jump to navigation

Explaining recommendations January 8, 2008

Posted by Andre Vellino in Collaborative filtering, Social networks.
trackback

I heartily agree with Paul Lamere’s comments about making recommendations transparent using explanations. Furthermore, I am convinced that explanations will also provide users with a feedback mechanism to control the recommender’s long-term behaviour. (For a nice survey of explanations in recommender systems, see this paper by Nava Tintarev.)

While Paul’s system of recommendation via social tagging (“tagomendations”) seems to work well for recommending music, I think that for recommending scholarly scientific articles, content analysis techniques (combined with collaborative filtering) are more appropriate. In part this is because the social tagging of, and indeed, the rating of “entertainment” is a form of self-expression. When you rate “Fargo” with 5 stars you are saying something which you want other people to know.

I can also see doing this with books (novels / biographies etc), but not with scholarly articles. For those items, tagging and rating are, at best, a behaviour that users learn will help the automated recommender system do a better job. That’s why I think implicit ratings are more likely to be of value in a scholarly digital library.

Comments»

1. lemire - January 8, 2008

There is no question about it. Implicit ratings are more appropriate for the personal space of a researcher, say.

I do not think it is sensible to expect students or researchers to rate papers in a consistent way.

What you have to leverage is the fact that most researchers live (or should live) in niches. This is a challenge: the researchers know their field well and they are unlikely to be impressed by run-of-the-mill recommendations. It is also an opportunity: you can “learn” pretty accurately what a researcher what to learn about.

Here is a fun application I would like to see:
– extract the reference sections of the papers I have written;
– add to the lot every paper I have written.
Then, tell me about any paper that cites these papers. Allow me to interact with the view: can I see only the authors I have cited before? Naturally, the more often I have cited an author, the more likely a recent paper by this author would be of interest to me… as long as you can determine whether the paper is “on topic” or not.

This could be pulled together rather easily by hacking Google Scholar.

Then you want to be told about papers that are similar to the papers you have written or cited.

Of course, what I am thinking about looks a lot like RSS feed or RSS feed item recommendation, but there are differences… the “items” invariably cite a lot of others items… people are interested in a very narrow set of “items”… there are few new items every day… and these items are very content-rich.

2. Paul - January 8, 2008

I totally agree that recommendations based upon social tags falls apart pretty quickly, even for ‘entertainment’ domains such as music when you get deep into the long tail where there are no longer enough tags to do make sense about the content. Content-based techniques are certainly going to be needed when their aren’t millions of tagger providing all of the metadata. To deal with this exact problem we’ve been working on autotagging systems, that will learn how to apply tags for items that don’t have any social tags. This yields a hybrid recommender that spans popular content and long tail content.

3. Andre Vellino - January 10, 2008

Ah, interesting. So Paul, would “tagomendation” on text-items using, say, keyphrase extraction, be equivalent to a hybrid content-based + CF recommender?

4. Andre Vellino - January 10, 2008

Daniel – I like your citation-based notifier. Could be a good one to add to CISTI’s next-gen portal.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: