jump to navigation

Google Blog Reader Recommender November 30, 2007

Posted by Andre Vellino in Collaborative filtering, Recommender service.

Don’t you love it when Google just slips in these new features like easter eggs? The new Blog Reader recommender appears to be using a hybrid collaborative-filtering + content (from query expansion), which seems to work pretty well. I like the fact that the relevance ranking doesn’t seemed to be biased by the number of subscriptions to the blog.


1. lemire - November 30, 2007

I noticed this new tool too. Fantastic.

I am not sure how well it works. As always, with recommender systems, the accuracy feels so-so. Very subjective matter.

BTW your feed offers only partial content now. Do you mean it that way?

2. Google Reader Recommender « The Mendicant Bug - November 30, 2007

[…] with the recommendations.  I’ve been hearing about this all day from a number of sources [here, here, and here, to name a few].  RSS […]

3. Andre Vellino - November 30, 2007

I thought that the Google Reader recommender was pretty good (for me), but that was judging from a small sample. I suspect it depends a lot on the particular usage you make of Google search and the Reader.

I think that for such a feature to be valuable the “window” of CF information that is used to do the recommendation has to be quite narrow. It will be interesting to see how it evolves.

4. lemire - December 1, 2007

Quality-wise, you subscribed to how many of the feeds that were suggested to you?

In my case, I might have subscribed to one or two feeds out of the dozens or so that were suggested to me. But what is bad is that Google seems to think that I work in Theoretical Computer Science and Mathematics. Well. No. I don’t.

So, qualitatively, it misses the mark.

To be fair, Google Reader does not know that I have a blog. It does not know about my research papers. All it knows of, are the blogs I subscribed to.

What is happening, I think, is that there are many TCS and Math. blogs out there. I am not sure this is entirely accurate, but it seems that these folks are very active in the blogosphere. My research area is database, and I know very few database research with blogs.

So, I end up subscribing to a bit too many Math. and TCS blogs. This, in turns, get Google Reader to suggest more of those, and then if I accept the recommendations, it will then reinforce its belief and voilà! Things will get worse for me, not better.

I think that people commonly ignore feedback loops in recommender systems. A recommender process is essentially a non linear process, so getting it a bit wrong early on can have a big impact.

What should it do? It should give me *more control* over what I want recommended to me. This flat list forced upon me is just no good. I can’t even browse over the recommendations in any sensible way.
Then I can course-correct the recommender process!

5. Andre Vellino - December 2, 2007

You’re right – I didn’t endorse the recommendations sufficiently to subscribe to them. But I’m glad they’ve been made available to me. One difference between us is that I like these TCS guys and what they talk about.

I think that my reluctance to subscribe to Google’s recommendations is that I still don’t have a good way of (automatically) judging the quality of the content of what is recommended. I want some reputation measure and right now the best one is still word of mouth.

I *completely* agree with you about controlling the criteria that the recommender comes up with. What I want is an interactive map. There are some things I like about what http://www.quintura.com/ has done with it’s tag-cloud map-thing. We’re not quite there yet, but we’re getting there (I’ll know it when I see it, but I can’t specify it.)

Yes a good control mechanism would allow you to avoid the narrowing feedback mechanism. For blogs, control over the content-analysis part of the recommender would be crucial. My blog feeds are a bit like my news feeds – I’m interested in a lot of different things and my criterion is primarily quality.

6. Paul - December 5, 2007

Also, note that the Google reader recommender takes into account my search history, not just my blog reading behavior. I’m not sure if that is a good idea. For instance, I was writing a bit of Applescript the other day. Not knowing anything about AS, I did a lot of Google searches. Now my script is done, and my interest in Applescript is now back to zero. But the first recommendation in Google reader today for me is an Applescript site. Google has made the assumption that what I search for is the same as what I want to read about. I’m not sure if that is always the case. (Perhaps I should start blogging about the address to the local department of motor vehicles, I may get lots of google reader referrals ….)

7. Andre Vellino - December 5, 2007

Search history has its drawbacks, as you point out, but one of its virtues is to introduce some degree of serendipity. If the recommender only relied on existing blog feed subscription patterns (i.e. used mostly CF), I expect the recommendations would be pretty static. If it uses the blog feed-content as well, then that would offer some variety in the recommendations, depending on the subject diversity of the subscribed blog content.

I *definitely* agree about the need for explanations – especially as a feedback mechanism to control the space of recommendations.

Since I don’t know much about music recommenders, I was curious to see what Pandora does by way of explanation feedback, as you suggest in your blog entry:


Unfortunately, I get the following message from the Pandora web site (http://www.pandora.com/)

“…due to licensing constraints, we can no longer allow access to Pandora for most listeners located outside of the U.S. ”

I could proxy my Canadian IP address so that Pandora thinks I’m logging in from the U.S. but that would be morally reprehensible :-).

8. Paul - December 5, 2007


Yes, the legal situation with regard to streaming music internationally is disappointing. There are, however, a number of screenshots of Pandora on Flickr:


that should give you the flavor of how Pandora explains their recommendations. My favorite is this one:

pandora Thieves qualities

The comment by the poster is a great example of why transparent recommendations are important:
“Technology loosing it’s “cold”:
Pandora feels like a smart friend to me. This friend can articulate the reasons I love some of the things I love most (songs) better than I can, but only because I have told it what I like. THis is one of my very favorite Prince songs and Pandora knows just why I like it so much. And I didn’t know how to say it so well. Makes the technology seem very warm and reflective of my feelings an identity. It’s an extension of the user, not a cold, isolating technology. I feel a part of Pandora some times. I’ll bet they LOVE this. “

9. Google Recommends Blogs: Another PageRank? - December 10, 2007

[…] Andre, and many others, have written good things about the Google Reader recommender system: if you read […]

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: