One of the activities assigned to us for the first week was to explore Hunch – a recommendation engine of sorts.
It works in a simple way – it asks you twenty questions, and based on your responses, recommends items from various categories (magazines, games, food, etc) that are suited for you. Amazon has been doing this for some time now, and Facebook and Twitter attempt to do something similar. They recommends friends.
I chose not to get very deep in the understanding of how recommendation engines work. At least not till I wrote this post. We are in Tuesday already; nearly half the week is over. So it is a good idea for me to finish this post, so that I can go back and study if my initial thoughts on this have any basis at all.
First few thoughts on Hunch, per se (I’ll follow this up with a few thoughts on general recommendation engines):
- Privacy: As some of my classmates have noted, requiring a Twitter or a Facebook OAuth is a put off. I think the primary reason for this is that you are not sure before you “get into Hunch” what it is about – and given that you possibly have a lot of personal information (at least in Facebook), you’d think twice before using FB to sign in. For the common user, who is concerned about privacy, it s not a very useful method – a direct signup feature would be useful.
- Cultural/Regional: Hunch is very American. So if you are not an American, it doesn’t work very well for you. These are cultural aspects which Hunch probably did not consider. If I state that I am a non-vegetarian, I get recommendations of steak-houses, but I do not eat red meat, for example. Further, when you look at the recommendations, even if I want to act on them, chances are – they are not available in my country. As a Mac user, Hunch made some very interesting recommendations regarding Mac products, however, music via iTunes is not sold in India. Apples knows why, we don’t. There are no Apples stores in India, either.
- Commercial: Next, the commercial angle; to an extent related to the regional and cultural aspect. The eCommerce sites are all US-specific – even for books. The most popular online bookstore, Flipkart.com, doesn’t find a mention.
- Accuracy: Finally, the quality of the recommendations – they are far from relevant. Now, when I look at the results, I might say, hey, I find that interesting, but when I see a majority of the recommendations far from what I might really like, the algorithm seems to fail.
More often than not, recommendation engines have two primary shortcomings (a) they are based on a very narrow mapping (simplified relationships), and (b) they refer to a very limited data-set (limitation of available data).
Personally, if I were Amazon, I’d rather allow folks to link my Goodreads site and extract information from there, than just map to who bought other books when they bought the book that I am viewing right now. Goodreads seems to work on a data set that has been created by the user, and the system. When I do a book-compatibility check with friends on Goodreads, I’d like to believe I have a more closer understanding of recommendations from friends: I’d trust Goodreads more than I’d trust Amazon. The fact that someone bought some books is less valuable to me, than the fact that someone I know, liked (or disliked) the books – explained with a reason.
Recommendations as they exist today, I conclude (as of this post, I hope my opinion will be enriched as I go through this course), are slaves to the limitation of databases and the relationships that software architects can define, within the construct of a (software) program. Which means that soft-attributes that are more human and analogue, so to speak, aren’t a component of the mapping exercise.
In education, this takes on a different dimension. When learning online, It is quite possible (and easy) to recommend what content other students have liked, rated, which you might consider for your own consumption. The inherent problem in this is similar to (what I think) is called (Iterative) Economic Inequality: The rich getting richer. Allow me to explain this through an example:
iTunes has a built-in smart-playlist called “Top 25 Most Played.” It is a simple algorithm: it takes the 25 songs with the most play-counts. If you use this play-list often enough, the playlist does not undergo any significant change. Simply because the playcounts of these 25 songs increase (because you keep listening to these songs). The only way to break the authority of this play list is to choose songs manually, or select a custom smart playlist. Over a period, the “Top 25 Most Played smart playlist” changes. I refer back to my earlier statement, that this kind of recommendation is slavish to the limitation of the imagination and the construct of the program (or the programmer.)
When you add the variables of: learning styles (for those of us who still believe in it), variable pace of learning, environments in which we learn, cultural contexts in which we learn, and the economic situations which we prepare for, the recommendations and similar gimmickry are fairly redundant.
In (final) conclusion, I think that recommendations, used the way they are today serve little (and often, negative) benefits to learning. There needs to be either a radical shift in the way we think about this model, or it needs to encompass several more variables (which sounds infeasible to me) for it to make any sense in learning and teaching.
It may be worthwhile for you to see the model with Livemocha and Getglue.