Sunday, May 14, 2006

Findory in the Wall Street Journal

In her article "Me, Me, Me", Jessica Mintz at the Wall Street Journal talks about the dream of a personalized newspaper and efforts to apply personalization to news.

Rojo, Newsvine, and Findory are given as examples of "a new generation of Web start-ups" that "help readers deal with the sheer volume of material that's out there" by learning from "the reading habits of their users" and "then [using] that data to make suggestions to individuals based on what others like them are reading."

An excerpt on Findory from the article:
Findory.com ... relies heavily on the behavior of its users, but doesn't require them to list their interests, select feeds or vote on stories.

Instead, it works on the same principle as Amazon.com's recommendations ... Findory ... looks at an individual's reading history, compares it with similar readers' tastes, and offers up links to stories that similar readers have enjoyed ...

Each time the user returns to the Findory home page after clicking on an article, he or she will find the page reconfigured with a different mix of stories.
Just read articles, that's it. Findory learns from the articles you read, adapts to your interests, and builds you a personalized front page. Findory gets better and better the more you read.

6 comments:

Eran Sandler said...

Findory uses the history of what I read as basis for the learning system. This means that under a large data set of what I read it can recommend in a very reasonable way what I would probably want to read.

Eliminating the need for the user to not tell about himself/herself anything lowers the barrier of getting the user to use the system, but the it will take sometime until the user will teach the system what her/she actually wants to read.

There is no visible way a user can say that he/she dislikes a certain topic or recommendation the system offers. Instead the user will probably have to go to topic that he/she does like so that their amount will be large enough to overcome the "bad" data.

I assume this is a "by design" feature, or more accurately, "by methodology" feature, since that is the user experience you want to give. No rating, no giving information up front.

I for example, would have prefered to tell you what I like/know/want to know. Even in the easiest smallest way.

It wouldn't have to be mandatory, simply a way for me to say "I like Arts, Technology and Sports".

I assume this would greatly help the system (and I also assume you are pre-categorizing the feeds which you crawl ;-) ).

Why didn't you go that way? Wouldn't it allow eliminating in a very strong manner the "noise" which can come from wrongful or uninteresting clicks on items that are not relevant?

I know rating is more problematic since it raises the bar for most users so I'm not even talking about that.

Greg Linden said...

Hi, Eran. Thanks for your comments on Findory! Let me take some time to respond.

On the speed of learning, that is a problem with some techniques, but Findory starts learning very quickly, personalizing the site even after someone has read just a couple articles.

On bad recommendations, readers can get an explanation of why an article was recommended by clicking the little sunburst icon next to the title of a recommended article.

As for the cost of mistakes, it isn't clear to me that a mistaken recommendation is any worse than the untargeted, uninteresting, and unpersonalized articles thrown up every day on websites like CNN.com, the vast majority of which are uninteresting to me. Do you think it is?

On making it mandatory to specify a few interesting categories, there's two reasons Findory does not do that. First, it is unnecessary, since we can infer your subject interests from the articles you read. Second, the vast majority of people wouldn't bother doing it.

That being said, this customization approach you suggest is the method used by sites like My Yahoo, Netvibes, Live.com, Bloglines, and many others. Findory is a different approach, helping readers find interesting news without having to provide any explicit information about their interests. Just read articles, that's it. It all just works.

Thanks again, Eran. I appreciate hearing your thoughts on Findory!

Anonymous said...

greg,

this is great. i blogged about findory and the wall street journal several weeks ago. my comment were on personalization, in general, neural networks, and usability. great stuff.

the link to that article can be found here:

http://www.shmula.com/?p=30

Eran Sandler said...

Hi greg, thanks for the long answer (well, my question was quite long as well ;-) ).

There is a difference between sites like CNN.com which shows you whatever the editor thought was interesting, than to have a news personalization site like Findory which is suppose to show you only interesting stuff. In CNN.com you just get whatever it is they show you. In Findory, since its a learning system I will get whatever it is the system thinks I'm interested in.

The main problem I was trying to raise earlier is that since Findory is based on my reading history, if (for some strange reason) I click and view items that I don't find that interesting I cannot tell the system. I can see why it was recommended to me (which is great, btw) but the only way I can "correct" the system (and "correct" being telling the system that I don't like this) is to never go there again using Findory.

So, in fact, a few "wrong clicks" can throw the system way off course for me and to correct that I will have to go more to links that I do find interesting.
This means that clicking on something that I didn't find that interesting in Findory is not like clicking on something not interesting in CNN.com (besides the fact that CNN.com doesn't learning and respond to that because its not a learning system ;-) ).

This also leads me to another thought about personolized systems. My main "problem" with personlizied systems is that I usually want to "exploration" factor. This means, that for some kind of a percentage from the offered content I would like to get things that I never even thought were interesting directly.

I know that not everyone wants that, but for me, a learning system has to have (if opted in) the ability to suggest new interesting things to me that I never thought I would like and after reading them they will become my main interest.

I guess that is one factor that Findory can handle better than "stated systems" (system that you need to state your interested to bootstarp) since there is always the ability to show something to someone and when clicking on the reason to know why this item was shown it will say something like "You have switched the exploration preference on, the system will not try to find new interesting stuff for you".

As far as I could tell in Findory, the sources for crawling are limited and someone is categorizing them manual to some predefined taxonomy and, IMO, it limits the system to get and handle new things that are not hard classified into this taxonomy.

How do you handle that?

p.s. sorry for bugging you too much about that ;-) If you don't want this whole discussion on your blog and you don't mind me poking questions at you we can move this into an Email disscussion :-)

Greg Linden said...

Hi, Eran. Findory does allow readers to delete items from their reading history if they accidentally read something they did not like.

On the exploration, that's the pigeonholing problem, yep. But I think you're making the mistaken assumption that Findory recommends very similar items (e.g. more articles on Iraq if you read an article on Iraq).

Instead, Findory tries to recommend surprising and non-obvious items, things you wouldn't have found easily on your own.

The point of personalization is to aid discovery and enhance serendipity. In well implemented systems, personalization reaches beyond the obvious and into the surprising.

Eran Sandler said...

Hi Greg.

Perhaps I wasn't clear enough. I'm not suggesting that going a lot to articles on Iraq will bring me a lot of Iraqi related items, that's for sure, BUT, it would assume it will strengthen information about politics, economies, foriegn affairs and stuff like that.

Plus, as far as I could tell (though I'm only guessing here) is that you don't do a textual semantic analysis of the text you are suggesting, therefore, if the article on Iraq was in the politics section and was something like "Big oppossion for preseident Bush to stay in Iraq" will not suggest me more Iraq related articles. Instead, it will assume that I'm interested in politics or something like that.