesting experiment, and one might not need access to the photo.net database to do it, though it would require extensive crawling (without image downloads though, should have minimal impact if the bot is well-behaved). So much of Photo.net was done by Phil's MIT students as final projects, perhaps this idea could be passed along to the proper class through the right channels...

[cd]

I support this suggestion entirely.

[nikos]

I have considered the crawling option, but even with a well-behaved bot I wouldn't do it without Brian's permission, because it's more than just adding to the server's load, it's actually compiling heaps of derived information from his database. But assuming that Brian would be OK with someone working on a project to do something with this information, then obviously HTTP crawling would be far less efficient than getting an incremental export of selected fields from the users and ratings table by Brian on a regular basis.

 

Regarding the concern of filtering to a narrow content subset of 'things you are comfortable with' I would have to disagree. There are excellent algorithms to weed out common choices. (for example weighting every photo's value as a component of one's preferences by a logarithmic relation to the popularity of the photo, essentially giving less value to common matches of taste) In fact a system like that would not only provide with relatively accurate suggestions in terms of similar taste, but would also let you surface hidden gems in the same way you do by the users' top rated list. The idea is: You can find a lot of hidden gems by a user whose taste you like, and you did it. Now, the system should be able to give you more users with similar taste, the existence of whom you didn't know before. And it could also be able to point you to the photos they have rated and few others have rated. (possible hidden gems)

 

It's not a heavy process in terms of data set size. It's just computationally intensive. But once you get the data set isolated outside the production server, any powerful PC can generate indices in a matter of hours (I guess)

[joe_garrick]

You can already do this by just clicking on the ratings list for the photo (the number of ratings shown is a hyperlink to the list of individual ratings unless there are less than five for the photo). Find people that rated it the way you did and jump to their pages. The other method I use is to scan the comments, looking for people that might have posted remarks similar to my own thoughts about a photo (what worked, what didn't).

 

I find it to be amazingly easy to find work I like on this site, despite the fact that there's more trash than treasure overall.

[nikos]

Sure I can find a person with the same opinion as I have for a single photo and then peruse his ratings. And sure, with time, I can identify people whose opinion I respect.

 

What I'm suggesting is a way to find a person who has rated 100 images with the same rating. A person who has commented on 30% of the images that I have commented too. And I do not want to spend my life clicking through the site to find it. And there is a technical solution to doing that, which is what I am suggesting. It is not an quick and easy one, but feasible and worth considering at some point.

[joe_garrick]

Nikos -

 

I do programming and work with SQL databases every day. What you're suggesting is not overly complicated, but it's apparent to me that it's also probably not imminent. I only posted my comment to present an alternative that - although not as efficient as an automated method - is already available.