Sunday, January 1st, 2006

New feature: Tag-based recommendations

Tag-based recommendations.

I have added a second set of book recommendations, this time based on the tags people apply to books. So if a book is tagged mostly “constitution” and “american history” it will match up with other books about the Constitution, favoring historical ones. The algorithm is quite complex, weighing the relative frequency of tags by book and overall, book popularity and more. I think you’ll find it’s quite solid.

Here are some examples:

The new system joins the existing system “people who own X also own Y.” More than solid this has the potential to provide truly inspired recommendations, crossing genres to get at what really binds people together around a book. For example, Brooks’ The Mythical Man Month triggers other software project-management books. But it also triggers a Dilbert book. People who do software project management have a lot of Dilbert books, and for good reason. Dilbert is, of course, a software engineer and half the jokes are about the stupidies of cubicles and “pointy headed bosses.” Or take the first Harry Potter. Amazon suggest five other books—all Harry Potters! LibraryThing also suggests Philip Pullman, Tolkien and—my favorite—Madeline L’Engle.

Connection uncertain.

Connection apparent.

Unfortunately, the system can backfire. Take Thomas a Kempis’ The Imitation of Christ, the spiritual reflections of a 15th century monk. This currently triggers Clifford the Big Red Dog. In this case duplicate copies are the problem, something that will shortly be fixed. But there’s a general problem when a book is held by only a few people. The results are statistically true, but not necessarily meaningful.

So there are now two systems, providing 25 recommendations in all (to Amazon’s five). They will certainly get better as users add more tags and books, but I would argue LibraryThing’s recommendations are already better than Amazon’s—richer, more inspired and based on what’s good to read NOT what’s selling this month. Community is, it seems, more powerful than commerce, even for commerce.

Finally, I have a third way of calculating recommendations, different from the other two. I don’t think I can talk how it works now, but it’s somewhere in between them in terms of noise vs. inspiration, and it works well when there are only a few books. The trick is that it’s very expensive—it takes forever to calculate! I think it will have to wait until I have a dedicated “thinking” machine. (If LibraryThing thing keeps growing, I’m forsee a stack of Mac Minis thinking all day.) I also plan to allow users to cheer and boo recommendations—Clifford boo!—and perhaps add their own.

And finally finally, I added another major new feature you may spot—RSS and HTML feeds all over the place. But I don’t want to step on my own toes and I have some quirks to work out before I announce it. Intrepid explorers are invited to let me know what you think in advance.

PS: Do you think I can get away with calling tag-based connections between books “contaguinity”?

UPDATE: Between yesterday and today Clifford slid off the Thomas a Kempis recommendations list, like a toaster that isn’t broken when you take it to the repair shop. [Yes, I know, repair shops vanished 30 years ago.] So I added them both to my library to force the connection. I’m NOT going to go looking for another example!

Labels: 1

0 Comments:

Leave a Reply